Skip to content

M
muzero
Commit ceb98db4 by Werner Duvaud
Options

Add console log

parent 2a7838e4
Showing with 127 additions and 94 deletions
README.md
......@@ -7,26 +7,26 @@
# MuZero General
A flexible, commented and [documented](https://github.com/werner-duvaud/muzero-general/wiki/MuZero-Documentation) implementation of MuZero based on the Google DeepMind [paper](https://arxiv.org/abs/1911.08265) and the associated [pseudocode](https://arxiv.org/src/1911.08265v1/anc/pseudocode.py).
It is designed to be easily adaptable for every games or reinforcement learning environments (like [gym](https://github.com/openai/gym)). You only need to edit the game file with the parameters and the game class. Please refer to the documentation and the [example](https://github.com/werner-duvaud/muzero-general/blob/master/games/cartpole.py).
It is designed to be easily adaptable for every games or reinforcement learning environments (like [gym](https://github.com/openai/gym)). You only need to edit the [game file](https://github.com/werner-duvaud/muzero-general/tree/master/games) with the parameters and the game class. Please refer to the documentation and the [example](https://github.com/werner-duvaud/muzero-general/blob/master/games/cartpole.py).
MuZero is a model based reinforcement learning algorithm, successor of AlphaZero. It learns to master games without knowing the rules. It only knows actions and then learn to play and master the game. It is at least more efficient than similar algorithms like [AlphaZero](https://arxiv.org/abs/1712.01815), [SimPLe](https://arxiv.org/abs/1903.00374) and [World Models](https://arxiv.org/abs/1803.10122).
It uses [PyTorch](https://github.com/pytorch/pytorch) and [Ray](https://github.com/ray-project/ray) for running the different components simultaneously. There is a complete GPU support.
There are four "actors" which are classes that run simultaneously in a dedicated thread.
The shared storage holds the latest neural network weights, the self-play uses those weights to generate self-play games and store them in the replay buffer. Finally, those games are used to train a network and store the weights in the shared storage. The circle is complete.
Those components are launched and managed from the MuZero class in muzero.py and the structure of the neural network is defined in models.py.
There are four components which are classes that run simultaneously in a dedicated thread.
The `shared storage` holds the latest neural network weights, the `self-play` uses those weights to generate self-play games and store them in the `replay buffer`. Finally, those played games are used to `train` a network and store the weights in the shared storage. The circle is complete. See [How it works](https://github.com/werner-duvaud/muzero-general/wiki/How-MuZero-works)
All performances are tracked and displayed in real time in tensorboard.
Those components are launched and managed from the MuZero class in `muzero.py` and the structure of the neural network is defined in `models.py`.
![lunarlander training preview](https://github.com/werner-duvaud/muzero-general/blob/master/pretrained/cartpole_training_summary.png)
All performances are tracked and displayed in real time in tensorboard.
![lunarlander training preview](https://github.com/werner-duvaud/muzero-general/blob/master/docs/cartpole_training_summary.png)
## Games already implemented with pretrained network available
* Lunar Lander
* Cartpole
![lunarlander training preview](https://github.com/werner-duvaud/muzero-general/blob/master/games/lunarlander_training_preview.png)
![lunarlander training preview](https://github.com/werner-duvaud/muzero-general/blob/master/docs/lunarlander_training_preview.png)
## Getting started
### Installation
......
docs/README.md 0 → 100644
# MuZero General Documentation
Please refer to the [GitHub wiki](https://github.com/werner-duvaud/muzero-general/wiki/MuZero-Documentation) and to the comments in the code.
