Introduction:

This libaray allows you to easily train agents built with Keras or PyTorch using reinforcement learning. You just need to have your agent class inherit from the RL or RL_pytorch class, and you can easily train your agent built with Keras or PyTorch. You can learn how to build an agent from the examples here. The README shows how to train, save, and restore agent built with Keras or PyTorch.

Installation:

To use this library, you need to download it and then unzip it to the site-packages folder of your Python environment.

dependent packages:

tensorflow>=2.16.1

pytorch>=2.3.1

gym<=0.25.2

matplotlib>=3.8.4

python requirement:

python>=3.10

Train:

Keras:

Agent built with Keras.

import tensorflow as tf
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.keras.DQN import DQN

model=DQN(4,128,2)
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10)
optimizer = tf.keras.optimizers.Adam()
train_loss = tf.keras.metrics.Mean(name='train_loss')
model.train(train_loss, optimizer, 100)

# If set criterion.
# model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10,trial_count=10,criterion=200)
# model.train(train_loss, optimizer, 100)

# If save the model at intervals of 10 episode, with a maximum of 2 saved file, and the file name is model.dat.
# model.path='model.dat'
# model.save_freq=10
# model. max_save_files=2
# model.train(train_loss, optimizer, 100)

# If save parameters only
# model.path='param.dat'
# model.save_freq=10
# model. max_save_files=2
# model.save_param_only=True
# model.train(train_loss, optimizer, 100)

# If save best only
# model.path='model.dat'
# model.save_best_only=True
# model.train(train_loss, optimizer, 100)

# visualize
# model.visualize_loss()
# model.visualize_reward()
# model.visualize_reward_loss()

# animate agent
# model.animate_agent(200)

# save
# model.save_param('param.dat')
# model.save('model.dat')

# Use PPO.
import tensorflow as tf
from Note_rl.policy import SoftmaxPolicy
from Note_rl.examples.keras.PPO import PPO

model=PPO(4,128,2,0.7,0.7)
model.set(policy=SoftmaxPolicy(),pool_size=10000,batch=64,update_steps=1000,PPO=True)
optimizer = [tf.keras.optimizers.Adam(1e-4),tf.keras.optimizers.Adam(5e-3)]
train_loss = tf.keras.metrics.Mean(name='train_loss')
model.train(train_loss, optimizer, 100)

# Use HER.
import tensorflow as tf
from Note_rl.noise import GaussianWhiteNoiseProcess
from Note_rl.examples.keras.DDPG_HER import DDPG

model=DDPG(128,0.1,0.98,0.005)
model.set(noise=GaussianWhiteNoiseProcess(),pool_size=10000,batch=256,criterion=-5,trial_count=10,HER=True)
optimizer = [tf.keras.optimizers.Adam(),tf.keras.optimizers.Adam()]
train_loss = tf.keras.metrics.Mean(name='train_loss')
model.train(train_loss, optimizer, 2000)

# Use Multi-agent reinforcement learning.
import tensorflow as tf
from Note_rl.policy import SoftmaxPolicy
from Note_rl.examples.keras.MADDPG import DDPG

model=DDPG(128,0.1,0.98,0.005)
model.set(policy=SoftmaxPolicy(),pool_size=3000,batch=32,trial_count=10,MARL=True)
optimizer = [tf.keras.optimizers.Adam(),tf.keras.optimizers.Adam()]
train_loss = tf.keras.metrics.Mean(name='train_loss')
model.train(train_loss, optimizer, 100)

# This technology uses Python’s multiprocessing module to speed up trajectory collection and storage, I call it Pool Network.
import tensorflow as tf
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.keras.pool_network.DQN import DQN

model=DQN(4,128,2,7)
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,update_batches=17)
optimizer = tf.keras.optimizers.Adam()
train_loss = tf.keras.metrics.Mean(name='train_loss')
model.train(train_loss, optimizer, 100, pool_network=True, processes=7)

PyTorch:

Agent built with PyTorch.

import torch
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.pytorch.DQN import DQN

model=DQN(4,128,2)
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10)
optimizer = torch.optim.Adam(model.param)
model.train(optimizer, 100)

# If set criterion.
# model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10,trial_count=10,criterion=200)
# model.train(optimizer, 100)

# If use prioritized replay.
# model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10,trial_count=10,criterion=200,PR=True,initial_TD=7,alpha=0.7)
# model.train(optimizer, 100)

# If save the model at intervals of 10 episode, with a maximum of 2 saved file, and the file name is model.dat.
# model.path='model.dat'
# model.save_freq=10
# model. max_save_files=2
# model.train(optimizer, 100)

