def initialize(self, env_spaces, share_memory=False,
global_B=1, env_ranks=None):
super().initialize(env_spaces, share_memory,
global_B=global_B, env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
class DqnAgent(EpsilonGreedyAgentMixin, BaseAgent):
def __call__(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
return q.cpu()
def initialize(self,
env_spaces,
share_memory=False,
global_B=1,
env_ranks=None):
super().initialize(env_spaces,
share_memory,
global_B=global_B,
env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
def to_device(self, cuda_idx=None):
super().to_device(cuda_idx)
self.target_model.to(self.device)
def state_dict(self):
return dict(model=self.model.state_dict(),
target=self.target_model.state_dict())
@torch.no_grad()
def step(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
q = q.cpu()
action = self.distribution.sample(q)
agent_info = AgentInfo(q=q)
# action, agent_info = buffer_to((action, agent_info), device="cpu")
return AgentStep(action=action, agent_info=agent_info)
def target(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
target_q = self.target_model(*model_inputs)
return target_q.cpu()
def update_target(self, tau=1):
update_state_dict(self.target_model, self.model.state_dict(), tau)
def initialize(self, env_spaces, share_memory=False, global_B=1, env_ranks=None):
"""Along with standard initialization, creates vector-valued epsilon
for exploration, if applicable, with a different epsilon for each
environment instance."""
self.model = self.ModelCls(
image_shape=env_spaces.observation.shape,
action_size=env_spaces.action.n,
**self.model_kwargs
)
if self.load_conv:
logger.log("Agent loading state dict: " + self.state_dict_filename)
loaded_state_dict = torch.load(
self.state_dict_filename, map_location=torch.device("cpu")
)
# From UL, saves snapshot: params["algo_state_dict"]["encoder"]
loaded_state_dict = loaded_state_dict.get(
"algo_state_dict", loaded_state_dict
)
loaded_state_dict = loaded_state_dict.get("encoder", loaded_state_dict)
# A bit onerous, but ensures that state dicts match:
conv_state_dict = OrderedDict(
[
(k.replace("conv.", "", 1), v)
for k, v in loaded_state_dict.items()
if k.startswith("conv.")
]
)
self.model.conv.load_state_dict(conv_state_dict)
logger.log("Agent loaded CONV state dict.")
elif self.load_all:
# From RL, saves snapshot: params["agent_state_dict"]
loaded_state_dict = torch.load(
self.state_dict_filename, map_location=torch.device("cpu")
)
self.load_state_dict(loaded_state_dict["agent_state_dict"])
logger.log("Agnet loaded FULL state dict.")
else:
logger.log("Agent NOT loading state dict.")
self.target_model = copy.deepcopy(self.model)
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
if share_memory:
self.model.share_memory()
self.shared_model = self.model
if self.initial_model_state_dict is not None:
raise NotImplementedError
self.env_spaces = env_spaces
self.share_memory = share_memory
def initialize(self, env_spaces, share_memory=False, global_B=1, env_ranks=None):
"""
初始化agent。这个函数在Sampler类(例如SerialSampler)中的 initialize() 里会被调用。
:param env_spaces: 参考 Env.spaces()，类型为 EnvSpaces 这样一个 namedtuple，包含observation space 和 action space两个属性。
:param share_memory: 为 True 时使得模型参数可以在多进程间共享，为 False 时不共享。
:param global_B: 在BatchSpec中，表示独立的trajectory的数量，即environment实例的数量。这里的global_B可能是指所有env的总数
:param env_ranks: 其含义参考我写的文章 https://www.codelast.com/?p=10932
"""
super().initialize(env_spaces, share_memory, global_B=global_B, env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs, **self.model_kwargs) # torch.nn.Module的子类
self.target_model.load_state_dict(self.model.state_dict()) # 加载PyTorch模型，开始的时候target network和main network一致
self.distribution = EpsilonGreedy(dim=env_spaces.action.n) # 按ε-greedy方法来探索，n是action的维度
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
def initialize(self, env_spaces, share_memory=False):
env_model_kwargs = self.make_env_to_model_kwargs(env_spaces)
self.model = self.ModelCls(**env_model_kwargs, **self.model_kwargs)
if share_memory:
self.model.share_memory()
self.shared_model = self.model
if self.initial_model_state_dict is not None:
self.model.load_state_dict(self.initial_model_state_dict)
self.target_model = self.ModelCls(**env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
self.env_spaces = env_spaces
self.env_model_kwargs = env_model_kwargs
self.share_memory = share_memory
super().initialize(env_spaces, share_memory)
def initialize(self,
env_spaces,
share_memory=False,
global_B=1,
env_ranks=None):
"""Along with standard initialization, creates vector-valued epsilon
for exploration, if applicable, with a different epsilon for each
environment instance."""
super().initialize(env_spaces,
share_memory,
global_B=global_B,
env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
class AtariDqnAgent(EpsilonGreedyAgentMixin, BaseAgent):
"""
Standard agent for DQN algorithms with epsilon-greedy exploration.
"""
def __init__(self,
ModelCls=AtariDqnModel,
model_kwargs=None,
load_conv=False,
load_all=False,
state_dict_filename=None,
store_latent=False,
**kwargs):
if model_kwargs is None:
model_kwargs = dict()
assert not (load_conv and load_all)
save__init__args(locals())
super().__init__(ModelCls=ModelCls, **kwargs)
def __call__(self, observation, prev_action, prev_reward):
"""Returns Q-values for states/observations (with grad)."""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q, _conv = self.model(*model_inputs)
return q.cpu()
def initialize(self,
env_spaces,
share_memory=False,
global_B=1,
env_ranks=None):
"""Along with standard initialization, creates vector-valued epsilon
for exploration, if applicable, with a different epsilon for each
environment instance."""
self.model = self.ModelCls(image_shape=env_spaces.observation.shape,
action_size=env_spaces.action.n,
**self.model_kwargs)
if self.load_conv:
logger.log("Agent loading state dict: " + self.state_dict_filename)
loaded_state_dict = torch.load(self.state_dict_filename,
map_location=torch.device('cpu'))
# From UL, saves snapshot: params["algo_state_dict"]["encoder"]
loaded_state_dict = loaded_state_dict.get("algo_state_dict",
loaded_state_dict)
loaded_state_dict = loaded_state_dict.get("encoder",
loaded_state_dict)
# A bit onerous, but ensures that state dicts match:
conv_state_dict = OrderedDict([
(k.replace("conv.", "", 1), v)
for k, v in loaded_state_dict.items() if k.startswith("conv.")
])
self.model.conv.load_state_dict(conv_state_dict)
logger.log("Agent loaded CONV state dict.")
elif self.load_all:
# From RL, saves snapshot: params["agent_state_dict"]
loaded_state_dict = torch.load(self.state_dict_filename,
map_location=torch.device('cpu'))
self.load_state_dict(loaded_state_dict["agent_state_dict"])
logger.log("Agnet loaded FULL state dict.")
else:
logger.log("Agent NOT loading state dict.")
self.target_model = copy.deepcopy(self.model)
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
if share_memory:
self.model.share_memory()
self.shared_model = self.model
if self.initial_model_state_dict is not None:
raise NotImplementedError
self.env_spaces = env_spaces
self.share_memory = share_memory
def to_device(self, cuda_idx=None):
super().to_device(cuda_idx)
self.target_model.to(self.device)
def state_dict(self):
return dict(model=self.model.state_dict(),
target=self.target_model.state_dict())
@torch.no_grad()
def step(self, observation, prev_action, prev_reward):
"""Computes Q-values for states/observations and selects actions by
epsilon-greedy. (no grad)"""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q, conv = self.model(*model_inputs)
q = q.cpu()
action = self.distribution.sample(q)
agent_info = AgentInfoConv(q=q,
conv=conv if self.store_latent else None)
# action, agent_info = buffer_to((action, agent_info), device="cpu")
return AgentStep(action=action, agent_info=agent_info)
def target(self, observation, prev_action, prev_reward):
"""Returns the target Q-values for states/observations."""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
target_q, _conv = self.target_model(*model_inputs)
return target_q.cpu()
def update_target(self, tau=1):
"""Copies the model parameters into the target model."""