\ No newline at end of file
games/cartpole.py
......@@ -15,6 +15,7 @@ class MuZeroConfig:
self.max_moves = 500 # Maximum number of moves if game is not finished before
self.num_simulations = 50 # Number of futur moves self-simulated
self.discount = 0.997 # Chronological discount of the reward
self.self_play_delay = None # Number of seconds to wait after each played game to adjust the self play / training ratio to avoid overfitting (Recommended is 13:1 see https://arxiv.org/abs/1902.04522 Appendix A)
# Root prior exploration noise
self.root_dirichlet_alpha = 0.25
......@@ -24,30 +25,28 @@ class MuZeroConfig:
self.pb_c_base = 19652
self.pb_c_init = 1.25
# If we already have some information about which values occur in the environment, we can use them to initialize the rescaling
# This is not strictly necessary, but establishes identical behaviour to AlphaZero in board games
self.min_known_bound = None
self.max_known_bound = None
### Network
self.encoding_size = 64
self.hidden_size = 32
# Training
self.results_path = "./pretrained" # Path to store the model weights
self.training_steps = 1000 # Total number of training steps (ie weights update according to a batch)
self.training_steps = 2000 # Total number of training steps (ie weights update according to a batch)
self.batch_size = 128 # Number of parts of games to train on at each training step
self.num_unroll_steps = 5 # Number of game moves to keep for every batch element
self.test_episodes = 2 # Number of game played to evaluate the network
self.checkpoint_interval = 10 # Number of training steps before using the model for sef-playing
self.window_size = 1000 # Number of self-play games to keep in the replay buffer
self.td_steps = 10 # Number of steps in the futur to take into account for calculating the target value
self.training_delay = 1 # Number of seconds to wait after each training to adjust the self play / training ratio to avoid overfitting (Recommended is 13:1 see https://arxiv.org/abs/1902.04522 Appendix A)
self.weight_decay = 1e-4 # L2 weights regularization
self.momentum = 0.9
# Test
self.test_episodes = 2 # Number of game played to evaluate the network
# Exponential learning rate schedule
self.lr_init = 0.005 # Initial learning rate
self.lr_init = 0.0005 # Initial learning rate
self.lr_decay_rate = 0.1
self.lr_decay_steps = 3500
......@@ -59,7 +58,7 @@ class MuZeroConfig:
Returns:
Positive float.
"""
if trained_steps < 0.5 * self.training_steps:
if trained_steps < 0.25 * self.training_steps:
return 1.0
elif trained_steps < 0.75 * self.training_steps:
return 0.5
......
games/lunarlander.py
......@@ -16,6 +16,7 @@ class MuZeroConfig:
self.max_moves = 500 # Maximum number of moves if game is not finished before
self.num_simulations = 50 # Number of futur moves self-simulated
self.discount = 0.997 # Chronological discount of the reward
self.self_play_delay = None # Number of seconds to wait after each played game to adjust the self play / training ratio to avoid overfitting (Recommended is 13:1 see https://arxiv.org/abs/1902.04522 Appendix A)
# Root prior exploration noise
self.root_dirichlet_alpha = 0.25
......@@ -25,30 +26,28 @@ class MuZeroConfig:
self.pb_c_base = 19652
self.pb_c_init = 1.25
# If we already have some information about which values occur in the environment, we can use them to initialize the rescaling
# This is not strictly necessary, but establishes identical behaviour to AlphaZero in board games
self.min_known_bound = None
self.max_known_bound = None
### Network
self.encoding_size = 64
self.hidden_size = 32
# Training
self.results_path = "./pretrained" # Path to store the model weights
self.training_steps = 20000 # Total number of training steps (ie weights update according to a batch)
self.training_steps = 2000 # Total number of training steps (ie weights update according to a batch)
self.batch_size = 128 # Number of parts of games to train on at each training step
self.num_unroll_steps = 5 # Number of game moves to keep for every batch element
self.test_episodes = 2 # Number of game played to evaluate the network
self.checkpoint_interval = 20 # Number of training steps before using the model for sef-playing
self.checkpoint_interval = 10 # Number of training steps before using the model for sef-playing
self.window_size = 1000 # Number of self-play games to keep in the replay buffer
self.td_steps = 100 # Number of steps in the futur to take into account for calculating the target value
self.td_steps = 10 # Number of steps in the futur to take into account for calculating the target value
self.training_delay = 8 # Number of seconds to wait after each training to adjust the self play / training ratio to avoid overfitting (Recommended is 13:1 see https://arxiv.org/abs/1902.04522 Appendix A)
self.weight_decay = 1e-4 # L2 weights regularization
self.momentum = 0.9
# Test
self.test_episodes = 2 # Number of game played to evaluate the network
# Exponential learning rate schedule
self.lr_init = 0.005 # Initial learning rate
self.lr_init = 0.0001 # Initial learning rate
self.lr_decay_rate = 0.1
self.lr_decay_steps = 3500
......@@ -60,7 +59,7 @@ class MuZeroConfig:
Returns:
Positive float.