# If save parameters only
# model.path='param.dat'
# model.save_freq=10
# model. max_save_files=2
# model.save_param_only=True
# model.train(optimizer, 100)

# If save best only
# model.path='model.dat'
# model.save_best_only=True
# model.train(optimizer, 100)

# visualize
# model.visualize_loss()
# model.visualize_reward()
# model.visualize_reward_loss()

# animate agent
# model.animate_agent(200)

# save
# model.save_param('param.dat')
# model.save('model.dat')

# Use HER.
import torch
from Note_rl.noise import GaussianWhiteNoiseProcess
from Note_rl.examples.pytorch.DDPG_HER import DDPG

model=DDPG(128,0.1,0.98,0.005)
model.set(noise=GaussianWhiteNoiseProcess(),pool_size=10000,batch=256,criterion=-5,trial_count=10,HER=True)
optimizer = [torch.optim.Adam(model.param[0]),torch.optim.Adam(model.param[1])]
model.train(optimizer, 2000)

# Use Multi-agent reinforcement learning.
import torch
from Note_rl.policy import SoftmaxPolicy
from Note_rl.examples.pytorch.MADDPG import DDPG

model=DDPG(128,0.1,0.98,0.005)
model.set(policy=SoftmaxPolicy(),pool_size=3000,batch=32,trial_count=10,MARL=True)
optimizer = [torch.optim.Adam(model.param[0]),torch.optim.Adam(model.param[1])]
model.train(optimizer, 100)

# This technology uses Python’s multiprocessing module to speed up trajectory collection and storage, I call it Pool Network.
import torch
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.pytorch.pool_network.DQN import DQN

model=DQN(4,128,2,7)
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_batches=17)
optimizer = torch.optim.Adam(model.param)
model.train(optimizer, 100, pool_network=True, processes=7)

# Use HER.
# This technology uses Python’s multiprocessing module to speed up trajectory collection and storage, I call it Pool Network.
# Furthermore use Python’s multiprocessing module to speed up getting a batch of data.
import torch
from Note_rl.noise import GaussianWhiteNoiseProcess
from Note_rl.examples.pytorch.pool_network.DDPG_HER import DDPG

model=DDPG(128,0.1,0.98,0.005,7)
model.set(noise=GaussianWhiteNoiseProcess(),pool_size=10000,batch=256,trial_count=10,HER=True)
optimizer = [torch.optim.Adam(model.param[0]),torch.optim.Adam(model.param[1])]
model.train(train_loss, optimizer, 2000, pool_network=True, processes=7, processes_her=4)

Distributed training:

MirroredStrategy:

Agent built with Keras.

import tensorflow as tf
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.keras.DQN import DQN

strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE_PER_REPLICA = 64
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

with strategy.scope():
  model=DQN(4,128,2)
  optimizer = tf.keras.optimizers.Adam()
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=64,update_steps=10)
model.distributed_training(GLOBAL_BATCH_SIZE, optimizer, strategy, 100)

# If set criterion.
# model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=GLOBAL_BATCH_SIZE,update_steps=10,trial_count=10,criterion=200)
# model.distributed_training(optimizer, strategy, 100)

# If save the model at intervals of 10 episode, with a maximum of 2 saved file, and the file name is model.dat.
# model.path='model.dat'
# model.save_freq=10
# model. max_save_files=2
# model.distributed_training(optimizer, strategy, 100)

# If save parameters only
# model.path='param.dat'
# model.save_freq=10
# model. max_save_files=2
# model.save_param_only=True
# model.distributed_training(optimizer, strategy, 100)

# If save best only
# model.path='model.dat'
# model.save_best_only=True
# model.distributed_training(optimizer, strategy, 100)

# visualize
# model.visualize_loss()
# model.visualize_reward()
# model.visualize_reward_loss()

# animate agent
# model.animate_agent(200)

# save
# model.save_param('param.dat')
# model.save('model.dat')

# Use PPO
import tensorflow as tf
from Note_rl.policy import SoftmaxPolicy
from Note_rl.examples.keras.PPO import PPO

strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE_PER_REPLICA = 64
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

with strategy.scope():
  model=PPO(4,128,2,0.7,0.7)
  optimizer = [tf.keras.optimizers.Adam(1e-4),tf.keras.optimizers.Adam(5e-3)]

model.set(policy=SoftmaxPolicy(),pool_size=10000,batch=GLOBAL_BATCH_SIZE,update_steps=1000,PPO=True)
model.distributed_training(optimizer, strategy, 100)