update_state_dict(self.target_model, self.model.state_dict(), tau)
class DqnAgent(EpsilonGreedyAgentMixin, BaseAgent):
def __call__(self, observation, prev_action, prev_reward):
"""
__call__使得一个class可以像一个method一样调用，即：假设agent为DqnAgent的一个对象，那么agent(observation, prev_action,
prev_reward)就等同于调用agent.__call__(observation, prev_action, prev_reward)
"""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward), device=self.device)
q = self.model(*model_inputs) # torch.nn.Module子类的实例，使用torch.nn.Module里定义的__call__调用，相当于计算模型输出(一个Tensor)
return q.cpu() # 将tensor移动到CPU(内存)
def initialize(self, env_spaces, share_memory=False, global_B=1, env_ranks=None):
"""
初始化agent。这个函数在Sampler类(例如SerialSampler)中的 initialize() 里会被调用。
:param env_spaces: 参考 Env.spaces()，类型为 EnvSpaces 这样一个 namedtuple，包含observation space 和 action space两个属性。
:param share_memory: 为 True 时使得模型参数可以在多进程间共享，为 False 时不共享。
:param global_B: 在BatchSpec中，表示独立的trajectory的数量，即environment实例的数量。这里的global_B可能是指所有env的总数
:param env_ranks: 其含义参考我写的文章 https://www.codelast.com/?p=10932
"""
super().initialize(env_spaces, share_memory, global_B=global_B, env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs, **self.model_kwargs) # torch.nn.Module的子类
self.target_model.load_state_dict(self.model.state_dict()) # 加载PyTorch模型，开始的时候target network和main network一致
self.distribution = EpsilonGreedy(dim=env_spaces.action.n) # 按ε-greedy方法来探索，n是action的维度
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
def to_device(self, cuda_idx=None):
"""
指定把模型数据(parameter和buffer)放在什么设备上(CPU/GPU)。
父类是指定self.model的数据放在哪个GPU上，在本子类中是指定self.target_model的数据放在哪。
:param cuda_idx: GPU编号
"""
super().to_device(cuda_idx)
self.target_model.to(self.device) # self.device在初始化的时候已经写死了是CPU，因此这里指定在CPU上运行
def state_dict(self):
"""
返回main network和target network两个网络的state数据。例如网络的weight，bias等。
:return: 一个dict。
"""
return dict(model=self.model.state_dict(), target=self.target_model.state_dict())
@torch.no_grad()
def step(self, observation, prev_action, prev_reward):
"""
在environment中走一步。environment类(例如AtariEnv也有一个step()，那个step()和这里的step()的主要区别是：这里的step()根据policy
network选取了一个action，而AtariEnv里的step()输入的action是已经选取好的action，并且AtariEnv的step()会计算reward，记录一些统计
信息等，这里的step()不会去计算reward。
这个函数在Collector类的collect_batch()函数中会被调用。
这里会发生policy network的前向传播过程(比较耗计算资源的操作)，即根据输入(例如observation)计算下一步要采取的action。
:param observation: 其义自明。
:param prev_action: 前一个action。
:param prev_reward: 之前累积的reward。
:return: 要采取的action(类型为torch.Tensor)，以及agent的信息(例如Q值)
"""
prev_action = self.distribution.to_onehot(prev_action) # 返回类型为 torch.Tensor
model_inputs = buffer_to((observation, prev_action, prev_reward), device=self.device) # 策略网络的输入(torch.Tensor)
q = self.model(*model_inputs) # self.model是torch.nn.Module的子类对象，这里是输入特征计算网络的输出，因此会发生NN的forward过程
q = q.cpu() # 把tensor移到CPU(内存)，返回torch.Tensor
action = self.distribution.sample(q) # 选择一个action(torch.Tensor)
agent_info = AgentInfo(q=q)
# action, agent_info = buffer_to((action, agent_info), device="cpu")
return AgentStep(action=action, agent_info=agent_info)
def target(self, observation, prev_action, prev_reward):
"""
计算Q值。
:param observation: 如其名。
:param prev_action: 前一个action。
:param prev_reward: 前一个reward。
:return: CPU(内存)里的Q值对应的Tensor。
"""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward), device=self.device)
# 计算Q值，self.target_model是一个torch.nn.Module子类的实例，使用torch.nn.Module里定义的__call__调用，相当于计算模型输出(一个Tensor)
target_q = self.target_model(*model_inputs)
return target_q.cpu() # 将tensor移动到CPU(内存)
def update_target(self, tau=1):
"""
更新target network，即把main network(self.model)的参数拷贝到target network(self.target_model)上。
当τ>0的时候会使用soft update算法来更新参数。
为了保持learning过程的稳定性以及高效性，target network不是实时更新，而是周期性地更新一次。例如，在DQN.optimize_agent()函数中，
会看到每隔一定的周期才会调用一次本函数的代码逻辑。
:param tau: soft update算法里的τ参数。
"""
update_state_dict(self.target_model, self.model.state_dict(), tau)
class DqnAgent(EpsilonGreedyAgentMixin, BaseAgent):
"""
Standard agent for DQN algorithms with epsilon-greedy exploration.
