Skip to content

M
Model-Transfer-Adaptability
Commit cba2e1d6 by Zhihong Ma
Options

fix: add quantization feature for loss trajectory mia

parent 68add01f
Showing with 5807 additions and 84 deletions
mzh/new_mzh/Loss_Trajectory_MIA/MIA.py
...@@ -8,7 +8,11 @@ import normal ...@@ -8,7 +8,11 @@ import normal
import dataset as DATA import dataset as DATA
from typing import TYPE_CHECKING, Callable, List, Optional, Tuple, Union from typing import TYPE_CHECKING, Callable, List, Optional, Tuple, Union
from sklearn import metrics from sklearn import metrics
import openpyxl
from global_var import GlobalVariables
import gol
import sys
class MLP_BLACKBOX(nn.Module): class MLP_BLACKBOX(nn.Module):
def __init__(self, dim_in): def __init__(self, dim_in):
super(MLP_BLACKBOX, self).__init__() super(MLP_BLACKBOX, self).__init__()
...@@ -88,7 +92,8 @@ def test_mia_attack_model(args, epoch, model, attack_test_loader, loss_fn, max_a ...@@ -88,7 +92,8 @@ def test_mia_attack_model(args, epoch, model, attack_test_loader, loss_fn, max_a
'fpr': fpr, 'fpr': fpr,
'tpr': tpr 'tpr': tpr
} }
np.save(f'./outputs/{args.data}_{args.model}_{args.model_distill}_trajectory_auc', save_data) # 不再save，不关注，暂时只关注acc
# np.save(f'./outputs/{args.data}_{args.model}_{args.model_distill}_trajectory_auc', save_data)
if accuracy > max_acc: if accuracy > max_acc:
max_acc = accuracy max_acc = accuracy
...@@ -177,17 +182,44 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'): ...@@ -177,17 +182,44 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'):
elif args.model == 'mobilenetv2': elif args.model == 'mobilenetv2':
model_name = '{}_mobilenetv2'.format(args.data) model_name = '{}_mobilenetv2'.format(args.data)
if args.quant_type is not None:
if args.quant_type == 'FLOAT':
title = '%s_%d_E%d' % (args.quant_type, args.num_bits, args.e_bits)
else:
title = '%s_%d' % (args.quant_type, args.num_bits)
# 在这里设置一下各种表
gol._init()
if args.quant_type != 'INT':
bias_list = utils.build_bias_list(args.quant_type)
gol.set_value(bias_list, is_bias=True)
if args.quant_type != 'INT':
plist = utils.build_list(args.quant_type, args.num_bits, args.e_bits)
gol.set_value(plist)
if args.mode == 'shadow': if args.mode == 'shadow':
cnn_model, cnn_params = normal.load_model(args, ori_model_path+'/shadow', model_name, epoch=args.epochs) cnn_model, cnn_params = normal.load_model(args, ori_model_path+'/shadow', model_name, epoch=args.epochs)
MODEL = cnn_model.to(device)
# 对于量化后的 这里应该要改吧，需要load量化后的模型来做推理
elif args.mode == 'target': elif args.mode == 'target':
cnn_model, cnn_params = normal.load_model(args, ori_model_path+'/target', model_name, epoch=args.epochs) cnn_model, cnn_params = normal.load_model(args, ori_model_path+'/target', model_name, epoch=args.epochs)
if args.quant_type is None:
MODEL = cnn_model.to(device) MODEL = cnn_model.to(device)
else:
ptq_file_prefix = 'networks/{}/{}'.format(args.seed, args.mode) + '/' + f'{args.data}_{args.model}/'
cnn_model.quantize(args.quant_type,args.num_bits,args.e_bits)
cnn_model.load_state_dict(torch.load(ptq_file_prefix + title + '.pt'))
MODEL = cnn_model.to(device)
MODEL.freeze()
print('Successfully load ptq model: ' + title)
print('pt file path:' + ptq_file_prefix + title + '.pt')
# params['task']中记录用什么数据集 # params['task']中记录用什么数据集
dataset = utils.get_dataset(cnn_params['task'], mode=args.mode, aug=True, batch_size=384) dataset = utils.get_dataset(cnn_params['task'], mode=args.mode, aug=True, batch_size=384)
if args.mode == 'target': if args.mode == 'target':
print('load target_dataset ... ') print('load target_dataset ... ')
train_loader = dataset.aug_target_train_loader train_loader = dataset.aug_target_train_loader
...@@ -218,13 +250,14 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'): ...@@ -218,13 +250,14 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'):
# distill model的output与label得到的loss trajectory 按一列列的形式组织为一个tensor # distill model的output与label得到的loss trajectory 按一列列的形式组织为一个tensor
batch_trajectory = get_trajectory(data, target, args, ori_model_path, device) batch_trajectory = get_trajectory(data, target, args, ori_model_path, device)
data, target = data.to(device), target.to(device) data, target = data.to(device), target.to(device)
batch_logit_target = MODEL(data) if args.quant_type is None:
batch_logit_target = MODEL(data)
else:
batch_logit_target = MODEL.quantize_inference(data)
# 每行的最大值 (predict label) # 最大值的位置 (predict label) 相当于是哪类
_, batch_predict_label = batch_logit_target.max(1) _, batch_predict_label = batch_logit_target.max(1)
batch_predicted_label = batch_predict_label.long().cpu().detach().numpy() batch_predicted_label = batch_predict_label.long().cpu().detach().numpy()
batch_original_label = target.long().cpu().detach().numpy() batch_original_label = target.long().cpu().detach().numpy()
# 相当于最后一次loss，可以拼接到loss trajectory末尾 # 相当于最后一次loss，可以拼接到loss trajectory末尾
...@@ -266,13 +299,25 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'): ...@@ -266,13 +299,25 @@ def build_trajectory_membership_dataset(args, ori_model_path, device='cpu'):
'nb_classes':dataset.num_classes 'nb_classes':dataset.num_classes
} }
# target model和shadow model的分别作为trajector_train_data, trajectory_test_data保存 # distill target model和distill shadow model的分别作为trajector_train_data, trajectory_test_data保存
# TODO 把(ptq) distill target model的存储路径进行修改
dataset_type = 'trajectory_train_data' if args.mode == 'shadow' else 'trajectory_test_data' dataset_type = 'trajectory_train_data' if args.mode == 'shadow' else 'trajectory_test_data'
utils.create_path(ori_model_path + f'/{args.mode}/{model_name}') utils.create_path(ori_model_path + f'/{args.mode}/{model_name}')
np.save(ori_model_path + f'/{args.mode}/{model_name}/{dataset_type}', data) model_name_ptq = model_name + '_' + title
utils.create_path(ori_model_path + f'/{args.mode}/{model_name_ptq}')
if args.quant_type is None:
np.save(ori_model_path + f'/{args.mode}/{model_name}/{dataset_type}', data)
else:
np.save(ori_model_path + f'/{args.mode}/{model_name_ptq}/{dataset_type}', data)
def trajectory_black_box_membership_inference_attack(args, models_path, device='cpu'): def trajectory_black_box_membership_inference_attack(args, models_path, device='cpu'):
if args.quant_type is not None:
if args.quant_type == 'FLOAT':
title = '%s_%d_E%d' % (args.quant_type, args.num_bits, args.e_bits)
else:
title = '%s_%d' % (args.quant_type, args.num_bits)
# if args.model == 'vgg': # if args.model == 'vgg':
# model_name = '{}_vgg16bn'.format(args.data) # model_name = '{}_vgg16bn'.format(args.data)
# elif args.model == 'mobilenet': # elif args.model == 'mobilenet':
...@@ -290,7 +335,7 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device=' ...@@ -290,7 +335,7 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device='
elif args.model == 'mobilenetv2': elif args.model == 'mobilenetv2':
model_name = '{}_mobilenetv2'.format(args.data) model_name = '{}_mobilenetv2'.format(args.data)
print(f"MODEL NAME IS :{model_name}")
# if args.model_distill == 'vgg': # if args.model_distill == 'vgg':
# model_distill_name = '{}_vgg16bn'.format(args.data) # model_distill_name = '{}_vgg16bn'.format(args.data)
# elif args.model_distill == 'mobilenet': # elif args.model_distill == 'mobilenet':
...@@ -309,24 +354,45 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device=' ...@@ -309,24 +354,45 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device='
elif args.model_distill == 'mobilenetv2': elif args.model_distill == 'mobilenetv2':
model_distill_name = '{}_mobilenetv2'.format(args.data) model_distill_name = '{}_mobilenetv2'.format(args.data)
print(f"MODEL DISTILL NAME IS :{model_distill_name}") if args.quant_type is None:
print(f"MODEL NAME IS :{model_name}")
print(f"MODEL DISTILL NAME IS :{model_distill_name}")
else:
print(f"MODEL NAME IS :{model_name}_{title}")