# Use HER.
import tensorflow as tf
from Note_rl.noise import GaussianWhiteNoiseProcess
from Note_rl.examples.keras.DDPG_HER import DDPG

strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE_PER_REPLICA = 256
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

with strategy.scope():
  model=DDPG(128,0.1,0.98,0.005)
  optimizer = [tf.keras.optimizers.Adam(),tf.keras.optimizers.Adam()]

model.set(noise=GaussianWhiteNoiseProcess(),pool_size=10000,batch=GLOBAL_BATCH_SIZE,criterion=-5,trial_count=10,HER=True)
model.distributed_training(optimizer, strategy, 2000)

# Use Multi-agent reinforcement learning.
import tensorflow as tf
from Note_rl.policy import SoftmaxPolicy
from Note_rl.examples.keras.MADDPG import DDPG

strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE_PER_REPLICA = 32
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

with strategy.scope():
  model=DDPG(128,0.1,0.98,0.005)
  optimizer = [tf.keras.optimizers.Adam(),tf.keras.optimizers.Adam()]

model.set(policy=SoftmaxPolicy(),pool_size=3000,batch=GLOBAL_BATCH_SIZE,trial_count=10,MARL=True)
model.distributed_training(optimizer, strategy, 100)

# This technology uses Python’s multiprocessing module to speed up trajectory collection and storage, I call it Pool Network.
import tensorflow as tf
from Note_rl.policy import EpsGreedyQPolicy
from Note_rl.examples.keras.pool_network.DQN import DQN

strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE_PER_REPLICA = 64
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

with strategy.scope():
  model=DQN(4,128,2,7)
  optimizer = tf.keras.optimizers.Adam()
model.set(policy=EpsGreedyQPolicy(0.01),pool_size=10000,batch=GLOBAL_BATCH_SIZE,update_batches=17)
model.distributed_training(optimizer, strategy, 100, pool_network=True, processes=7)

MultiWorkerMirroredStrategy:

import tensorflow as tf
from Note.RL import rl
from Note_rl.examples.keras.pool_network.DQN import DQN
import sys
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
os.environ.pop('TF_CONFIG', None)
if '.' not in sys.path:
  sys.path.insert(0, '.')

tf_config = {
    'cluster': {
        'worker': ['localhost:12345', 'localhost:23456']
    },
    'task': {'type': 'worker', 'index': 0}
}

strategy = tf.distribute.MultiWorkerMirroredStrategy()
per_worker_batch_size = 64
num_workers = len(tf_config['cluster']['worker'])
global_batch_size = per_worker_batch_size * num_workers

with strategy.scope():
  multi_worker_model = DQN(4,128,2)
  optimizer = tf.keras.optimizers.Adam()

multi_worker_model.set(policy=rl.EpsGreedyQPolicy(0.01),pool_size=10000,batch=global_batch_size,update_batches=17)
multi_worker_model.distributed_training(optimizer, strategy, num_episodes=100,
                    pool_network=True, processes=7)

# If set criterion.
# model.set(policy=rl.EpsGreedyQPolicy(0.01),pool_size=10000,batch=global_batch_size,update_steps=10,trial_count=10,criterion=200)
# multi_worker_model.distributed_training(optimizer, strategy, num_episodes=100,
#                    pool_network=True, processes=7)

# If save the model at intervals of 10 episode, with a maximum of 2 saved file, and the file name is model.dat.
# model.path='model.dat'
# model.save_freq=10
# model. max_save_files=2
# multi_worker_model.distributed_training(optimizer, strategy, num_episodes=100,
#                    pool_network=True, processes=7)

# If save parameters only
# model.path='param.dat'
# model.save_freq=10
# model. max_save_files=2
# model.save_param_only=True
# multi_worker_model.distributed_training(optimizer, strategy, num_episodes=100,
#                    pool_network=True, processes=7)

# If save best only
# model.path='model.dat'
# model.save_best_only=True
# multi_worker_model.distributed_training(optimizer, strategy, num_episodes=100,
#                    pool_network=True, processes=7)

# visualize
# model.visualize_loss()
# model.visualize_reward()
# model.visualize_reward_loss()

# animate agent
# model.animate_agent(200)

# save
# model.save_param('param.dat')
# model.save('model.dat')

Name		Name	Last commit message	Last commit date
Latest commit History 105 Commits
Note_rl		Note_rl
LICENSE		LICENSE
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Introduction:

Installation:

Train:

Keras:

PyTorch:

Distributed training:

MirroredStrategy:

MultiWorkerMirroredStrategy:

About

Releases

Packages

Languages

License

NoteDance/Note_rl

Folders and files

Latest commit

History

Repository files navigation

Introduction:

Installation:

Train:

Keras:

PyTorch:

Distributed training:

MirroredStrategy:

MultiWorkerMirroredStrategy:

About

Topics

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages