def test_policy_agent(self):
policy_agent = PolicyAgent(self.net)
action = policy_agent(self.states, self.device)
self.assertIsInstance(action, list)
self.assertEqual(action[0], 1)
def __init__(self,
env: str,
gamma: float = 0.99,
lr: float = 1e-4,
batch_size: int = 32,
entropy_beta: float = 0.01,
batch_episodes: int = 4,
*args,
**kwargs) -> None:
"""
PyTorch Lightning implementation of `Vanilla Policy Gradient
<https://papers.nips.cc/paper/
1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf>`_
Paper authors: Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
Model implemented by:
- `Donal Byrne <https://github.com/djbyrne>`
Example:
>>> from pl_bolts.models.rl.vanilla_policy_gradient_model import PolicyGradient
...
>>> model = PolicyGradient("PongNoFrameskip-v4")
Train::
trainer = Trainer()
trainer.fit(model)
Args:
env: gym environment tag
gamma: discount factor
lr: learning rate
batch_size: size of minibatch pulled from the DataLoader
batch_episodes: how many episodes to rollout for each batch of training
entropy_beta: dictates the level of entropy per batch
.. note::
This example is based on:
https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On-Second-Edition\
/blob/master/Chapter11/04_cartpole_pg.py
.. note:: Currently only supports CPU and single GPU training with `distributed_backend=dp`
"""
super().__init__()
# self.env = wrappers.make_env(self.hparams.env) # use for Atari
self.env = ToTensor(gym.make(env)) # use for Box2D/Control
self.env.seed(123)
self.obs_shape = self.env.observation_space.shape
self.n_actions = self.env.action_space.n
self.net = None
self.build_networks()
self.agent = PolicyAgent(self.net)
self.source = NStepExperienceSource(env=self.env,
agent=self.agent,
n_steps=10)
self.gamma = gamma
self.lr = lr
self.batch_size = batch_size
self.batch_episodes = batch_episodes
self.entropy_beta = entropy_beta
self.baseline = 0
# Metrics
self.reward_sum = 0
self.env_steps = 0
self.total_steps = 0
self.total_reward = 0
self.episode_count = 0
self.reward_list = []
for _ in range(100):
self.reward_list.append(torch.tensor(0, device=self.device))
self.avg_reward = 0
def __init__(
self,
env: str,
gamma: float = 0.99,
lr: float = 0.01,
batch_size: int = 8,
n_steps: int = 10,
avg_reward_len: int = 100,
entropy_beta: float = 0.01,
epoch_len: int = 1000,
num_batch_episodes: int = 4,
**kwargs
) -> None:
"""
PyTorch Lightning implementation of `REINFORCE
<https://papers.nips.cc/paper/
1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf>`_
Paper authors: Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
Model implemented by:
- `Donal Byrne <https://github.com/djbyrne>`
Example:
>>> from pl_bolts.models.rl.reinforce_model import Reinforce
...
>>> model = Reinforce("CartPole-v0")
Train::
trainer = Trainer()
trainer.fit(model)
Args:
env: gym environment tag
gamma: discount factor
lr: learning rate
batch_size: size of minibatch pulled from the DataLoader
n_steps: number of stakes per discounted experience
entropy_beta: entropy coefficient
epoch_len: how many batches before pseudo epoch
num_batch_episodes: how many episodes to rollout for each batch of training
avg_reward_len: how many episodes to take into account when calculating the avg reward
Note:
This example is based on:
https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On-Second-Edition/blob/master/Chapter11/02_cartpole_reinforce.py
Note:
Currently only supports CPU and single GPU training with `distributed_backend=dp`
"""
super().__init__()
if not _GYM_AVAILABLE:
raise ModuleNotFoundError('This Module requires gym environment which is not installed yet.')
# Hyperparameters
self.lr = lr
self.batch_size = batch_size
self.batches_per_epoch = self.batch_size * epoch_len
self.entropy_beta = entropy_beta
self.gamma = gamma
self.n_steps = n_steps
self.num_batch_episodes = num_batch_episodes
self.save_hyperparameters()
# Model components
self.env = gym.make(env)
self.net = MLP(self.env.observation_space.shape, self.env.action_space.n)
self.agent = PolicyAgent(self.net)
# Tracking metrics
self.total_steps = 0
self.total_rewards = [0]
self.done_episodes = 0
self.avg_rewards = 0
self.reward_sum = 0.0
self.batch_episodes = 0
self.avg_reward_len = avg_reward_len
self.batch_states = []
self.batch_actions = []
self.batch_qvals = []
self.cur_rewards = []
self.state = self.env.reset()
def __init__(self,
env: str,
gamma: float = 0.99,
lr: float = 0.01,
batch_size: int = 8,
n_steps: int = 10,
avg_reward_len: int = 100,
entropy_beta: float = 0.01,
epoch_len: int = 1000,
num_batch_episodes: int = 4,
**kwargs) -> None:
"""
Args:
env: gym environment tag
gamma: discount factor
lr: learning rate
batch_size: size of minibatch pulled from the DataLoader
n_steps: number of stakes per discounted experience
entropy_beta: entropy coefficient
epoch_len: how many batches before pseudo epoch
num_batch_episodes: how many episodes to rollout for each batch of training
avg_reward_len: how many episodes to take into account when calculating the avg reward
"""
super().__init__()
if not _GYM_AVAILABLE:
raise ModuleNotFoundError(
'This Module requires gym environment which is not installed yet.'