print(f"MODEL DISTILL NAME IS :{model_distill_name}_{title}")
cnn = model_name cnn = model_name
print(f'------------------model: {model_name}-------------------') # print(f'------------------model: {model_name}-------------------')
orgin_model_name = model_name orgin_model_name = model_name
model_name_ptq = model_name + '_' +title
# 唯一的attack model路径 (因为对target model量化后蒸馏并不会影响攻击者的训练)
save_path = models_path + '/attack/' + model_name save_path = models_path + '/attack/' + model_name
# if args.quant_type is None:
# save_path = models_path + '/attack/' + model_name
# else:
# save_path = models_path + '/attack/' + model_name_ptq
utils.create_path(save_path) utils.create_path(save_path)
best_prec1 = 0.0 best_prec1 = 0.0
best_auc = 0.0 best_auc = 0.0
epoch = 0
AttackModelTrainSet = np.load(models_path + f'/shadow/{model_name}/trajectory_train_data.npy', allow_pickle=True).item() AttackModelTrainSet = np.load(models_path + f'/shadow/{model_name}/trajectory_train_data.npy', allow_pickle=True).item()
AttackModelTestSet = np.load(models_path + f'/target/{model_name}/trajectory_test_data.npy', allow_pickle=True).item()
train_set = torch.utils.data.TensorDataset( # 支持load (ptq) distill target model loss trajectory
if args.quant_type is None:
AttackModelTestSet = np.load(models_path + f'/target/{model_name}/trajectory_test_data.npy', allow_pickle=True).item()
else:
AttackModelTestSet = np.load(models_path + f'/target/{model_name_ptq}/trajectory_test_data.npy', allow_pickle=True).item()
# not load trained model, need to train a new one
if args.load_attack is False:
train_set = torch.utils.data.TensorDataset(
torch.from_numpy(np.array(AttackModelTrainSet['model_loss_ori'], dtype='f')), torch.from_numpy(np.array(AttackModelTrainSet['model_loss_ori'], dtype='f')),
torch.from_numpy(np.array(AttackModelTrainSet['model_trajectory'], dtype='f')), torch.from_numpy(np.array(AttackModelTrainSet['model_trajectory'], dtype='f')),
# 转换为one-hot编码 # 转换为one-hot编码
...@@ -335,7 +401,8 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device=' ...@@ -335,7 +401,8 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device='
# train/test i.e. in or out # train/test i.e. in or out
torch.from_numpy(np.array(check_and_transform_label_format(AttackModelTrainSet['predicted_status'], nb_classes=2, return_one_hot=True)[:,:2])).type(torch.long), torch.from_numpy(np.array(check_and_transform_label_format(AttackModelTrainSet['predicted_status'], nb_classes=2, return_one_hot=True)[:,:2])).type(torch.long),
torch.from_numpy(np.array(AttackModelTrainSet['member_status'])).type(torch.long),) torch.from_numpy(np.array(AttackModelTrainSet['member_status'])).type(torch.long),)
attack_train_loader = torch.utils.data.DataLoader(train_set, batch_size=128, shuffle=True)
test_set = torch.utils.data.TensorDataset( test_set = torch.utils.data.TensorDataset(
torch.from_numpy(np.array(AttackModelTestSet['model_loss_ori'], dtype='f')), torch.from_numpy(np.array(AttackModelTestSet['model_loss_ori'], dtype='f')),
torch.from_numpy(np.array(AttackModelTestSet['model_trajectory'], dtype='f')), torch.from_numpy(np.array(AttackModelTestSet['model_trajectory'], dtype='f')),
...@@ -344,43 +411,103 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device=' ...@@ -344,43 +411,103 @@ def trajectory_black_box_membership_inference_attack(args, models_path, device='
# train/test i.e. in or out # train/test i.e. in or out
torch.from_numpy(np.array(check_and_transform_label_format(AttackModelTestSet['predicted_status'], nb_classes=2, return_one_hot=True)[:,:2])).type(torch.long), torch.from_numpy(np.array(check_and_transform_label_format(AttackModelTestSet['predicted_status'], nb_classes=2, return_one_hot=True)[:,:2])).type(torch.long),
torch.from_numpy(np.array(AttackModelTestSet['member_status'])).type(torch.long),) torch.from_numpy(np.array(AttackModelTestSet['member_status'])).type(torch.long),)
attack_train_loader = torch.utils.data.DataLoader(train_set, batch_size=128, shuffle=True)
attack_test_loader = torch.utils.data.DataLoader(test_set, batch_size=128, shuffle=True) attack_test_loader = torch.utils.data.DataLoader(test_set, batch_size=128, shuffle=True)
print(f'-------------------"Loss Trajectory"------------------') print(f'-------------------"Loss Trajectory"------------------')
# 训练Attack Model # 训练Attack Model
attack_model = MLP_BLACKBOX(dim_in = args.epochs_distill + 1) attack_model = MLP_BLACKBOX(dim_in = args.epochs_distill + 1)
attack_optimizer = torch.optim.SGD(attack_model.parameters(), lr=0.01, momentum=0.9, weight_decay=0.0001)
attack_model = attack_model.to(device) attack_model = attack_model.to(device)
attack_optimizer = torch.optim.SGD(attack_model.parameters(), lr=0.01, momentum=0.9, weight_decay=0.0001)
loss_fn = nn.CrossEntropyLoss() loss_fn = nn.CrossEntropyLoss()
max_auc = 0 max_auc = 0
max_acc = 0 max_acc = 0
for epoch in range(100): # not load trained Attack Model
train_loss, train_prec1 = train_mia_attack_model(args, epoch, attack_model, attack_train_loader, attack_optimizer, loss_fn, device) if args.load_attack is False:
val_loss, val_prec1, val_auc, max_auc, max_acc = test_mia_attack_model(args, epoch, attack_model, attack_test_loader, loss_fn, max_auc, max_acc, device) print("You should not retrain the Attack Model !!!")
is_best_prec1 = val_prec1 > best_prec1 sys.exit()
is_best_auc = val_auc > best_auc for epoch in range(100):
if is_best_prec1: train_loss, train_prec1 = train_mia_attack_model(args, epoch, attack_model, attack_train_loader, attack_optimizer, loss_fn, device)
best_prec1 = val_prec1 val_loss, val_prec1, val_auc, max_auc, max_acc = test_mia_attack_model(args, epoch, attack_model, attack_test_loader, loss_fn, max_auc, max_acc, device)
if is_best_auc: is_best_prec1 = val_prec1 > best_prec1
best_auc = val_auc is_best_auc = val_auc > best_auc
if epoch % 10 == 0: if is_best_prec1:
print(('epoch:{} \t train_loss:{:.4f} \t test_loss:{:.4f} \t train_prec1:{:.4f} \t test_prec1:{:.4f} \t val_prec1:{:.4f} \t val_auc:{:.4f}') best_prec1 = val_prec1
torch.save(attack_model.state_dict(), save_path + '/' + 'trajectory' + '.pkl')
print(f"Save Best ACC Attack Model: acc:{val_prec1}, auc:{val_auc}")
if is_best_auc:
best_auc = val_auc
if epoch % 10 == 0:
print(('epoch:{} \t train_loss:{:.4f} \t test_loss:{:.4f} \t train_prec1:{:.4f} \t test_prec1:{:.4f} \t val_prec1:{:.4f} \t val_auc:{:.4f}')
.format(epoch, train_loss, val_loss, .format(epoch, train_loss, val_loss,
train_prec1, val_prec1, val_prec1, val_auc)) train_prec1, val_prec1, val_prec1, val_auc))
print('Max AUC: ', max_auc)
print('Max ACC: ', max_acc/100) print(f'Load Trained Attack Model:')
# 这里可以改 仅存储is_best_auc or is_best_prec1的 attack_model.load_state_dict(torch.load(save_path + '/' + 'trajectory' + '.pkl'))
torch.save(attack_model.state_dict(), save_path + '/' + 'trajectory' + '.pkl') val_loss, val_prec1, val_auc, max_auc, max_acc = test_mia_attack_model(args, epoch, attack_model, attack_test_loader, loss_fn, max_auc, max_acc, device)
print(('Test load Attack Model \t test_loss:{:.4f} \t test_prec1:{:.4f} \t val_prec1:{:.4f} \t val_auc:{:.4f}').format( val_loss, val_prec1, val_prec1, val_auc))
# 在 test_mia_attack_model存储的最优的数据
data_auc = np.load(f'./outputs/{args.data}_{args.model}_{args.model_distill}_trajectory_auc.npy', allow_pickle=True).item() print('Max AUC: ', max_auc)
for i in range(len(data_auc['fpr'])): print('Max ACC: ', max_acc/100)
if data_auc['fpr'][i] > 0.001: # 这里可以改 仅存储is_best_auc or is_best_prec1的
print('TPR at 0.1% FPR: {:.1%}'.format(data_auc['tpr'][i-1])) # 在 test_mia_attack_model存储的最优的数据
break data_auc = np.load(f'./outputs/{args.data}_{args.model}_{args.model_distill}_trajectory_auc.npy', allow_pickle=True).item()
for i in range(len(data_auc['fpr'])):
if data_auc['fpr'][i] > 0.001:
print('TPR at 0.1% FPR: {:.1%}'.format(data_auc['tpr'][i-1]))
break
# load trained Attack Model
# 对于量化后，只会load trained attack model，不会重新训练.
else:
print(f'Load Trained Attack Model:')
attack_model.load_state_dict(torch.load(save_path + '/' + 'trajectory' + '.pkl'))
val_loss, val_prec1, val_auc, max_auc, max_acc = test_mia_attack_model(args, epoch, attack_model, attack_test_loader, loss_fn, max_auc, max_acc, device)
print(('Test load Attack Model \t test_loss:{:.4f} \t test_prec1:{:.4f} \t val_prec1:{:.4f} \t val_auc:{:.4f}').format( val_loss, val_prec1, val_prec1, val_auc))
print('Max AUC: ', max_auc)
print('Max ACC: ', max_acc/100)
filename =f'{args.model}_result.xlsx'
try:
# 如果文件已经存在，则加载它
workbook = openpyxl.load_workbook(filename)
except FileNotFoundError:
# 如果文件不存在，则创建一个新的Excel工作簿
workbook = openpyxl.Workbook()
if args.data not in workbook.sheetnames:
# 如果工作表不存在，则创建一个新的工作表
worksheet = workbook.create_sheet(title=args.data)
# 在工作表中写入表头
worksheet.cell(row=1,column=1,value='FP32-acc')
# worksheet.cell(row=1,column=2,value=top1)
worksheet.cell(row=3,column=1,value='title')
worksheet.cell(row=3,column=2,value='js_div')
worksheet.cell(row=3,column=4,value='ptq_acc')
worksheet.cell(row=3,column=5,value='acc_loss')
worksheet.cell(row=3,column=6,value='AUC')
worksheet.cell(row=3,column=7,value='ACC')
else:
worksheet = workbook[args.data]
# 全精度下的
if args.quant_type is None:
worksheet.cell(row=1,column=4,value='origin_auc')
worksheet.cell(row=1,column=5,value=max_auc)
worksheet.cell(row=1,column=7,value='origin_acc')
worksheet.cell(row=1,column=8,value=max_acc/100)
else:
idx = GlobalVariables.title_list.index(title)
idx += 4
worksheet.cell(row=idx,column=1,value=title)
worksheet.cell(row=idx,column=6,value=max_auc)
worksheet.cell(row=idx,column=7,value=max_acc/100)
workbook.save(filename)
# #
def get_trajectory(data, target, args, model_path, device='cpu'): def get_trajectory(data, target, args, model_path, device='cpu'):
...@@ -408,6 +535,11 @@ def get_trajectory(data, target, args, model_path, device='cpu'): ...@@ -408,6 +535,11 @@ def get_trajectory(data, target, args, model_path, device='cpu'):
# 创建一个形状为 (data.shape[0], 1) 的 NumPy 数组 predicted_label，并将其初始化为 -1 data.shape[0]即batch_size # 创建一个形状为 (data.shape[0], 1) 的 NumPy 数组 predicted_label，并将其初始化为 -1 data.shape[0]即batch_size
predicted_label = np.array([-1]).repeat(data.shape[0],0).reshape(data.shape[0],1) predicted_label = np.array([-1]).repeat(data.shape[0],0).reshape(data.shape[0],1)
# TODO 需要适配由PTQ Target Model得到的Distill Target Model
# 虽然是通过mode == shadow和target来区分，但load的model是ditill model
for s in range(1): for s in range(1):
trajectory_current = None trajectory_current = None
model_path_current = 'networks/{}'.format(s) model_path_current = 'networks/{}'.format(s)
...@@ -417,14 +549,29 @@ def get_trajectory(data, target, args, model_path, device='cpu'): ...@@ -417,14 +549,29 @@ def get_trajectory(data, target, args, model_path, device='cpu'):
if args.mode == 'shadow': if args.mode == 'shadow':
cnn_model_target, cnn_params_target = normal.load_model(args, model_path_current+'/distill_shadow', model_name, epoch=i) cnn_model_target, cnn_params_target = normal.load_model(args, model_path_current+'/distill_shadow', model_name, epoch=i)
elif args.mode == 'target': elif args.mode == 'target':
cnn_model_target, cnn_params_target = normal.load_model(args, model_path_current+'/distill_target', model_name, epoch=i) if args.quant_type is None:
cnn_model_target, cnn_params_target = normal.load_model(args, model_path_current+'/distill_target', model_name, epoch=i)
# TODO 调整load的路径，把(ptq) Distill Target Model的权值参数load进来
else:
if args.quant_type == 'FLOAT':
title = '%s_%d_E%d' % (args.quant_type, args.num_bits, args.e_bits)
else:
title = '%s_%d' % (args.quant_type, args.num_bits)
cnn_model_target, cnn_params_target = normal.load_model(args, model_path_current+'/distill_target', model_name + '_' + title, epoch=i)
MODEL_target = cnn_model_target.to(device) MODEL_target = cnn_model_target.to(device)
# data是一个包含batch个向量的list # data是一个包含batch个向量的list
data = data.to(device) data = data.to(device)
# label # label
target = target.to(device) target = target.to(device)
# 获取target model的输出 # 获取target model的输出
# 都是distill model的output，因此不需要quantize_inference
logit_target = MODEL_target(data) logit_target = MODEL_target(data)
# 看target model的output与label的loss （batch list中的各个数据分别算） # 看target model的output与label的loss （batch list中的各个数据分别算）
loss = [F.cross_entropy(logit_target_i.unsqueeze(0), target_i.unsqueeze(0)) for (logit_target_i, target_i) in zip(logit_target, target)] loss = [F.cross_entropy(logit_target_i.unsqueeze(0), target_i.unsqueeze(0)) for (logit_target_i, target_i) in zip(logit_target, target)]
# list -> nparray 一列的 # list -> nparray 一列的
......
mzh/new_mzh/Loss_Trajectory_MIA/div.py 0 → 100644
import openpyxl
from global_var import GlobalVariables
from utils import *
import gol
import argparse
import numpy as np
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='DIV_TrajectoryMIA')
parser.add_argument('--model', type=str, default='resnet18', help=['resnet18','resnet50','resnet152','mobilenetv2'])
parser.add_argument('--data', type=str, default='cifar10', help=['cinic10', 'cifar10', 'cifar100', 'gtsrb'])
args = parser.parse_args()
data_path = f'networks/0/target/{args.data}_{args.model}/trajectory_test_data.npy'
# 统一计算所有的js
gol._init()
quant_type_list = ['INT','POT','FLOAT']
filename =f'{args.model}_result.xlsx'
workbook = openpyxl.load_workbook(filename)
worksheet = workbook[args.data]
for quant_type in quant_type_list:
num_bit_list = numbit_list(quant_type)
for num_bits in num_bit_list:
e_bit_list = ebit_list(quant_type,num_bits)
for e_bits in e_bit_list:
if quant_type == 'FLOAT':
title = '%s_%d_E%d' % (quant_type, num_bits, e_bits)
else:
title = '%s_%d' % (quant_type, num_bits)
model_name_ptq = f'{args.data}_{args.model}_{title}'
p_data_path = f'networks/0/target/{model_name_ptq}/trajectory_test_data.npy'
dataSet = np.load(data_path, allow_pickle=True).item()
p_dataSet = np.load(p_data_path, allow_pickle=True).item()
data = torch.from_numpy(np.array(dataSet['model_trajectory'], dtype='f')).transpose(0,1)
p_data = torch.from_numpy(np.array(p_dataSet['model_trajectory'], dtype='f')).transpose(0,1)
# data = torch.from_numpy(np.array(dataSet['model_trajectory'], dtype='f'))
# p_data = torch.from_numpy(np.array(p_dataSet['model_trajectory'], dtype='f'))
div = js_div(data,p_data)
div = div.item()
if div<0:
div = 0
print(f"js div of {model_name_ptq}: {div}")
idx = GlobalVariables.title_list.index(title)
idx += 4
worksheet.cell(row=idx,column=2,value=div)
workbook.save(filename)
mzh/new_mzh/Loss_Trajectory_MIA/global_var.py
class GlobalVariables: class GlobalVariables:
SELF_INPLANES = 0 SELF_INPLANES = 0
title_list = ['INT_2','INT_3','INT_4','INT_5','INT_6','INT_7','INT_8','INT_9','INT_10','INT_11','INT_12','INT_13','INT_14','INT_15','INT_16','POT_2','POT_3','POT_4','POT_5','POT_6','POT_7','POT_8','FLOAT_3_E1','FLOAT_4_E1','FLOAT_4_E2','FLOAT_5_E1','FLOAT_5_E2','FLOAT_5_E3','FLOAT_6_E1','FLOAT_6_E2','FLOAT_6_E3','FLOAT_6_E4','FLOAT_7_E1','FLOAT_7_E2','FLOAT_7_E3','FLOAT_7_E4','FLOAT_7_E5','FLOAT_8_E1','FLOAT_8_E2','FLOAT_8_E3','FLOAT_8_E4','FLOAT_8_E5','FLOAT_8_E6']
\ No newline at end of file
mzh/new_mzh/Loss_Trajectory_MIA/main.py
...@@ -44,6 +44,12 @@ if __name__ == '__main__': ...@@ -44,6 +44,12 @@ if __name__ == '__main__':
parser.add_argument('--model_distill', type=str, default='resnet18',help=['resnet18','resnet50','resnet152','mobilenetv2']) parser.add_argument('--model_distill', type=str, default='resnet18',help=['resnet18','resnet50','resnet152','mobilenetv2'])
parser.add_argument('--epochs_distill', type=int, default=100) parser.add_argument('--epochs_distill', type=int, default=100)
parser.add_argument('--mia_type', type=str, help=['build-dataset', 'black-box']) parser.add_argument('--mia_type', type=str, help=['build-dataset', 'black-box'])
parser.add_argument('--load_attack',action='store_true', help='load a trained attack model')
parser.add_argument('--store_ptq',action='store_true', help='store a ptq model')
parser.add_argument('--quant_type', type=str, choices=['INT', 'POT', 'FLOAT'], default=None,help='choose a ptq mode for target model')
parser.add_argument("--num_bits",type=int,default=0)
parser.add_argument("--e_bits",type=int,default=0)
parser.add_argument('--load_ptq',action='store_true', help='load a ptq target model')
args = parser.parse_args() args = parser.parse_args()
utils.set_random_seeds(args.seed) utils.set_random_seeds(args.seed)
......
mzh/new_mzh/Loss_Trajectory_MIA/normal.py
...@@ -6,26 +6,184 @@ import numpy as np ...@@ -6,26 +6,184 @@ import numpy as np
import pickle import pickle
import utils import utils
from architectures import * from architectures import *
import os.path as osp
import dataset
import gol
import openpyxl
from global_var import GlobalVariables
def direct_quantize(args, model, data, device='cpu'):
print("====Begin Direct Quantize!====")
if not hasattr(model, 'augment_training') or model.augment_training:
if args.mode == 'distill_target':
print('load aug_target_dataset ... ')
train_loader = data.aug_target_train_loader
elif args.mode == 'distill_shadow':
print('load aug_shadow_dataset ...')
train_loader = data.aug_shadow_train_loader
else:
if args.mode == 'distill_target':
print('load target_dataset ... ')
train_loader = data.target_train_loader
elif args.mode == 'distill_shadow':
print('load shadow_dataset ...')
train_loader = data.shadow_train_loader
model.eval()
count = 0
for x, y, idx in train_loader:
# cnn_training_step(model, optimizer, x, y, device)
b_x = x.to(device)
b_y = y.to(device)
output = model.quantize_forward(b_x)
count += 1
# TODO 可能对不同数据集需要修改
if count % 500 == 0:
break
print('direct quantization finish')
def full_inference(args, model, data, device='cpu'):
print(f'model.augument_training:{model.augment_training}')
if not hasattr(model, 'augment_training') or model.augment_training:
if args.mode == 'distill_target':
print('load aug_target_dataset ... ')
test_loader = data.aug_target_test_loader
elif args.mode == 'distill_shadow':
print('load aug_shadow_dataset ...')
test_loader = data.aug_shadow_test_loader
else:
if args.mode == 'distill_target':
print('load target_dataset ... ')
test_loader = data.target_test_loader
elif args.mode == 'distill_shadow':
print('load shadow_dataset ...')
test_loader = data.shadow_test_loader
model.eval()
top1 = dataset.AverageMeter()
top5 = dataset.AverageMeter()
with torch.no_grad():
for batch in test_loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
output = model(b_x)
prec1, prec5 = dataset.accuracy(output, b_y, topk=(1, 5))
top1.update(prec1[0], b_x.size(0))
top5.update(prec5[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top5_acc = top5.avg.data.cpu().numpy()[()]
print('\nTest set: Full Model Accuracy: {:.2f}%'.format(top1_acc))
return top1_acc, top5_acc
def quantize_inference(args, model, data, device='cpu'):
if not hasattr(model, 'augment_training') or model.augment_training:
if args.mode == 'distill_target':
print('load aug_target_dataset ... ')
test_loader = data.aug_target_test_loader
elif args.mode == 'distill_shadow':
print('load aug_shadow_dataset ...')
test_loader = data.aug_shadow_test_loader
else:
if args.mode == 'distill_target':
print('load target_dataset ... ')
test_loader = data.target_test_loader
elif args.mode == 'distill_shadow':
print('load shadow_dataset ...')
test_loader = data.shadow_test_loader
model.eval()
top1 = dataset.AverageMeter()
top5 = dataset.AverageMeter()
with torch.no_grad():
for batch in test_loader:
b_x = batch[0].to(device)
b_y = batch[1].to(device)
output = model.quantize_inference(b_x)
prec1, prec5 = dataset.accuracy(output, b_y, topk=(1, 5))
top1.update(prec1[0], b_x.size(0))
top5.update(prec5[0], b_x.size(0))
top1_acc = top1.avg.data.cpu().numpy()[()]
top5_acc = top5.avg.data.cpu().numpy()[()]
print('\nTest set: PTQ Model Accuracy: {:.2f}%'.format(top1_acc))
return top1_acc, top5_acc
# def direct_quantize(model, test_loader,device):
# for i, (data, target) in enumerate(test_loader, 1):
# data = data.to(device)
# output = model.quantize_forward(data).cpu()
# if i % 500 == 0:
# break
# print('direct quantization finish')
# def full_inference(model, test_loader, device):
# correct = 0
# for i, (data, target) in enumerate(test_loader, 1):
# data = data.to(device)
# output = model(data).cpu()
# pred = output.argmax(dim=1, keepdim=True)
# # print(pred)
# correct += pred.eq(target.view_as(pred)).sum().item()
# print('\nTest set: Full Model Accuracy: {:.2f}%'.format(100. * correct / len(test_loader.dataset)))
# return 100. * correct / len(test_loader.dataset)
# def quantize_inference(model, test_loader, device):
# correct = 0
# for i, (data, target) in enumerate(test_loader, 1):
# data = data.to(device)
# output = model.quantize_inference(data).cpu()
# pred = output.argmax(dim=1, keepdim=True)
# correct += pred.eq(target.view_as(pred)).sum().item()
# print('Test set: Quant Model Accuracy: {:.2f}%'.format(100. * correct / len(test_loader.dataset)))
# return 100. * correct / len(test_loader.dataset)
# 对 是否distill，有不同的train方法 # 对 是否distill，有不同的train方法
# 之前已经创建了model并把params config一起存储到了相应路径，此处先把model和params config load出来再trian # 之前已经创建了model并把params config一起存储到了相应路径，此处先把model和params config load出来再trian
# model_path_tar, model_path_dis = 'networks/{}/{}'.format(args.seed, args.mode)
# untrained_model_tar, untrained_model_dis => model_name = '{}_mobilenetv2'.format(args.data)...
def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untrained_model_dis = None, device='cpu'): def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untrained_model_dis = None, device='cpu'):
print('Training models...') print('Training models...')
# 蒸馏训练 # 蒸馏训练
# load_model返回的是model
if 'distill' in args.mode: if 'distill' in args.mode:
# load进来还没train的target model (create的时候就会save untrained) # load进来还没train(epoch=0)的distill model (create的时候就会save untrained)
# 把name传进来便于load
trained_model, model_params = load_model(args, model_path_dis, untrained_model_dis, epoch=0) trained_model, model_params = load_model(args, model_path_dis, untrained_model_dis, epoch=0)
# load进来已经训好的target model # load进来已经训好的target model
# TODO load进来一个trained Target Model (args.epochs已经是最后一个训练epoch)
trained_model_tar, model_params_tar = load_model(args, model_path_tar, untrained_model_tar, epoch=args.epochs) trained_model_tar, model_params_tar = load_model(args, model_path_tar, untrained_model_tar, epoch=args.epochs)
trained_model_tar.to(device)
# 正常训练 # 正常训练
else: else:
# load进来还没train的target model (create的时候就会save untrained) # load进来还没train的target model (create的时候就会save untrained)
trained_model, model_params = load_model(args, model_path_tar, untrained_model_tar, epoch=0) trained_model, model_params = load_model(args, model_path_tar, untrained_model_tar, epoch=0)
print(model_params) print(model_params)
# 获得以划分好的且做了数据增强dataset trained_model.to(device)
# 获得以划分好的且做了数据增强dataset distill model dataset
dataset = utils.get_dataset(model_params['task'], args.mode, aug=True) dataset = utils.get_dataset(model_params['task'], args.mode, aug=True)
if args.mode == 'distill_target':
dataset_t = utils.get_dataset(model_params['task'],mode='target',aug=True)
elif args.mode == 'distill_shadow':
dataset_t = utils.get_dataset(model_params['task'],mode='shadow',aug=True)
# 做一系列超参数设置 # 做一系列超参数设置
# TODO 调整学习策略 看看改进后的效果 # TODO 调整学习策略 看看改进后的效果
learning_rate = model_params['learning_rate'] learning_rate = model_params['learning_rate']
...@@ -36,16 +194,109 @@ def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untr ...@@ -36,16 +194,109 @@ def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untr
optimization_params = (learning_rate, weight_decay, momentum) optimization_params = (learning_rate, weight_decay, momentum)
optimizer, scheduler = utils.get_full_optimizer(trained_model, optimization_params, args) optimizer, scheduler = utils.get_full_optimizer(trained_model, optimization_params, args)
# model_name : dataset + model的形式
if 'distill' in args.mode: if 'distill' in args.mode:
trained_model_name = untrained_model_dis trained_model_name = untrained_model_dis
else: else:
trained_model_name = untrained_model_tar trained_model_name = untrained_model_tar
print('Training: {}...'.format(trained_model_name)) # TODO 增加量化
trained_model.to(device) if args.quant_type is not None:
print('Training: {}... with PTQ Target Model'.format(trained_model_name))
else:
print('Training: {}... with Target Model'.format(trained_model_name))
# 具体训练 # 具体训练
# metric中可以不断append，记录每一步的具体数据 # metric中可以不断append，记录每一步的具体数据
gol._init()
# TODO
if args.quant_type is not None:
top1, top5 = full_inference(args,trained_model_tar, dataset_t, device)
# 对trained_model_tar进行PTQ量化
if args.quant_type is not None and 'distill' in args.mode:
if args.quant_type != 'INT':
bias_list = utils.build_bias_list(args.quant_type)
gol.set_value(bias_list, is_bias=True)
if args.quant_type == 'FLOAT':
title = '%s_%d_E%d' % (args.quant_type, args.num_bits, args.e_bits)
else:
title = '%s_%d' % (args.quant_type, args.num_bits)
print('\nPTQ: '+title)
# ptq target model存储路径
ptq_file_prefix = model_path_tar + '/' + f'{args.data}_{args.model}/'
# 设置量化表
if args.quant_type != 'INT':
plist = utils.build_list(args.quant_type, args.num_bits, args.e_bits)
gol.set_value(plist)
# 判断是否需要载入
if args.load_ptq is True and osp.exists(ptq_file_prefix + title + '.pt'):
trained_model_tar.quantize(args.quant_type,args.num_bits,args.e_bits)
trained_model_tar.load_state_dict(torch.load(ptq_file_prefix + title + '.pt'))
trained_model_tar.to(device)
print('Successfully load ptq model: ' + title)
else:
trained_model_tar.quantize(args.quant_type,args.num_bits,args.e_bits)
# TODO
trained_model_tar.eval()
direct_quantize(args,trained_model_tar, dataset_t, device)
if args.store_ptq == True:
torch.save(trained_model_tar.state_dict(), ptq_file_prefix + title + '.pt')
trained_model_tar.freeze()
# TODO
ptq_top1, ptq_top5 = quantize_inference(args,trained_model_tar, dataset_t, device)
acc_loss = (top1 - ptq_top1) / top1
print(f"Target Model Quantization Finished, Acc Loss:{acc_loss}")
filename =f'{args.model}_result.xlsx'
try:
# 如果文件已经存在，则加载它
workbook = openpyxl.load_workbook(filename)
except FileNotFoundError:
# 如果文件不存在，则创建一个新的Excel工作簿
workbook = openpyxl.Workbook()
if args.data not in workbook.sheetnames:
# 如果工作表不存在，则创建一个新的工作表
worksheet = workbook.create_sheet(title=args.data)
# 在工作表中写入表头
worksheet.cell(row=1,column=1,value='FP32-acc')
worksheet.cell(row=1,column=2,value=top1)
worksheet.cell(row=3,column=1,value='title')
worksheet.cell(row=3,column=2,value='js_div')
worksheet.cell(row=3,column=4,value='ptq_acc')
worksheet.cell(row=3,column=5,value='acc_loss')
worksheet.cell(row=3,column=6,value='AUC')
worksheet.cell(row=3,column=7,value='ACC')
else:
worksheet = workbook[args.data]
worksheet.cell(row=1,column=2,value=top1)
idx = GlobalVariables.title_list.index(title)
idx += 4
worksheet.cell(row=idx,column=1,value=title)
worksheet.cell(row=idx,column=4,value=ptq_top1)
worksheet.cell(row=idx,column=5,value=acc_loss)
workbook.save(filename)
# 前面为了load untrained distill model，所以未改distill_target model的路径model_path_dis
if 'distill' in args.mode: if 'distill' in args.mode:
metrics = trained_model.train_func(args, trained_model_tar, trained_model, dataset, num_epochs, optimizer, scheduler, model_params, model_path_dis, trained_model_name, device=device) metrics = trained_model.train_func(args, trained_model_tar, trained_model, dataset, num_epochs, optimizer, scheduler, model_params, model_path_dis, trained_model_name, device=device)
else: else:
...@@ -62,11 +313,18 @@ def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untr ...@@ -62,11 +313,18 @@ def train(args, model_path_tar, untrained_model_tar, model_path_dis = None, untr
model_params['total_time'] = total_training_time model_params['total_time'] = total_training_time
print('Training took {} seconds...'.format(total_training_time)) print('Training took {} seconds...'.format(total_training_time))
# 存储训练后的模型权值参数和model_params # 存储训练后的模型权值参数和model_params (区分了用到的target model是否量化/什么量化)
if 'distill' in args.mode: # model_path_dis = 'networks/{}/{}'.format(args.seed, args.mode)
save_model(trained_model, model_params, model_path_dis, trained_model_name, epoch=num_epochs) trained_model_name_ptq = trained_model_name + '_' +title
# 在networks/{}/{}'.format(args.seed, args.mode)/trained_model_name + title 下存储
if args.quant_type is not None and 'distill' in args.mode:
save_model(trained_model, model_params, model_path_dis, trained_model_name_ptq, epoch=num_epochs)
else: else:
save_model(trained_model, model_params, model_path_tar, trained_model_name, epoch=num_epochs) if 'distill' in args.mode:
save_model(trained_model, model_params, model_path_dis, trained_model_name, epoch=num_epochs)
else:
save_model(trained_model, model_params, model_path_tar, trained_model_name, epoch=num_epochs)
# 配置模型信息,创建并训练模型 # 配置模型信息,创建并训练模型
def train_models(args, model_path_tar, model_path_dis, device='cpu'): def train_models(args, model_path_tar, model_path_dis, device='cpu'):
...@@ -107,6 +365,9 @@ def train_models(args, model_path_tar, model_path_dis, device='cpu'): ...@@ -107,6 +365,9 @@ def train_models(args, model_path_tar, model_path_dis, device='cpu'):
cnn_dis = create_mobilenetv2(model_path_dis, args) cnn_dis = create_mobilenetv2(model_path_dis, args)