"""
if trained_steps < 0.5 * self.training_steps:
if trained_steps < 0.25 * self.training_steps:
return 1.0
elif trained_steps < 0.75 * self.training_steps:
return 0.5
......
models.py
import torch
class FullyConnectedNetwork(torch.nn.Module):
def __init__(
self, input_size, layers_sizes, output_size, activation=torch.nn.Tanh()
......@@ -25,6 +26,7 @@ class FullyConnectedNetwork(torch.nn.Module):
x = layer(x)
return x
# TODO: unified residual network
class MuZeroNetwork(torch.nn.Module):
def __init__(self, observation_size, action_space_size, encoding_size, hidden_size):
......
muzero.py
import copy
import datetime
import importlib
import os
import time
......@@ -21,7 +20,8 @@ class MuZero:
Main class to manage MuZero.
Args:
game_name (str): Name of the game module, it should match the name of a .py file in the "./games" directory.
game_name (str): Name of the game module, it should match the name of a .py file
in the "./games" directory.
Example:
>>> muzero = MuZero("cartpole")
......@@ -46,18 +46,16 @@ class MuZero:
raise err
# Fix random generator seed for reproductibility
# TODO: check if results do not change from one run to another
numpy.random.seed(self.config.seed)
torch.manual_seed(self.config.seed)
# Used to initialize components when continuing a former training
# Initial weights used to initialize components
self.muzero_weights = models.MuZeroNetwork(
self.config.observation_shape,
len(self.config.action_space),
self.config.encoding_size,
self.config.hidden_size,
).get_weights()
self.training_steps = 0
def train(self):
ray.init()
......@@ -68,16 +66,12 @@ class MuZero:
# Initialize workers
training_worker = trainer.Trainer.remote(
copy.deepcopy(self.muzero_weights),
self.training_steps,
self.config,
# Train on GPU if available
"cuda" if torch.cuda.is_available() else "cpu",
)
shared_storage_worker = shared_storage.SharedStorage.remote(
copy.deepcopy(self.muzero_weights),
self.training_steps,
self.game_name,
self.config,
copy.deepcopy(self.muzero_weights), self.game_name, self.config,
)
replay_buffer_worker = replay_buffer.ReplayBuffer.remote(self.config)
self_play_workers = [
......@@ -107,7 +101,7 @@ class MuZero:
# Loop for monitoring in real time the workers
print(
"Run tensorboard --logdir ./ and go to http://localhost:6006/ to track the training performance"
"\nTraining...\nRun tensorboard --logdir ./ and go to http://localhost:6006/ to see in real time the training performance.\n"
)
counter = 0
infos = ray.get(shared_storage_worker.get_infos.remote())
......@@ -126,15 +120,21 @@ class MuZero:
"2.Workers/Training steps", infos["training_step"], counter
)
writer.add_scalar("3.Loss/1.Total loss", infos["total_loss"], counter)
writer.add_scalar("3.Loss details/Value loss", infos["value_loss"], counter)
writer.add_scalar(
"3.Loss details/Reward loss", infos["reward_loss"], counter
)
writer.add_scalar(
"3.Loss details/Policy loss", infos["policy_loss"], counter
writer.add_scalar("3.Loss/Value loss", infos["value_loss"], counter)
writer.add_scalar("3.Loss/Reward loss", infos["reward_loss"], counter)
writer.add_scalar("3.Loss/Policy loss", infos["policy_loss"], counter)
print(
"Last test reward: {0:.2f}. Training step: {1}/{2}. Played games: {3}. Loss: {4:.2f}".format(
infos["total_reward"],
infos["training_step"],
self.config.training_steps,
ray.get(replay_buffer_worker.get_self_play_count.remote()),
infos["total_loss"],
),
end="\r",
)
counter += 1
time.sleep(1)
time.sleep(3)
self.muzero_weights = ray.get(shared_storage_worker.get_weights.remote())
ray.shutdown()
......@@ -142,6 +142,7 @@ class MuZero:
"""
Test the model in a dedicated thread.
"""
print("Testing...")
ray.init()
self_play_workers = self_play.SelfPlay.remote(
copy.deepcopy(self.muzero_weights), self.Game(), self.config, "cpu",
......@@ -149,18 +150,17 @@ class MuZero:
test_rewards = []
with torch.no_grad():
for _ in range(self.config.test_episodes):
history = ray.get(self_play_workers.self_play.remote(0, render))
history = ray.get(self_play_workers.play_game.remote(0, render))
test_rewards.append(sum(history.rewards))
ray.shutdown()
return test_rewards
def load_model(self, path=None, training_step=0):
# TODO: why pretrained model is degradated during the new train
def load_model(self, path=None):
if not path:
path = os.path.join(self.config.results_path, self.game_name)
try:
self.muzero_weights = torch.load(path)
self.training_step = training_step
print("Using weights from {}".format(path))
except FileNotFoundError:
print("There is no model saved in {}.".format(path))
......@@ -168,5 +168,6 @@ class MuZero:
if __name__ == "__main__":
muzero = MuZero("cartpole")
muzero.train()
# muzero.load_model()
muzero.load_model()
muzero.test()
notebook.ipynb
{
"nbformat": 4,
"nbformat_minor": 2,
"nbformat_minor": 0,
"metadata": {
"language_info": {
"name": "python",
"codemirror_mode": {
"name": "ipython",
"version": 3
}
"colab": {
"name": "Untitled3.ipynb",
"provenance": []
},
"orig_nbformat": 2,
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"npconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": 3
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
}
},
"cells": [
{
......@@ -24,13 +18,22 @@
"metadata": {},
"outputs": [],
"source": [
"# Google colab imports\n",
"# Google colab stuffs\n",
"!pip install -r requirements.txt\n",
"!pip uninstall -y pyarrow\n",
"# If you have an import issue with ray, restart the environment (execution menu)\n",
"\n",
"%load_ext tensorboard\n",
"# If you have an import issue with ray in google colab, restart the environment (execution menu)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# You must have the repository imported along with your notebook. \n",
"# For google colab, click on \">\" buttton (left) and import files (muzero.py, self_play.py, ...).\n",
"\n",
"import muzero as mz"
]
},
......@@ -42,7 +45,6 @@
"source": [
"#Train on cartpole game\n",
"muzero = mz.MuZero(\"cartpole\")\n",
"muzero.load_model()\n",
"muzero.train()\n",
"muzero.test()"
]
......
pretrained/cartpole
No preview for this file type
pretrained/lunarlander
No preview for this file type
replay_buffer.py
......@@ -62,7 +62,8 @@ def sample_game(buffer):
"""
Sample game from buffer either uniformly or according to some priority.
"""
# TODO: sample with probability link to the highest difference between real and predicted value (see paper appendix Training)
# TODO: sample with probability link to the highest difference between real and
# predicted value (see paper appendix Training)
return numpy.random.choice(buffer)
......@@ -76,7 +77,8 @@ def sample_position(game_history):
def make_target(game_history, state_index, num_unroll_steps, td_steps):
"""
The value target is the discounted root value of the search tree td_steps into the future, plus the discounted sum of all rewards until then.
The value target is the discounted root value of the search tree td_steps into the
future, plus the discounted sum of all rewards until then.
"""
target_values, target_rewards, target_policies = [], [], []
for current_index in range(state_index, state_index + num_unroll_steps + 1):
......
self_play.py
import math
import time
import copy
import numpy
import ray
import torch
......@@ -12,6 +13,7 @@ class SelfPlay:
"""
Class which run in a dedicated thread to play games and save them to the replay-buffer.
"""
def __init__(self, initial_weights, game, config, device):
self.config = config
self.game = game
......@@ -25,11 +27,14 @@ class SelfPlay:
)
self.model.set_weights(initial_weights)
self.model.to(torch.device(device))
self.model.eval()
def continuous_self_play(self, shared_storage, replay_buffer, test_mode=False):
with torch.no_grad():
while True:
self.model.set_weights(ray.get(shared_storage.get_weights.remote()))
self.model.set_weights(
copy.deepcopy(ray.get(shared_storage.get_weights.remote()))
)
# Take the best action (no exploration) in test mode
temperature = (
......@@ -41,8 +46,7 @@ class SelfPlay:
]
)
)
game_history = self.self_play(temperature, False)
game_history = self.play_game(temperature, False)
# Save to the shared storage
if test_mode:
......@@ -52,7 +56,10 @@ class SelfPlay:
if not test_mode:
replay_buffer.save_game.remote(game_history)
def self_play(self, temperature, render):
if not test_mode and self.config.self_play_delay:
time.sleep(self.config.self_play_delay)
def play_game(self, temperature, render):
"""
Play one game with actions based on the Monte Carlo tree search at each moves.
"""
......@@ -83,7 +90,8 @@ class SelfPlay:
def select_action(node, temperature):
"""
Select action according to the vivist count distribution and the temperature.
The temperature is changed dynamically with the visit_softmax_temperature function in the config.
The temperature is changed dynamically with the visit_softmax_temperature function
in the config.
"""
visit_counts = numpy.array(
[[child.visit_count, action] for action, child in node.children.items()]
......@@ -120,8 +128,10 @@ class MCTS:
def run(self, model, observation, add_exploration_noise):
"""
At the root of the search tree we use the representation function to obtain a hidden state given the current observation.
We then run a Monte Carlo Tree Search using only action sequences and the model learned by the network.
At the root of the search tree we use the representation function to obtain a
hidden state given the current observation.
We then run a Monte Carlo Tree Search using only action sequences and the model
learned by the network.
"""
root = Node(0)
observation = (
......@@ -153,7 +163,8 @@ class MCTS:
last_action = action
search_path.append(node)
# Inside the search tree we use the dynamics function to obtain the next hidden state given an action and the previous hidden state
# Inside the search tree we use the dynamics function to obtain the next hidden
# state given an action and the previous hidden state
parent = search_path[-2]
value, reward, policy_logits, hidden_state = model.recurrent_inference(
parent.hidden_state,
......@@ -194,10 +205,12 @@ class MCTS:
def backpropagate(self, search_path, value, min_max_stats):
"""
At the end of a simulation, we propagate the evaluation all the way up the tree to the root.
At the end of a simulation, we propagate the evaluation all the way up the tree
to the root.
"""
for node in search_path:
# Always the same player, the other players minds should be modeled in network because environment do not act always in the best way to make you lose
# Always the same player, the other players minds should be modeled in network
# because environment do not act always in the best way to make you lose
node.value_sum += value # if node.to_play == to_play else -value
node.visit_count += 1
min_max_stats.update(node.value())
......@@ -225,7 +238,8 @@ class Node:
def expand(self, actions, reward, policy_logits, hidden_state):
"""
We expand a node using the value, reward and policy prediction obtained from the neural network.
We expand a node using the value, reward and policy prediction obtained from the
neural network.
"""
self.reward = reward
self.hidden_state = hidden_state
......@@ -236,7 +250,8 @@ class Node:
def add_exploration_noise(self, dirichlet_alpha, exploration_fraction):
"""
At the start of each search, we add dirichlet noise to the prior of the root to encourage the search to explore new actions.
At the start of each search, we add dirichlet noise to the prior of the root to
encourage the search to explore new actions.
"""
actions = list(self.children.keys())
noise = numpy.random.dirichlet([dirichlet_alpha] * len(actions))
......
shared_storage.py
......@@ -2,21 +2,25 @@ import ray
import torch
import os
@ray.remote
class SharedStorage:
"""
Class which run in a dedicated thread to store the network weights and some information.
"""
def __init__(self, weights, training_step, game_name, config):
def __init__(self, weights, game_name, config):
self.config = config
self.game_name = game_name
self.weights = weights
self.infos = {'training_step': training_step,
'total_reward': 0,
'total_loss': 0,
'value_loss': 0,
'reward_loss': 0,
'policy_loss': 0}
self.infos = {
"training_step": 0,
"total_reward": 0,
"total_loss": 0,
"value_loss": 0,
"reward_loss": 0,
"policy_loss": 0,
}
def get_weights(self):
return self.weights
......@@ -25,6 +29,7 @@ class SharedStorage:
self.weights = weights
if not path:
path = os.path.join(self.config.results_path, self.game_name)
torch.save(self.weights, path)
def get_infos(self):
......
trainer.py
......@@ -10,11 +10,13 @@ import models
@ray.remote(num_gpus=1)
class Trainer:
"""
Class which run in a dedicated thread to train a neural network and save it in the shared storage.
Class which run in a dedicated thread to train a neural network and save it
in the shared storage.
"""
def __init__(self, initial_weights, initial_training_step, config, device):
def __init__(self, initial_weights, config, device):
self.config = config
self.training_step = initial_training_step
self.training_step = 0
# Initialize the network
self.model = models.MuZeroNetwork(
......@@ -55,6 +57,9 @@ class Trainer:
shared_storage_worker.set_infos.remote("reward_loss", reward_loss)
shared_storage_worker.set_infos.remote("policy_loss", policy_loss)
if self.config.training_delay:
time.sleep(self.config.training_delay)
def update_weights(self, batch):
"""
Perform one training step.
......@@ -66,7 +71,6 @@ class Trainer:
for param_group in self.optimizer.param_groups:
param_group["lr"] = lr
(
observation_batch,
action_batch,
......
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