"""
def __call__(self, observation, prev_action, prev_reward):
"""Returns Q-values for states/observations (with grad)."""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
return q.cpu()
def initialize(self,
env_spaces,
share_memory=False,
global_B=1,
env_ranks=None):
"""Along with standard initialization, creates vector-valued epsilon
for exploration, if applicable, with a different epsilon for each
environment instance."""
super().initialize(env_spaces,
share_memory,
global_B=global_B,
env_ranks=env_ranks)
self.target_model = self.ModelCls(**self.env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
if env_ranks is not None:
self.make_vec_eps(global_B, env_ranks)
def to_device(self, cuda_idx=None):
super().to_device(cuda_idx)
self.target_model.to(self.device)
def state_dict(self):
return dict(model=self.model.state_dict(),
target=self.target_model.state_dict())
@torch.no_grad()
def step(self, observation, prev_action, prev_reward):
"""Computes Q-values for states/observations and selects actions by
epsilon-greedy. (no grad)"""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
q = q.cpu()
action = self.distribution.sample(q)
agent_info = AgentInfo(q=q)
# action, agent_info = buffer_to((action, agent_info), device="cpu")
return AgentStep(action=action, agent_info=agent_info)
def target(self, observation, prev_action, prev_reward):
"""Returns the target Q-values for states/observations."""
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
target_q = self.target_model(*model_inputs)
return target_q.cpu()
def update_target(self, tau=1):
"""Copies the model parameters into the target model."""
update_state_dict(self.target_model, self.model.state_dict(), tau)
class DqnAgent(EpsilonGreedyAgentMixin, BaseAgent):
def __call__(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
return q.cpu()
def initialize(self, env_spaces, share_memory=False):
env_model_kwargs = self.make_env_to_model_kwargs(env_spaces)
self.model = self.ModelCls(**env_model_kwargs, **self.model_kwargs)
if share_memory:
self.model.share_memory()
self.shared_model = self.model
if self.initial_model_state_dict is not None:
self.model.load_state_dict(self.initial_model_state_dict)
self.target_model = self.ModelCls(**env_model_kwargs,
**self.model_kwargs)
self.target_model.load_state_dict(self.model.state_dict())
self.distribution = EpsilonGreedy(dim=env_spaces.action.n)
self.env_spaces = env_spaces
self.env_model_kwargs = env_model_kwargs
self.share_memory = share_memory
super().initialize(env_spaces, share_memory)
def initialize_cuda(self, cuda_idx=None, ddp=False):
if cuda_idx is None:
return # CPU
if self.shared_model is not None:
self.model = self.ModelCls(**self.env_model_kwargs,
**self.model_kwargs)
self.model.load_state_dict(self.shared_model.state_dict())
self.device = torch.device("cuda", index=cuda_idx)
self.model.to(self.device)
if ddp:
self.model = DDP(self.model,
device_ids=[cuda_idx],
output_device=cuda_idx)
logger.log("Initialized DistributedDataParallel agent model "
f"on device: {self.device}.")
else:
logger.log(f"Initialized agent model on device: {self.device}.")
self.target_model.to(self.device)
def make_env_to_model_kwargs(self, env_spaces):
raise NotImplementedError
def state_dict(self):
return dict(model=self.model.state_dict(),
target=self.target_model.state_dict())
@torch.no_grad()
def step(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
q = self.model(*model_inputs)
q = q.cpu()
action = self.distribution.sample(q)
agent_info = AgentInfo(q=q)
# action, agent_info = buffer_to((action, agent_info), device="cpu")
return AgentStep(action=action, agent_info=agent_info)
def target(self, observation, prev_action, prev_reward):
prev_action = self.distribution.to_onehot(prev_action)
model_inputs = buffer_to((observation, prev_action, prev_reward),
device=self.device)
target_q = self.target_model(*model_inputs)
return target_q.cpu()
def update_target(self):
self.target_model.load_state_dict(self.model.state_dict())