# load untrained model和model_params，开始训练 # load untrained model和model_params，开始训练
# model_path_tar,model_path_dis = 'networks/{}/{}'.format(args.seed, args.mode)
# cnn_tar, cnn_dis => model_name = '{}_mobilenetv2'.format(args.data)...
train(args, model_path_tar, cnn_tar, model_path_dis, cnn_dis, device = device) train(args, model_path_tar, cnn_tar, model_path_dis, cnn_dis, device = device)
else: else:
train(args, model_path_tar, cnn_tar, device=device) train(args, model_path_tar, cnn_tar, device=device)
...@@ -189,19 +450,6 @@ def create_vgg16bn(model_path, args): ...@@ -189,19 +450,6 @@ def create_vgg16bn(model_path, args):
return model_name return model_name
# def create_mobile(model_path, args):
# print('Creating MobileNet untrained {} models...'.format(args.data))
# model_params = get_data_params(args.data)
# model_name = '{}_mobilenet'.format(args.data)
# model_params['network_type'] = 'mobilenet'
# model_params['cfg'] = [64, (128,2), 128, (256,2), 256, (512,2), 512, 512, 512, 512, 512, (1024,2), 1024]
# model_params['augment_training'] = True
# model_params['init_weights'] = True
# get_lr_params(model_params, args)
# model_name = save_networks(args, model_name, model_params, model_path)
# return model_name
def create_resnet56(models_path, args): def create_resnet56(models_path, args):
print('Creating resnet56 untrained {} models...'.format(args.data)) print('Creating resnet56 untrained {} models...'.format(args.data))
...@@ -310,7 +558,7 @@ def create_wideresnet32_4(models_path, args): ...@@ -310,7 +558,7 @@ def create_wideresnet32_4(models_path, args):
return model_name return model_name
# 实例化model，并调用save_model存储 # 实例化model，并调用save_model存储，只在create model的时候用到
def save_networks(args, model_name, model_params, model_path): def save_networks(args, model_name, model_params, model_path):
print('Saving CNN...') print('Saving CNN...')
model_params['base_model'] = model_name model_params['base_model'] = model_name
...@@ -343,6 +591,10 @@ def save_networks(args, model_name, model_params, model_path): ...@@ -343,6 +591,10 @@ def save_networks(args, model_name, model_params, model_path):
def save_model(model, model_params, model_path, model_name, epoch=-1): def save_model(model, model_params, model_path, model_name, epoch=-1):
if not os.path.exists(model_path): if not os.path.exists(model_path):
os.makedirs(model_path) os.makedirs(model_path)
# model_name就是 数据集_网络架构
# 'networks/{}/{}'.format(args.seed, args.mode) + model_name
# 若是ptq的，加了title，否则，只是args.data + args.model
network_path = model_path + '/' + model_name network_path = model_path + '/' + model_name
if not os.path.exists(network_path): if not os.path.exists(network_path):
os.makedirs(network_path) os.makedirs(network_path)
......
mzh/new_mzh/Loss_Trajectory_MIA/readme.md
## Loss Trajectory MIA ## Loss Trajectory MIA
#### Update 2023.6.1
1\. 增加的内容：
- 使用量化后的Target Model训练Distill Target Model (简称为Q-Disitll)
- 通过Q-Distill的loss trajectory构建Attack Model的测试集
- 使用Attack Model攻击量化后的Target Model
- 使用JS散度衡量量化Target Model前后得到的Distill Target Model的Loss Trajectory数据的相似度
2\. 思路
- 攻击场景假设：
首先继承了之前叙述的攻击场景(假设攻击者知道Target Model的结构，且对数据集有一定了解，因此将Shadow Model，Distill Target Model，Distill Shadow Model都采用了于Target Model相同的结构，将CIFAR10数据集切割成多个不相交的部分，分别作为Target Model，Shadow Model，Distill Target Model，Distill Shadow Model的训练、测试集)；
还假设了攻击者不知道或不在意Target Model是否为量化模型，仍然采用FP32的Shadow Model，Distill Target Model，Distill Shadow Model.
保持原有的Attack Model，攻击量化后的Target Model，通过Acc和Auc指标的变化来检验模型迁移的隐私安全性。
- 实验设计
由攻击场景假设，我们需要改动的部分集中在了Distill Target Model的训练，Attack Model的测试集数据构建，Attack Model的测试这三方面。
Distill Target Model在训练时，学习的Target Model的output由原来的FP32模型的输出改为量化后的Target Model的输出，对量化后的Target Model进行蒸馏，把Distill Target Model重新训练。
由于攻击者的Shadow Model和Distill Shadow Model仍然采用FP32的模型，因此不需要重新训练，相应的，Attack Model是通过Distill Shadow Model的Loss Trajectory构建的训练集训练的，Attack Model也不需要重新训练。
分别将经过各种量化后的Distill Target Model的Loss Trajectory构造成Attack Model的测试数据集，并对Attack Model进行测试，记录Best Acc和相应的Auc.
计算量化前后得到的Loss Trajectory的JS散度，作为相似度衡量依据。
- 预期结果
由于Attack Model的训练是由Shadow Model和Distill Shadow Model决定的，他们都是FP32的模型，因此量化后的Target Model与FP32的Target Model越相似，则Distill Model的Loss Trajectory越接近，Attack Model的攻击效果应该会越好，迁移后的隐私安全性越差。
3\. 实验结果
在ResNet18 + CIFAR10上进行了实验，得到的数据在**resnet18_result.xlsx**中。以下为拟合曲线，经过尝试，仍是分子分母次数为2的有理数函数效果最好。
acc_loss - js
<img src = "fig/acc_loss_curve.png" class="h-90 auto">
auc_loss - js
<img src = "fig/auc_loss_curve.png" class="h-90 auto">
在js距离较小(loss trajectory较为相似)时，acc_loss和auc_loss相对波动较大，说明对于量化后的Target Model，即使与FP32的Target Model较为相似，也可能具有隐私安全性提升。
4\. 问题
Q1： 运行速度慢。对每一种量化，都需要经过Target Model量化，训练一个相应的Distill Target Model，构造Attack Model测试集，测试Attack Model，流程很长。
A1： 目前准备减少一些比较慢的测试点，主要是FLOAT系列的，包括FLOAT_6_E2, FLOAT_7_E2, FLOAT_7_E3, FLOAT8_E2, FLOAT_8_E3, FLOAT_8_E6. (主要是FLOAT的量化的比较慢，实际上distill model在Resnet50+CIFAR10的情况下每个epoch的训练用时小于1min，distill model的训练耗时在使用CIFAR10,100时不是非常大，如果使用CINIC-10，则distill model的训练耗时也变得过于大了)
​ POT_7, POT_8也可以考虑去掉。上述计划去掉的数据点都是信息比较重复，对曲线趋势影响小，而且量化耗时比较长的。
Q2：对各个模型、数据集实验的流程较长。
**注**：.sh训练脚本集中放到了sh文件夹中，其中``train_attack_18_10_ptq_FP_S1， train_attack_18_10_ptq_FP_S2，train_attack_18_10_ptq_FP_S3，train_attack_18_10_ptq_FP_S4，train_attack_18_10_ptq_INT_S1，train_attack_18_10_ptq_INT_S2，train_attack_18_10_ptq_POT，train_div.sh ``等是本次更新用到的。
<br><br>
#### Update 2023.5.28 #### Update 2023.5.28
1\. 思路 1\. 思路
...@@ -12,7 +79,7 @@ ...@@ -12,7 +79,7 @@
但传统方法无法区分那些虽然不是训练集的数据，但模型loss仍然很小的输入。文章作者发现了在模型的训练过程中，对于这类小loss数据，在训练集或不在训练集会在训练过程的loss收敛速度和进程方面有差异。如图所示，他们有着不同的Loss轨迹(Loss Trajectory) 但传统方法无法区分那些虽然不是训练集的数据，但模型loss仍然很小的输入。文章作者发现了在模型的训练过程中，对于这类小loss数据，在训练集或不在训练集会在训练过程的loss收敛速度和进程方面有差异。如图所示，他们有着不同的Loss轨迹(Loss Trajectory)
![p1](fig\p1.png) <img src = "fig/p1.png" class="h-90 auto">
因为小loss数据如果不在训练集中，一般是比较简单的图片，其训练loss会下降的很快，因此其loss轨迹在训练中期是低于在训练集中的数据的。通过捕获Loss Trajectory的区别，可以实现效果更好的MIA. 因为小loss数据如果不在训练集中，一般是比较简单的图片，其训练loss会下降的很快，因此其loss轨迹在训练中期是低于在训练集中的数据的。通过捕获Loss Trajectory的区别，可以实现效果更好的MIA.
...@@ -102,7 +169,7 @@ ...@@ -102,7 +169,7 @@
Attack Model: Attack Model:
``` ```
epoch:90 train_loss:0.0045 test_loss:0.0050 train_prec1:0.7016 test_prec1:0.6216 val_prec1:0.6216 val_auc:0.6929 epoch:90 train_loss:0.0045 test_loss:0.0050 train_prec1:0.7016 test_prec1:0.6216 val_prec1:0.6216 val_auc:0.6929
Max AUC: 0.70787291 Max AUC: 0.70787291
Max ACC: 0.63575 Max ACC: 0.63575
TPR at 0.1% FPR: 1.2% TPR at 0.1% FPR: 1.2%
...@@ -161,7 +228,7 @@ ...@@ -161,7 +228,7 @@
Attack Model: Attack Model:
``` ```
val_prec1:0.6158 val_auc:0.6675 val_prec1:0.6158 val_auc:0.6675
Max AUC: 0.67598043 Max AUC: 0.67598043
Max ACC: 0.6204500000000001 Max ACC: 0.6204500000000001
TPR at 0.1% FPR: 1.2% TPR at 0.1% FPR: 1.2%
...@@ -220,7 +287,7 @@ ...@@ -220,7 +287,7 @@
Attack Model: Attack Model:
``` ```
epoch:90 train_loss:0.0046 test_loss:0.0048 train_prec1:0.7022 test_prec1:0.6547 val_prec1:0.6547 val_auc:0.7353 epoch:90 train_loss:0.0046 test_loss:0.0048 train_prec1:0.7022 test_prec1:0.6547 val_prec1:0.6547 val_auc:0.7353
Max AUC: 0.7494320150000001 Max AUC: 0.7494320150000001
Max ACC: 0.6719499999999999 Max ACC: 0.6719499999999999
TPR at 0.1% FPR: 2.2% TPR at 0.1% FPR: 2.2%
...@@ -279,7 +346,7 @@ ...@@ -279,7 +346,7 @@
Attack Model: Attack Model:
``` ```
epoch:90 train_loss:0.0044 test_loss:0.0051 train_prec1:0.7212 test_prec1:0.6013 val_prec1:0.6013 val_auc:0.6869 epoch:90 train_loss:0.0044 test_loss:0.0051 train_prec1:0.7212 test_prec1:0.6013 val_prec1:0.6013 val_auc:0.6869
Max AUC: 0.70338137 Max AUC: 0.70338137
Max ACC: 0.63095 Max ACC: 0.63095
TPR at 0.1% FPR: 2.6% TPR at 0.1% FPR: 2.6%
...@@ -338,7 +405,7 @@ ...@@ -338,7 +405,7 @@
Attack Model: Attack Model:
``` ```
epoch:90 train_loss:0.0036 test_loss:0.0038 train_prec1:0.8490 test_prec1:0.8244 val_prec1:0.8244 val_auc:0.9037 epoch:90 train_loss:0.0036 test_loss:0.0038 train_prec1:0.8490 test_prec1:0.8244 val_prec1:0.8244 val_auc:0.9037
Max AUC: 0.9115403099999998 Max AUC: 0.9115403099999998
Max ACC: 0.8295999999999999 Max ACC: 0.8295999999999999
TPR at 0.1% FPR: 0.0% TPR at 0.1% FPR: 0.0%
...@@ -457,7 +524,7 @@ ...@@ -457,7 +524,7 @@
Attack Model: Attack Model:
``` ```
val_prec1:0.8617 val_auc:0.9357 val_prec1:0.8617 val_auc:0.9357
Max AUC: 0.94010921 Max AUC: 0.94010921
Max ACC: 0.8649 Max ACC: 0.8649
TPR at 0.1% FPR: 7.1% TPR at 0.1% FPR: 7.1%
...@@ -692,7 +759,7 @@ ...@@ -692,7 +759,7 @@
Attack Model: Attack Model:
``` ```
epoch:90 train_loss:0.0042 test_loss:0.0044 train_prec1:0.7580 test_prec1:0.7336 val_prec1:0.7336 val_auc:0.8117 epoch:90 train_loss:0.0042 test_loss:0.0044 train_prec1:0.7580 test_prec1:0.7336 val_prec1:0.7336 val_auc:0.8117
Max AUC: 0.81900534 Max AUC: 0.81900534
Max ACC: 0.7394 Max ACC: 0.7394
TPR at 0.1% FPR: 3.0% TPR at 0.1% FPR: 3.0%
...@@ -773,8 +840,3 @@ Q2：如何进行预测？ ...@@ -773,8 +840,3 @@ Q2：如何进行预测？
A2：可以考虑将loss trajectory作为property，去计算in，out数据(member_status可以作为in, out的标记)的loss trajectory的相似度然后预测? 可以用Distill Target Model的Loss Trajectory与Attack Acc(AUC...)之类的配对构成数据点(也可以增加Distill Shadow Model的作为补充)。 A2：可以考虑将loss trajectory作为property，去计算in，out数据(member_status可以作为in, out的标记)的loss trajectory的相似度然后预测? 可以用Distill Target Model的Loss Trajectory与Attack Acc(AUC...)之类的配对构成数据点(也可以增加Distill Shadow Model的作为补充)。
预计的效果是相似度越低，则攻击成功率越高 预计的效果是相似度越低，则攻击成功率越高
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_152_10.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_152 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet152 --data cifar10"
python main.py --mode target --model resnet152 --data cifar10
echo "python main.py --mode shadow --model resnet152 --data cifar10"
python main.py --mode shadow --model resnet152 --data cifar10
echo "python main.py --mode distill_target --model resnet152 --data cifar10"
python main.py --mode distill_target --model resnet152 --data cifar10
echo "python main.py --mode distill_shadow --model resnet152 --data cifar10"
python main.py --mode distill_shadow --model resnet152 --data cifar10
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar10"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar10
echo "python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cifar10"
python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cifar10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_152_100.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_152_10 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet152 --data cifar100"
python main.py --mode target --model resnet152 --data cifar100
echo "python main.py --mode shadow --model resnet152 --data cifar100"
python main.py --mode shadow --model resnet152 --data cifar100
echo "python main.py --mode distill_target --model resnet152 --data cifar100"
python main.py --mode distill_target --model resnet152 --data cifar100
echo "python main.py --mode distill_shadow --model resnet152 --data cifar100"
python main.py --mode distill_shadow --model resnet152 --data cifar100
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar100"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar100
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar100"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cifar100
echo "python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cifar100"
python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cifar100
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_152_cinic.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_CIN_10 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 3-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-long # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet152 --data cinic10"
python main.py --mode target --model resnet152 --data cinic10
echo "python main.py --mode shadow --model resnet152 --data cinic10"
python main.py --mode shadow --model resnet152 --data cinic10
echo "python main.py --mode distill_target --model resnet152 --data cinic10"
python main.py --mode distill_target --model resnet152 --data cinic10
echo "python main.py --mode distill_shadow --model resnet152 --data cinic10"
python main.py --mode distill_shadow --model resnet152 --data cinic10
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cinic10"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cinic10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cinic10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet152 --model_distill resnet152 --data cinic10
echo "python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cinic10"
python main.py --action 1 --mia_type black-box --model resnet152 --model_distill resnet152 --data cinic10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10"
python main.py --mode distill_target --model resnet18 --data cifar10
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_100.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_C100_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar100"
# python main.py --mode target --model resnet18 --data cifar100
# echo "python main.py --mode shadow --model resnet18 --data cifar100"
# python main.py --mode shadow --model resnet18 --data cifar100
# echo "python main.py --mode distill_target --model resnet18 --data cifar100"
# python main.py --mode distill_target --model resnet18 --data cifar100
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar100"
# python main.py --mode distill_shadow --model resnet18 --data cifar100
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar100"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar100
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_100_load.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_C100_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar100"
# python main.py --mode target --model resnet18 --data cifar100
# echo "python main.py --mode shadow --model resnet18 --data cifar100"
# python main.py --mode shadow --model resnet18 --data cifar100
# echo "python main.py --mode distill_target --model resnet18 --data cifar100"
# python main.py --mode distill_target --model resnet18 --data cifar100
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar100"
# python main.py --mode distill_shadow --model resnet18 --data cifar100
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar100
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar100 --load_attack"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar100 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_load.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J L_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
# echo "python main.py --mode distill_target --model resnet18 --data cifar10"
# python main.py --mode distill_target --model resnet18 --data cifar10
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_FP_S1.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 3
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 3
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 3
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_FP_S2.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 4
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 4
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 4
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_FP_S3.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 5
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 5
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 5
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_FP_S4.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 6
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 6
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 6
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT4.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
# echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 4"
# python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 4
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 4"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 4
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT5.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
# echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 5"
# python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 5
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 5"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 5
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 5"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 5
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT6.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
# echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 6"
# python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 6
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 6"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 6
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 6"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 6
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT7.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 7"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 7
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 7"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 7
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 7"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 7
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT8.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 8"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 8"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 8
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT_S1.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 3
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 4
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 5"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 5
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 6"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 6
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 7"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 7
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# # CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 6
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 7"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 7
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 8"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 8
# ATTACK
# for test full precision mia result
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 5"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 6"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 6
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 7"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 7
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 8"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 8
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_INT_S2.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 9"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 9
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 10"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 11"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 11
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 12"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 12
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 13"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 13
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 14"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 14
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 15"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 15
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 16"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 16
# # CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 9"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 9
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 11"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 11
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 12"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 12
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 13"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 13
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 14"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 14
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 15"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 15
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 16"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type INT --num_bits 16
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 9"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 9
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 10"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 10
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 11"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 11
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 12"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 12
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 13"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 13
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 14"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 14
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 15"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 15
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 16"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type INT --num_bits 16
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 2"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 2
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 3
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 4
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 5"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 5
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 6"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 6
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 7"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 7
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 8"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 8
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 6
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 7"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 7
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 8"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --quant_type POT --num_bits 8
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 2"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 3"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 4"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 5"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 6"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 6
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 7"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 7
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 8"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack --quant_type POT --num_bits 8
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT3.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 3"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 3
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT4.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 4"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 4
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT5.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 5"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 5
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT6.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 6"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 6
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT7.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 7"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 7
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_10_ptq_POT8.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# echo "python main.py --mode target --model resnet18 --data cifar10"