)
# Hyperparameters
self.lr = lr
self.batch_size = batch_size
self.batches_per_epoch = self.batch_size * epoch_len
self.entropy_beta = entropy_beta
self.gamma = gamma
self.n_steps = n_steps
self.num_batch_episodes = num_batch_episodes
self.save_hyperparameters()
# Model components
self.env = gym.make(env)
self.net = MLP(self.env.observation_space.shape,
self.env.action_space.n)
self.agent = PolicyAgent(self.net)
# Tracking metrics
self.total_steps = 0
self.total_rewards = [0]
self.done_episodes = 0
self.avg_rewards = 0
self.reward_sum = 0.0
self.batch_episodes = 0
self.avg_reward_len = avg_reward_len
self.batch_states = []
self.batch_actions = []
self.batch_qvals = []
self.cur_rewards = []
self.state = self.env.reset()
def __init__(self,
env: str,
gamma: float = 0.99,
lr: float = 0.01,
batch_size: int = 8,
n_steps: int = 10,
avg_reward_len: int = 100,
num_envs: int = 4,
entropy_beta: float = 0.01,
epoch_len: int = 1000,
**kwargs) -> None:
"""
PyTorch Lightning implementation of `Vanilla Policy Gradient
<https://papers.nips.cc/paper/
1713-policy-gradient-methods-for-reinforcement-learning-with-function-approximation.pdf>`_
Paper authors: Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
Model implemented by:
- `Donal Byrne <https://github.com/djbyrne>`
Example:
>>> from pl_bolts.models.rl.vanilla_policy_gradient_model import VanillaPolicyGradient
...
>>> model = VanillaPolicyGradient("PongNoFrameskip-v4")
Train::
trainer = Trainer()
trainer.fit(model)
Args:
env: gym environment tag
gamma: discount factor
lr: learning rate
batch_size: size of minibatch pulled from the DataLoader
batch_episodes: how many episodes to rollout for each batch of training
entropy_beta: dictates the level of entropy per batch
avg_reward_len: how many episodes to take into account when calculating the avg reward
Note:
This example is based on:
https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On-Second-Edition/blob/master/Chapter11/04_cartpole_pg.py
Note:
Currently only supports CPU and single GPU training with `distributed_backend=dp`
"""
super().__init__()
# Hyperparameters
self.lr = lr
self.batch_size = batch_size * num_envs
self.batches_per_epoch = self.batch_size * epoch_len
self.entropy_beta = entropy_beta
self.gamma = gamma
self.n_steps = n_steps
self.save_hyperparameters()
# Model components
self.env = [gym.make(env) for _ in range(num_envs)]
self.net = MLP(self.env[0].observation_space.shape,
self.env[0].action_space.n)
self.agent = PolicyAgent(self.net)
self.exp_source = DiscountedExperienceSource(self.env,
self.agent,
gamma=gamma,
n_steps=self.n_steps)
# Tracking metrics
self.total_steps = 0
self.total_rewards = [0]
self.done_episodes = 0
self.avg_rewards = 0
self.reward_sum = 0.0
self.baseline = 0
self.avg_reward_len = avg_reward_len
def __init__(self,
env: str,
gamma: float = 0.99,
lr: float = 1e-4,
batch_size: int = 32,
entropy_beta: float = 0.01,
batch_episodes: int = 4,
*args,
**kwargs) -> None:
"""
PyTorch Lightning implementation of `Vanilla Policy Gradient
<https://papers.nips.cc/paper/
1713-policy-gradient-methods-for-rl-learning-with-function-approximation.pdf>`_
Paper authors: Richard S. Sutton, David McAllester, Satinder Singh, Yishay Mansour
Model implemented by:
- `Donal Byrne <https://github.com/djbyrne>`
Example:
>>> from pl_bolts.models.rl.vanilla_policy_gradient.model import PolicyGradient
...
>>> model = PolicyGradient("PongNoFrameskip-v4")
Train::
trainer = Trainer()
trainer.fit(model)
Args:
env: gym environment tag
gamma: discount factor
lr: learning rate
batch_size: size of minibatch pulled from the DataLoader
batch_episodes: how many episodes to rollout for each batch of training
entropy_beta: dictates the level of entropy per batch
"""
super().__init__()
# self.env = wrappers.make_env(self.hparams.env) # use for Atari
self.env = ToTensor(gym.make(env)) # use for Box2D/Control
self.env.seed(123)
self.obs_shape = self.env.observation_space.shape
self.n_actions = self.env.action_space.n
self.net = None
self.build_networks()
self.agent = PolicyAgent(self.net)
self.gamma = gamma
self.lr = lr
self.batch_size = batch_size
self.batch_episodes = batch_episodes
self.total_reward = 0
self.episode_reward = 0
self.episode_count = 0
self.episode_steps = 0
self.total_episode_steps = 0
self.entropy_beta = entropy_beta
self.reward_list = []
for _ in range(100):
self.reward_list.append(0)
self.avg_reward = 0