tune_nnvm_mobile_gpu.py 14.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
"""
Auto-tuning a convolutional network for Mobile GPU
====================================================
**Author**: `Lianmin Zheng <https://https://github.com/merrymercy>`_

Auto-tuning for a specific device is critical for getting the best
performance. This is a tutorial about how to tune a whole convolutional
network.

The operator implementation for Mobile GPU in TVM is written in template form.
The template has many tunable knobs (tile factor, vectorization, unrolling, etc).
We will tune all convolution, depthwise convolution and dense operators
in the neural network. After tuning, we produce a log file which stores
the best knob values for all required operators. When the tvm compiler compiles
these operators, it will query this log file to get the best knob values.

We also released pre-tuned parameters for some arm devices. You can go to
`Mobile GPU Benchmark <https://github.com/dmlc/tvm/wiki/Benchmark#mobile-gpu>`_
to see the results.
"""

######################################################################
# Install dependencies
# --------------------
# To use the autotvm package in tvm, we need to install some extra dependencies.
# (change "3" to "2" if you use python2):
#
# .. code-block:: bash
#
#   pip3 install --user psutil xgboost tornado
#
# To make tvm run faster during tuning, it is recommended to use cython
# as FFI of tvm. In the root directory of tvm, execute
# (change "3" to "2" if you use python2):
#
# .. code-block:: bash
#
#   pip3 install --user cython
#   sudo make cython3
#
# Now return to python code. Import packages.

import os

import numpy as np

import nnvm.testing
import nnvm.compiler
import tvm
from tvm import autotvm
from tvm.autotvm.tuner import XGBTuner, GATuner, RandomTuner, GridSearchTuner
from tvm.contrib.util import tempdir
import tvm.contrib.graph_runtime as runtime

#################################################################
# Define network
# --------------
# First we need to define the network in nnvm symbol API.
# We can load some pre-defined network from :code:`nnvm.testing`.
# We can also load models from MXNet, ONNX and TensorFlow (see NNVM
# tutorials :ref:`tutorial-nnvm` for more details).

def get_network(name, batch_size):
    """Get the symbol definition and random weight of a network"""
    input_shape = (batch_size, 3, 224, 224)
    output_shape = (batch_size, 1000)

    if "resnet" in name:
        n_layer = int(name.split('-')[1])
        net, params = nnvm.testing.resnet.get_workload(num_layers=n_layer, batch_size=batch_size)
    elif "vgg" in name:
        n_layer = int(name.split('-')[1])
        net, params = nnvm.testing.vgg.get_workload(num_layers=n_layer, batch_size=batch_size)
    elif name == 'mobilenet':
        net, params = nnvm.testing.mobilenet.get_workload(batch_size=batch_size)
    elif name == 'squeezenet_v1.1':
        net, params = nnvm.testing.squeezenet.get_workload(batch_size=batch_size, version='1.1')
    elif name == 'inception_v3':
        input_shape = (1, 3, 299, 299)
        net, params = nnvm.testing.inception_v3.get_workload(batch_size=batch_size)
    elif name == 'custom':
        # an example for custom network
        from nnvm.testing import utils
        net = nnvm.sym.Variable('data')
        net = nnvm.sym.conv2d(net, channels=4, kernel_size=(3,3), padding=(1,1))
        net = nnvm.sym.flatten(net)
        net = nnvm.sym.dense(net, units=1000)
        net, params = utils.create_workload(net, batch_size, (3, 224, 224))
    elif name == 'mxnet':
        # an example for mxnet model
        from mxnet.gluon.model_zoo.vision import get_model
        block = get_model('resnet18_v1', pretrained=True)
        net, params = nnvm.frontend.from_mxnet(block)
        net = nnvm.sym.softmax(net)
    else:
        raise ValueError("Unsupported network: " + name)

    return net, params, input_shape, output_shape


#################################################################
# Start RPC Tracker
# -----------------
# TVM uses RPC session to communicate with ARM boards.
# During tuning, the tuner will send the generated code to the board and
# measure the speed of code on the board.
#
# To scale up the tuning, TVM uses RPC Tracker to manage distributed devices.
# The RPC Tracker is a centralized master node. We can register all devices to
# the tracker. For example, if we have 10 phones, we can register all of them
# to the tracker, and run 10 measurements in parallel, accelerating the tuning process.
#
# To start an RPC tracker, run this command on the host machine. The tracker is
# required during the whole tuning process, so we need to open a new terminal for
# this command:
#
# .. code-block:: bash
#
#   python -m tvm.exec.rpc_tracker --host=0.0.0.0 --port=9190
#
# The expected output is
#
# .. code-block:: bash
#
#   INFO:RPCTracker:bind to 0.0.0.0:9190

#################################################################
# Register devices to RPC Tracker
# -----------------------------------
# Now we can register our devices to the tracker. The first step is to
# build tvm runtime for the ARM devices.
#
# * For Linux:
#   Follow this section :ref:`build-tvm-runtime-on-device` to build
#   tvm runtime on the device. Then register the device to tracker by
#
#   .. code-block:: bash
#
#     python -m tvm.exec.rpc_server --tracker=[HOST_IP]:9190 --key=rk3399
#
#   (replace :code:`[HOST_IP]` with the IP address of your host machine)
#
# * For Android:
#   Follow this `readme page <https://github.com/dmlc/tvm/tree/master/apps/android_rpc>`_ to
#   install tvm rpc apk on the android device. Make sure you can pass the android rpc test.
#   Then you have already registred your device. During tuning, you have to go to developer option
#   and enable "Keep screen awake during changing" and charge your phone to make it stable.
#
# After registering devices, we can confirm it by querying rpc_tracker
#
# .. code-block:: bash
#
#   python -m tvm.exec.query_rpc_tracker --host=0.0.0.0 --port=9190
#
# For example, if we have 2 Huawei mate10 pro, 11 Raspberry Pi 3B and 2 rk3399,
# the output can be
#
# .. code-block:: bash
#
#    Queue Status
#    ----------------------------------
#    key          total  free  pending
#    ----------------------------------
#    mate10pro    2      2     0
#    rk3399       2      2     0
#    rpi3b        11     11    0
#    ----------------------------------
#
# You can register multiple devices to the tracker to accelerate the measurement in tuning.

###########################################
# Set Tuning Options
# ------------------
# Before tuning, we should apply some configurations. Here I use an RK3399 board
# as example. In your setting, you should modify the target and device_key accordingly.
# set :code:`use_android` to True if you use android phone.

#### DEVICE CONFIG ####

target = tvm.target.create('opencl -device=mali')

# Replace "aarch64-linux-gnu" with the correct target of your board.
# This target host is used for cross compilation. You can query it by :code:`gcc -v` on your device.
target_host = 'llvm -target=aarch64-linux-gnu'

# Also replace this with the device key in your tracker
device_key = 'rk3399'

# Set this to True if you use android phone
use_android = False

#### TUNING OPTION ####
network = 'resnet-18'
log_file = "%s.%s.log" % (device_key, network)
dtype = 'float32'

tuning_option = {
    'log_filename': log_file,

    'tuner': 'xgb',
    'n_trial': 1000,
    'early_stopping': 450,

    'measure_option': autotvm.measure_option(
        builder=autotvm.LocalBuilder(
            build_func='ndk' if use_android else 'default'),
        runner=autotvm.RPCRunner(
            device_key, host='localhost', port=9190,
            number=10,
            timeout=5,
        ),
    ),
}

####################################################################
#
# .. note:: How to set tuning options
#
#   In general, the default values provided here work well.
#   If you have enough time budget, you can set :code:`n_trial`, :code:`early_stopping` larger,
#   which makes the tuning run longer.
#   If your device runs very slow or your conv2d operators have many GFLOPs, considering to
#   set timeout larger.
#

###################################################################
# Begin Tuning
# ------------
# Now we can extract tuning tasks from the network and begin tuning.
# Here, we provide a simple utility function to tune a list of tasks.
# This function is just an initial implementation which tunes them in sequential order.
# We will introduce a more sophisticated tuning scheduler in the future.

# You can skip the implementation of this function for this tutorial.
def tune_tasks(tasks,
               measure_option,
               tuner='xgb',
               n_trial=1000,
               early_stopping=None,
               log_filename='tuning.log',
               use_transfer_learning=True,
               try_winograd=True):
    if try_winograd:
        for i in range(len(tasks)):
            try:  # try winograd template
                tsk = autotvm.task.create(tasks[i].name, tasks[i].args,
                                          tasks[i].target, tasks[i].target_host, 'winograd')
                tasks.append(tsk)
            except Exception:
                pass

    # create tmp log file
    tmp_log_file = log_filename + ".tmp"
    if os.path.exists(tmp_log_file):
        os.remove(tmp_log_file)

    for i, tsk in enumerate(reversed(tasks)):
        prefix = "[Task %2d/%2d] " % (i+1, len(tasks))

        # create tuner
        if tuner == 'xgb' or tuner == 'xgb-rank':
            tuner_obj = XGBTuner(tsk, loss_type='rank')
        elif tuner == 'ga':
            tuner_obj = GATuner(tsk, pop_size=50)
        elif tuner == 'random':
            tuner_obj = RandomTuner(tsk)
        elif tuner == 'gridsearch':
            tuner_obj = GridSearchTuner(tsk)
        else:
            raise ValueError("Invalid tuner: " + tuner)

        if use_transfer_learning:
            if os.path.isfile(tmp_log_file):
                tuner_obj.load_history(autotvm.record.load_from_file(tmp_log_file))

        # do tuning
        tuner_obj.tune(n_trial=min(n_trial, len(tsk.config_space)),
                       early_stopping=early_stopping,
                       measure_option=measure_option,
                       callbacks=[
                           autotvm.callback.progress_bar(n_trial, prefix=prefix),
                           autotvm.callback.log_to_file(tmp_log_file)])

    # pick best records to a cache file
    autotvm.record.pick_best(tmp_log_file, log_filename)
    os.remove(tmp_log_file)


########################################################################
# Finally, we launch tuning jobs and evaluate the end-to-end performance.

def tune_and_evaluate(tuning_opt):
    # extract workloads from nnvm graph
    print("Extract tasks...")
    net, params, input_shape, out_shape = get_network(network, batch_size=1)
    tasks = autotvm.task.extract_from_graph(net, target=target, target_host=target_host,
                                            shape={'data': input_shape}, dtype=dtype,
                                            symbols=(nnvm.sym.conv2d, nnvm.sym.dense))

    # run tuning tasks
    print("Tuning...")
    tune_tasks(tasks, **tuning_opt)

    # compile kernels with history best records
    with autotvm.apply_history_best(log_file):
        print("Compile...")
        with nnvm.compiler.build_config(opt_level=3):
            graph, lib, params = nnvm.compiler.build(
                net, target=target, target_host=target_host,
                shape={'data': input_shape}, params=params, dtype=dtype)

        # export library
        tmp = tempdir()
        if use_android:
            from tvm.contrib import ndk
            filename = "net.so"
            lib.export_library(tmp.relpath(filename), ndk.create_shared)
        else:
            filename = "net.tar"
            lib.export_library(tmp.relpath(filename))

        # upload module to device
        print("Upload...")
        remote = autotvm.measure.request_remote(device_key, 'localhost', 9190,
                                                timeout=10000)
        remote.upload(tmp.relpath(filename))
        rlib = remote.load_module(filename)

        # upload parameters to device
        ctx = remote.context(str(target), 0)
        rparams = {k: tvm.nd.array(v, ctx) for k, v in params.items()}
        data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))
        module = runtime.create(graph, rlib, ctx)
        module.set_input('data', data_tvm)
        module.set_input(**rparams)

        # evaluate
        print("Evaluate inference time cost...")
        ftimer = module.module.time_evaluator("run", ctx, number=50, repeat=3)
        prof_res = np.array(ftimer().results) * 1000  # convert to millisecond
        print("Mean inference time (std dev): %.2f ms (%.2f ms)" %
              (np.mean(prof_res), np.std(prof_res)))

# We do not run the tuning in our webpage server since it takes too long.
# Uncomment the following line to run it by yourself.

# tune_and_evaluate(tuning_option)

######################################################################
# Sample Output
# -------------
# The tuning needs to compile many programs and extract feature from them.
# So a high performance CPU is recommended.
# One sample output is listed below. It takes about 3 hours on a 32T AMD Ryzen Threadripper.
#
# .. code-block:: bash
#
#    Extract tasks...
#    Tuning...
#    [Task  1/17]  Current/Best:   12.22/  36.05 GFLOPS | Progress: (32/1000) | 42.12 s
#
#    (The following part is running, will update it later).

######################################################################
#
# .. note:: **Experiencing Difficulties?**
#
#   The auto tuning module is error-prone. If you always see " 0.00/ 0.00 GFLOPS",
#   then there must be something wrong.
#
#   First, make sure you set the correct configuration of your device.
#   Then, you can print debug information by adding these lines in the beginning
#   of the script. It will print every measurement result, where you can find useful
#   error messages.
#
#   .. code-block:: python
#
#      import logging
#      logging.getLogger('autotvm').setLevel(logging.DEBUG)
#
#   Finally, always feel free to ask our community for help on https://discuss.tvm.ai