# python main.py --mode target --model resnet18 --data cifar10
# echo "python main.py --mode shadow --model resnet18 --data cifar10"
# python main.py --mode shadow --model resnet18 --data cifar10
echo "python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 8"
python main.py --mode distill_target --model resnet18 --data cifar10 --store_ptq --quant_type POT --num_bits 8
# echo "python main.py --mode distill_shadow --model resnet18 --data cifar10"
# python main.py --mode distill_shadow --model resnet18 --data cifar10
# echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10"
# python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10
# echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
# python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack"
# python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cifar10 --load_attack
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_18_cinic.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_CIN_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-long # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet18 --data cinic10"
python main.py --mode target --model resnet18 --data cinic10
echo "python main.py --mode shadow --model resnet18 --data cinic10"
python main.py --mode shadow --model resnet18 --data cinic10
echo "python main.py --mode distill_target --model resnet18 --data cinic10"
python main.py --mode distill_target --model resnet18 --data cinic10
echo "python main.py --mode distill_shadow --model resnet18 --data cinic10"
python main.py --mode distill_shadow --model resnet18 --data cinic10
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cinic10"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cinic10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cinic10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet18 --model_distill resnet18 --data cinic10
echo "python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cinic10"
python main.py --action 1 --mia_type black-box --model resnet18 --model_distill resnet18 --data cinic10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_50 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet50 "
python main.py --mode target --model resnet50
echo "python main.py --mode shadow --model resnet50 "
python main.py --mode shadow --model resnet50
echo "python main.py --mode distill_target --model resnet50 "
python main.py --mode distill_target --model resnet50
echo "python main.py --mode distill_shadow --model resnet50 "
python main.py --mode distill_shadow --model resnet50
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50 "
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 "
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 "
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_100.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_50_10 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet50 --data cifar100"
python main.py --mode target --model resnet50 --data cifar100
echo "python main.py --mode shadow --model resnet50 --data cifar100"
python main.py --mode shadow --model resnet50 --data cifar100
echo "python main.py --mode distill_target --model resnet50 --data cifar100"
python main.py --mode distill_target --model resnet50 --data cifar100
echo "python main.py --mode distill_shadow --model resnet50 --data cifar100"
python main.py --mode distill_shadow --model resnet50 --data cifar100
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar100"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar100
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar100"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar100
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar100"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar100
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_FP_S1.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 5 --e_bits 3
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 5 --e_bits 3
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 3 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 3 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 4 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 5 --e_bits 3
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_FP_S2.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 6 --e_bits 4
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 6 --e_bits 4
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 6 --e_bits 4
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_FP_S3.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 7 --e_bits 5
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 7 --e_bits 5
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 5"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 7 --e_bits 5
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_FP_S4.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type FLOAT --num_bits 8 --e_bits 6
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type FLOAT --num_bits 8 --e_bits 6
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 1"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 1
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 4"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 5"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 6"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type FLOAT --num_bits 8 --e_bits 6
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_INT_S1.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 3
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 4"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 4
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 5"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 5
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 6"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 6
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 7"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 7
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 8"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 8
# # CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 6
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 7"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 7
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 8"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 8
# ATTACK
# for test full precision mia result
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 4"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 5"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 6"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 6
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 7"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 7
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 8"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 8
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_INT_S2.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 9"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 9
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 10"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 10
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 11"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 11
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 12"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 12
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 13"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 13
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 14"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 14
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 15"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 15
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 16"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type INT --num_bits 16
# # CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 9"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 9
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 11"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 11
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 12"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 12
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 13"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 13
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 14"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 14
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 15"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 15
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 16"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type INT --num_bits 16
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 9"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 9
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 10"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 10
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 11"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 11
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 12"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 12
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 13"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 13
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 14"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 14
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 15"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 15
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 16"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type INT --num_bits 16
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_10_ptq_POT.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J I_Tra_10_18 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-06:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
# TRAIN DISTILL MODEL
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 2"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 2
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 3"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 3
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 4"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 4
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 5"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 5
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 6"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 6
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 7"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 7
echo "python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 8"
python main.py --mode distill_target --model resnet50 --data cifar10 --store_ptq --quant_type POT --num_bits 8
# CONSTRUCT TEST DATASET
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 2"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 3"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 3
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 4"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 4
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 5"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 5
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 6"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 6
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 7"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 7
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 8"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cifar10 --quant_type POT --num_bits 8
# ATTACK
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 2"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 2
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 3"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 3
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 4"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 4
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 5"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 5
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 6"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 6
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 7"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 7
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 8"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cifar10 --load_attack --quant_type POT --num_bits 8
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_50_cinic.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_CIN_50 # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 1-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-long # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model resnet50 --data cinic10"
python main.py --mode target --model resnet50 --data cinic10
echo "python main.py --mode shadow --model resnet50 --data cinic10"
python main.py --mode shadow --model resnet50 --data cinic10
echo "python main.py --mode distill_target --model resnet50 --data cinic10"
python main.py --mode distill_target --model resnet50 --data cinic10
echo "python main.py --mode distill_shadow --model resnet50 --data cinic10"
python main.py --mode distill_shadow --model resnet50 --data cinic10
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cinic10"
python main.py --action 1 --mode shadow --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cinic10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cinic10"
python main.py --action 1 --mode target --mia_type build-dataset --model resnet50 --model_distill resnet50 --data cinic10
echo "python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cinic10"
python main.py --action 1 --mia_type black-box --model resnet50 --model_distill resnet50 --data cinic10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_mobile_10.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_Mobile # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model mobilenetv2"
python main.py --mode target --model mobilenetv2
echo "python main.py --mode shadow --model mobilenetv2"
python main.py --mode shadow --model mobilenetv2
echo "python main.py --mode distill_target --model mobilenetv2"
python main.py --mode distill_target --model mobilenetv2
echo "python main.py --mode distill_shadow --model mobilenetv2"
python main.py --mode distill_shadow --model mobilenetv2
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2 "
python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2
echo "python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2"
python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2
echo "python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2"
python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_mobile_100.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_100_Mobile # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model mobilenetv2 --data cifar100"
python main.py --mode target --model mobilenetv2 --data cifar100
echo "python main.py --mode shadow --model mobilenetv2 --data cifar100"
python main.py --mode shadow --model mobilenetv2 --data cifar100
echo "python main.py --mode distill_target --model mobilenetv2 --data cifar100"
python main.py --mode distill_target --model mobilenetv2 --data cifar100
echo "python main.py --mode distill_shadow --model mobilenetv2 --data cifar100"
python main.py --mode distill_shadow --model mobilenetv2 --data cifar100
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --data cifar100 --model_distill mobilenetv2"
python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2 --data cifar100
echo "python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --data cifar100 --model_distill mobilenetv2"
python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2 --data cifar100
echo "python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2 --data cifar100"
python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2 --data cifar100
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_attack_mobile_cinic.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J Tra_CIN_Mobile # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 3-00:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-long # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python main.py --mode target --model mobilenetv2 --data cinic10"
python main.py --mode target --model mobilenetv2 --data cinic10
echo "python main.py --mode shadow --model mobilenetv2 --data cinic10"
python main.py --mode shadow --model mobilenetv2 --data cinic10
echo "python main.py --mode distill_target --model mobilenetv2 --data cinic10"
python main.py --mode distill_target --model mobilenetv2 --data cinic10
echo "python main.py --mode distill_shadow --model mobilenetv2 --data cinic10"
python main.py --mode distill_shadow --model mobilenetv2 --data cinic10
echo "python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --data cinic10 --model_distill mobilenetv2"
python main.py --action 1 --mode shadow --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2 --data cinic10
echo "python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --data cinic10 --model_distill mobilenetv2"
python main.py --action 1 --mode target --mia_type build-dataset --model mobilenetv2 --model_distill mobilenetv2 --data cinic10
echo "python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2 --data cinic10"
python main.py --action 1 --mia_type black-box --model mobilenetv2 --model_distill mobilenetv2 --data cinic10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_div.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J JS_DIV # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python div.py --model resnet18 --data cifar10"
python div.py --model resnet18 --data cifar10
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/sh/train_plot.sh 0 → 100644
#!/bin/bash
#- Job parameters
# (TODO)
# Please modify job name
#SBATCH -J PLOT # The job name
#SBATCH -o ./info/ret-%j.out # Write the standard output to file named 'ret-<job_number>.out'
#SBATCH -e ./info/ret-%j.err # Write the standard error to file named 'ret-<job_number>.err'
#- Resources
# (TODO)
# Please modify your requirements
#SBATCH -p nv-gpu # Submit to 'nv-gpu' Partitiion
#SBATCH -t 0-12:00:00 # Run for a maximum time of 0 days, 12 hours, 00 mins, 00 secs
#SBATCH --nodes=1 # Request N nodes
#SBATCH --gres=gpu:1 # Request M GPU per node
#SBATCH --gres-flags=enforce-binding # CPU-GPU Affinity
#SBATCH --qos=gpu-normal # Request QOS Type
###
### The system will alloc 8 or 16 cores per gpu by default.
### If you need more or less, use following:
### #SBATCH --cpus-per-task=K # Request K cores
###
###
### Without specifying the constraint, any available nodes that meet the requirement will be allocated
### You can specify the characteristics of the compute nodes, and even the names of the compute nodes
###
### #SBATCH --nodelist=gpu-v00 # Request a specific list of hosts
### #SBATCH --constraint="Volta|RTX8000" # Request GPU Type: Volta(V100 or V100S) or RTX8000
###
#- Log information
echo "Job start at $(date "+%Y-%m-%d %H:%M:%S")"
echo "Job run at:"
echo "$(hostnamectl)"
#- Load environments
source /tools/module_env.sh
source ~/pyt1.5/bin/activate
module list # list modules loaded
##- Tools
module load cluster-tools/v1.0
module load slurm-tools/v1.0
module load cmake/3.15.7
module load git/2.17.1
module load vim/8.1.2424
##- language
module load python3/3.6.8
##- CUDA
module load cuda-cudnn/11.1-8.1.1
##- virtualenv
# source xxxxx/activate
echo $(module list) # list modules loaded
echo $(which gcc)
echo $(which python)
echo $(which python3)
cluster-quota # nas quota
nvidia-smi --format=csv --query-gpu=name,driver_version,power.limit # gpu info
#- Warning! Please not change your CUDA_VISIBLE_DEVICES
#- in `.bashrc`, `env.sh`, or your job script
echo "Use GPU ${CUDA_VISIBLE_DEVICES}" # which gpus
#- The CUDA_VISIBLE_DEVICES variable is assigned and specified by SLURM
#- Job step
# [EDIT HERE(TODO)]
sleep 2s
hostname
echo "python plot.py --model resnet50 --model_distill resnet50 --data cifar100"
python plot.py --model resnet50 --model_distill resnet50 --data cifar100
#- End
echo "Job end at $(date "+%Y-%m-%d %H:%M:%S")"
mzh/new_mzh/Loss_Trajectory_MIA/utils.py
...@@ -188,21 +188,32 @@ def cnn_train(args, model, data, epochs, optimizer, scheduler, model_params, mod ...@@ -188,21 +188,32 @@ def cnn_train(args, model, data, epochs, optimizer, scheduler, model_params, mod
# 做model distill的每一步具体训练 (在对trained Target/Shadow Model通过KL散度做distill) # 做model distill的每一步具体训练 (在对trained Target/Shadow Model通过KL散度做distill)
def cnn_training_step_dis(model, model_dis, optimizer, data, labels, device='cpu'): def cnn_training_step_dis(args,model, model_dis, optimizer, data, labels, device='cpu'):
b_x = data.to(device) b_x = data.to(device)
# 不会用label的 # 不会用label的
b_y_1 = labels.to(device) b_y_1 = labels.to(device)
output = model_dis(b_x) output = model_dis(b_x)
# distill model和 target/shadow model 做蒸馏 学习 loss对比的不再是标签，而是output (target/shadow model的output应该是采用的trained model的吧，optimizer不会对其参数更新，只会更新distill model的权值参数) # distill model和 target/shadow model 做蒸馏 学习 loss对比的不再是标签，而是output (target/shadow model的output应该是采用的trained model的吧，optimizer不会对其参数更新，只会更新distill model的权值参数)
b_y = model(b_x) if args.quant_type is None:
b_y = model(b_x)
else:
b_y = model.quantize_inference(b_x)
loss = nn.KLDivLoss(reduction='batchmean')(F.log_softmax(output, dim=1), F.softmax(b_y, dim=1)) loss = nn.KLDivLoss(reduction='batchmean')(F.log_softmax(output, dim=1), F.softmax(b_y, dim=1))
optimizer.zero_grad() optimizer.zero_grad()
loss.backward() loss.backward()
optimizer.step() optimizer.step()
# 对distill model的训练 # 对distill model的训练
def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, model_params, model_path, trained_model_name, device='cpu'): def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, model_params, model_path, trained_model_name, device='cpu'):
metrics = {'epoch_times':[], 'test_top1_acc':[], 'test_top5_acc':[], 'train_top1_acc':[], 'train_top5_acc':[], 'lrs':[]} metrics = {'epoch_times':[], 'test_top1_acc':[], 'test_top5_acc':[], 'train_top1_acc':[], 'train_top5_acc':[], 'lrs':[]}
if args.quant_type is not None:
if args.quant_type == 'FLOAT':
title = '%s_%d_E%d' % (args.quant_type, args.num_bits, args.e_bits)
else:
title = '%s_%d' % (args.quant_type, args.num_bits)
for epoch in range(1, epochs+1): for epoch in range(1, epochs+1):
...@@ -225,7 +236,7 @@ def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, mo ...@@ -225,7 +236,7 @@ def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, mo
print('Epoch: {}/{}'.format(epoch, epochs)) print('Epoch: {}/{}'.format(epoch, epochs))
print('Cur lr: {}'.format(cur_lr)) print('Cur lr: {}'.format(cur_lr))
for i, (x, y, idx) in enumerate(train_loader): for i, (x, y, idx) in enumerate(train_loader):
cnn_training_step_dis(model, model_dis, optimizer, x, y, device) cnn_training_step_dis(args,model, model_dis, optimizer, x, y, device)
end_time = time.time() end_time = time.time()
top1_test, top5_test = cnn_test(model_dis, test_loader, device) top1_test, top5_test = cnn_test(model_dis, test_loader, device)
...@@ -255,7 +266,13 @@ def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, mo ...@@ -255,7 +266,13 @@ def cnn_train_dis(args, model, model_dis, data, epochs, optimizer, scheduler, mo
total_training_time = sum(model_params['epoch_times']) total_training_time = sum(model_params['epoch_times'])
model_params['total_time'] = total_training_time model_params['total_time'] = total_training_time
print('Training took {} seconds...'.format(total_training_time)) print('Training took {} seconds...'.format(total_training_time))
save_model(model_dis, model_params, model_path, trained_model_name, epoch=epoch) # TODO 这里要改 需要把Full precision / 各种ptq 对应的Distill Target Model的各个epoch的数据存储到相应的文件夹中
if args.quant_type is None:
save_model(model_dis, model_params, model_path, trained_model_name, epoch=epoch)
else:
trained_model_name_ptq = trained_model_name + '_' +title
save_model(model_dis, model_params, model_path, trained_model_name_ptq, epoch=epoch)
return metrics return metrics
...@@ -296,3 +313,184 @@ def get_full_optimizer(model, lr_params, args): ...@@ -296,3 +313,184 @@ def get_full_optimizer(model, lr_params, args):
scheduler = CosineAnnealingLR(optimizer, args.epochs) scheduler = CosineAnnealingLR(optimizer, args.epochs)
return optimizer, scheduler return optimizer, scheduler
# origin my utils
def js_div(p_output, q_output, get_softmax=True):
"""
Function that measures JS divergence between target and output logits:
"""
KLDivLoss = nn.KLDivLoss(reduction='sum')
if get_softmax:
p_output = F.softmax(p_output)
q_output = F.softmax(q_output)
log_mean_output = ((p_output + q_output)/2).log()
return (KLDivLoss(log_mean_output, p_output) + KLDivLoss(log_mean_output, q_output))/2
def kl_div(p_output, q_output):
"""
Function that measures KL divergence between target and output logits:
"""
return F.kl_div(F.log_softmax(p_output, dim=1), F.softmax(q_output, dim=1))
def ebit_list(quant_type, num_bits):
if quant_type == 'FLOAT':
e_bit_list = list(range(1,num_bits-1))
else:
e_bit_list = [0]
return e_bit_list
def numbit_list(quant_type):
if quant_type == 'INT':
num_bit_list = list(range(2,17))
elif quant_type == 'POT':
num_bit_list = list(range(2,9))
else:
num_bit_list = list(range(2,9))
# num_bit_list = [8]
return num_bit_list
def build_bias_list(quant_type):
if quant_type == 'POT':
return build_pot_list(8) #
else:
return build_float_list(16,7)
def build_list(quant_type, num_bits, e_bits):
if quant_type == 'POT':
return build_pot_list(num_bits)
else:
return build_float_list(num_bits,e_bits)
def build_pot_list(num_bits):
plist = [0.]
for i in range(-2 ** (num_bits-1) + 2, 1):
# i最高到0，即pot量化最大值为1
plist.append(2. ** i)
plist.append(-2. ** i)
plist = torch.Tensor(list(set(plist)))
# plist = plist.mul(1.0 / torch.max(plist))
return plist
def build_float_list(num_bits,e_bits):
m_bits = num_bits - 1 - e_bits
plist = [0.]
# 相邻尾数的差值
dist_m = 2 ** (-m_bits)
e = -2 ** (e_bits - 1) + 1
for m in range(1, 2 ** m_bits):
frac = m * dist_m # 尾数部分
expo = 2 ** e # 指数部分
flt = frac * expo
plist.append(flt)
plist.append(-flt)
for e in range(-2 ** (e_bits - 1) + 2, 2 ** (e_bits - 1) + 1):
expo = 2 ** e
for m in range(0, 2 ** m_bits):
frac = 1. + m * dist_m
flt = frac * expo
plist.append(flt)
plist.append(-flt)
plist = torch.Tensor(list(set(plist)))
return plist
def fold_ratio(layer, par_ratio, flop_ratio):
idx = -1
for name in layer:
idx = idx + 1
# layer是for name, param in model.named_parameters()中提取出来的，一定是有downsample的
if 'bn' in name or 'sample.1' in name:
par_ratio[idx-1] += par_ratio[idx]
flop_ratio[idx-1] += flop_ratio[idx]
return par_ratio,flop_ratio
def fold_model(model):
idx = -1
module_list = []
# print('fold model:')
for name, module in model.named_modules():
# print(name+'-- +')
idx += 1
module_list.append(module)
# 这里之前忘记考虑downsampl里的conv了，导致少融合了一些
if 'bn' in name or 'sample.1' in name:
# print(name+'-- *')
module_list[idx-1] = fold_bn(module_list[idx-1],module) # 在这里修改了
return model
# def fold_model(model):
# last_conv = None
# last_bn = None
# for name, module in model.named_modules():
# if isinstance(module, nn.Conv2d):
# # 如果当前模块是卷积层，则将其 "fold" 到上一个 BN 层中
# if last_bn is not None:
# last_conv = fold_bn(last_conv, last_bn)
# last_bn = None
# last_conv = module
# elif isinstance(module, nn.BatchNorm2d):
# # 如果当前模块是 BN 层，则将其 "fold" 到上一个卷积层中
# last_bn = module
# if last_conv is not None:
# last_conv = fold_bn(last_conv, last_bn)
# last_bn = None
# # 处理最后一个 BN 层
# if last_bn is not None:
# last_conv = fold_bn(last_conv, last_bn)
# return model
def fold_bn(conv, bn):
# 获取 BN 层的参数
gamma = bn.weight.data
beta = bn.bias.data
mean = bn.running_mean
var = bn.running_var
eps = bn.eps
std = torch.sqrt(var + eps)
feat = bn.num_features
# 获取卷积层的参数
weight = conv.weight.data
if conv.bias is not None:
bias = conv.bias.data
if bn.affine:
gamma_ = gamma / std
weight = weight * gamma_.view(feat, 1, 1, 1)
if conv.bias is not None:
bias = gamma_ * bias - gamma_ * mean + beta
else:
bias = beta - gamma_ * mean
else:
gamma_ = 1 / std
weight = weight * gamma_
if conv.bias is not None:
bias = gamma_ * bias - gamma_ * mean
else:
bias = -gamma_ * mean
# 设置新的 weight 和 bias
conv.weight.data = weight
# 适用于bias=none的
if conv.bias is None:
conv.bias = nn.Parameter(bias)
else:
conv.bias.data = bias
return conv
\ No newline at end of file
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment