循环状态空间模型（Recurrent State Space Models, RSSM）最初由 Danijar Hafer 等人在论文《Learning Latent Dynamics for Planning from Pixels》中提出。该模型在现代基于模型的强化学习（Model-Based Reinforcement Learning, MBRL）中发挥着关键作用，其主要目标是构建可靠的环境动态预测模型。通过这些学习得到的模型，智能体能够模拟未来轨迹并进行前瞻性的行为规划。
下面我们就来用一个实际案例来介绍RSSM。
环境配置
环境配置是实现过程中的首要步骤。我们这里用易于使用的 Gym API。为了提高实现效率，设计了多个模块化的包装器（wrapper），用于初始化参数并将观察结果调整为指定格式。
InitialWrapper 的设计允许在不执行任何动作的情况下进行特定数量的观察，同时支持在返回观察结果之前多次重复同一动作。这种设计对于响应具有显著延迟特性的环境特别有效。
PreprocessFrame 包装器负责将观察结果转换为正确的数据类型（本文中使用 numpy 数组），并支持灰度转换功能。
class InitialWrapper(gym.Wrapper):
def __init__(self, env: gym.Env, no_ops: int = 0, repeat: int = 1):
super(InitialWrapper, self).__init__(env)
self.repeat = repeat
self.no_ops = no_ops
self.op_counter = 0
def step(self, action: ActType) -> Tuple[ObsType, float, bool, bool, dict]:
if self.op_counter < self.no_ops:
obs, reward, done, info = self.env.step(0)
self.op_counter += 1
total_reward = 0.0
done = False
for _ in range(self.repeat):
obs, reward, done, info = self.env.step(action)
total_reward += reward
if done:
break
return obs, total_reward, done, info
class PreprocessFrame(gym.ObservationWrapper):
def __init__(self, env: gym.Env, new_shape: Sequence[int] = (128, 128, 3), grayscale: bool = False):
super(PreprocessFrame, self).__init__(env)
self.shape = new_shape
self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=self.shape, dtype=np.float32)
self.grayscale = grayscale
if self.grayscale:
self.observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(*self.shape[:-1], 1), dtype=np.float32)
def observation(self, obs: torch.Tensor) -> torch.Tensor:
obs = obs.astype(np.uint8)
new_frame = cv.resize(obs, self.shape[:-1], interpolation=cv.INTER_AREA)
if self.grayscale:
new_frame = cv.cvtColor(new_frame, cv.COLOR_RGB2GRAY)
new_frame = np.expand_dims(new_frame, -1)
torch_frame = torch.from_numpy(new_frame).float()
torch_frame = torch_frame / 255.0
return torch_frame
def make_env(env_name: str, new_shape: Sequence[int] = (128, 128, 3), grayscale: bool = True, **kwargs):
env = gym.make(env_name, **kwargs)
env = PreprocessFrame(env, new_shape, grayscale=grayscale)
return envmake_env 函数用于创建一个具有指定配置参数的环境实例。
模型架构
RSSM 的实现依赖于多个关键模型组件。具体来说，需要实现以下四个核心模块：
- 原始观察编码器（Encoder）
- 动态模型（Dynamics Model）：通过确定性状态 h 和随机状态 s 对编码观察的时间依赖性进行建模
- 解码器（Decoder）：将随机状态和确定性状态映射回原始观察空间
- 奖励模型（Reward Model）：将随机状态和确定性状态映射到奖励值
RSSM 模型组件结构图。模型包含随机状态 s 和确定性状态 h。
编码器实现
编码器采用简单的卷积神经网络（CNN）结构，将输入图像降维到一维嵌入表示。实现中使用了 BatchNorm 来提升训练稳定性。
class EncoderCNN(nn.Module):
def __init__(self, in_channels: int, embedding_dim: int = 2048, input_shape: Tuple[int, int] = (128, 128)):
super(EncoderCNN, self).__init__()
# 定义卷积层结构
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=3, stride=2, padding=1)
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1)
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1)
self.conv4 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1)
self.fc1 = nn.Linear(self._compute_conv_output((in_channels, input_shape[0], input_shape[1])), embedding_dim)
# 批标准化层
self.bn1 = nn.BatchNorm2d(32)
self.bn2 = nn.BatchNorm2d(64)
self.bn3 = nn.BatchNorm2d(128)
self.bn4 = nn.BatchNorm2d(256)
def _compute_conv_output(self, shape: Tuple[int, int, int]):
with torch.no_grad():
x = torch.randn(1, shape[0], shape[1], shape[2])
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = self.conv4(x)
return x.shape[1] * x.shape[2] * x.shape[3]
def forward(self, x):
x = torch.relu(self.conv1(x))
x = self.bn1(x)
x = torch.relu(self.conv2(x))
x = self.bn2(x)
x = torch.relu(self.conv3(x))
x = self.bn3(x)
x = self.conv4(x)
x = self.bn4(x)
x = x.view(x.size(0), -1)
x = self.fc1(x)
return x解码器实现
解码器遵循传统自编码器架构设计，其功能是将编码后的观察结果重建回原始观察空间。
class DecoderCNN(nn.Module):
def __init__(self, hidden_size: int, state_size: int, embedding_size: int,
use_bn: bool = True, output_shape: Tuple[int, int] = (3, 128, 128)):
super(DecoderCNN, self).__init__()
self.output_shape = output_shape
self.embedding_size = embedding_size
# 全连接层进行特征变换
self.fc1 = nn.Linear(hidden_size + state_size, embedding_size)
self.fc2 = nn.Linear(embedding_size, 256 * (output_shape[1] // 16) * (output_shape[2] // 16))
# 反卷积层进行上采样
self.conv1 = nn.ConvTranspose2d(256, 128, kernel_size=3, stride=2, padding=1, output_padding=1) # ×2
self.conv2 = nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1) # ×2
self.conv3 = nn.ConvTranspose2d(64, 32, kernel_size=3, stride=2, padding=1, output_padding=1) # ×2
self.conv4 = nn.ConvTranspose2d(32, output_shape[0], kernel_size=3, stride=2, padding=1, output_padding=1)
# 批标准化层
self.bn1 = nn.BatchNorm2d(128)
self.bn2 = nn.BatchNorm2d(64)
self.bn3 = nn.BatchNorm2d(32)
self.use_bn = use_bn
def forward(self, h: torch.Tensor, s: torch.Tensor):
x = torch.cat([h, s], dim=-1)
x = self.fc1(x)
x = torch.relu(x)
x = self.fc2(x)
x = x.view(-1, 256, self.output_shape[1] // 16, self.output_shape[2] // 16)
if self.use_bn:
x = torch.relu(self.bn1(self.conv1(x)))
x = torch.relu(self.bn2(self.conv2(x)))
x = torch.relu(self.bn3(self.conv3(x)))
else:
x = torch.relu(self.conv1(x))
x = torch.relu(self.conv2(x))
x = torch.relu(self.conv3(x))
x = self.conv4(x)
return x奖励模型实现
奖励模型采用了一个三层前馈神经网络结构，用于将随机状态 s 和确定性状态 h 映射到正态分布参数，进而通过采样获得奖励预测。
class RewardModel(nn.Module):
def __init__(self, hidden_dim: int, state_dim: int):
super(RewardModel, self).__init__()
self.fc1 = nn.Linear(hidden_dim + state_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, hidden_dim)
self.fc3 = nn.Linear(hidden_dim, 2)
def forward(self, h: torch.Tensor, s: torch.Tensor):
x = torch.cat([h, s], dim=-1)
x = torch.relu(self.fc1(x))
x = torch.relu(self.fc2(x))
x = self.fc3(x)
return x动态模型的实现
动态模型是 RSSM 架构中最复杂的组件，需要同时处理先验和后验状态转移模型：
- 后验转移模型：在能够访问真实观察的情况下使用（主要在训练阶段），用于在给定观察和历史状态的条件下近似随机状态的后验分布。
- 先验转移模型：用于近似先验状态分布，仅依赖于前一时刻状态，不依赖于观察。这在无法获取后验观察的推理阶段使用。
这两个模型均通过单层前馈网络进行参数化，输出各自正态分布的均值和对数方差，用于状态 s 的采样。该实现采用了简单的网络结构，但可以根据需要扩展为更复杂的架构。
确定性状态采用门控循环单元（GRU）实现。其输入包括：
- 前一时刻的隐藏状态
- 独热编码动作
- 前一时刻随机状态 s（根据是否可以获取观察来选择使用后验或先验状态）
这些输入信息足以让模型了解动作历史和系统状态。以下是具体实现代码：
class DynamicsModel(nn.Module):
def __init__(self, hidden_dim: int, action_dim: int, state_dim: int, embedding_dim: int, rnn_layer: int = 1):
super(DynamicsModel, self).__init__()
self.hidden_dim = hidden_dim
# 递归层实现，支持多层 GRU
self.rnn = nn.ModuleList([nn.GRUCell(hidden_dim, hidden_dim) for _ in range(rnn_layer)])
# 状态动作投影层
self.project_state_action = nn.Linear(action_dim + state_dim, hidden_dim)
# 先验网络：输出正态分布参数
self.prior = nn.Linear(hidden_dim, state_dim * 2)
self.project_hidden_action = nn.Linear(hidden_dim + action_dim, hidden_dim)
# 后验网络：输出正态分布参数
self.posterior = nn.Linear(hidden_dim, state_dim * 2)
self.project_hidden_obs = nn.Linear(hidden_dim + embedding_dim, hidden_dim)
self.state_dim = state_dim
self.act_fn = nn.ReLU()
def forward(self, prev_hidden: torch.Tensor, prev_state: torch.Tensor, actions: torch.Tensor,
obs: torch.Tensor = None, dones: torch.Tensor = None):
"""
动态模型的前向传播
参数:
prev_hidden: RNN的前一隐藏状态，形状 (batch_size, hidden_dim)
prev_state: 前一随机状态，形状 (batch_size, state_dim)
actions: 独热编码动作序列，形状 (sequence_length, batch_size, action_dim)
obs: 编码器输出的观察嵌入，形状 (sequence_length, batch_size, embedding_dim)
dones: 终止状态标志
"""
B, T, _ = actions.size() # 用于无观察访问时的推理
# 初始化存储列表
hiddens_list = []
posterior_means_list = []
posterior_logvars_list = []
prior_means_list = []
prior_logvars_list = []
prior_states_list = []
posterior_states_list = []
# 存储初始状态
hiddens_list.append(prev_hidden.unsqueeze(1))
prior_states_list.append(prev_state.unsqueeze(1))
posterior_states_list.append(prev_state.unsqueeze(1))
# 时序展开
for t in range(T - 1):
# 提取当前时刻状态和动作
action_t = actions[:, t, :]
obs_t = obs[:, t, :] if obs is not None else torch.zeros(B, self.embedding_dim, device=actions.device)
state_t = posterior_states_list[-1][:, 0, :] if obs is not None else prior_states_list[-1][:, 0, :]
state_t = state_t if dones is None else state_t * (1 - dones[:, t, :])
hidden_t = hiddens_list[-1][:, 0, :]
# 状态动作组合
state_action = torch.cat([state_t, action_t], dim=-1)
state_action = self.act_fn(self.project_state_action(state_action))
# RNN 状态更新
for i in range(len(self.rnn)):
hidden_t = self.rnn[i](state_action, hidden_t)
# 先验分布计算
hidden_action = torch.cat([hidden_t, action_t], dim=-1)
hidden_action = self.act_fn(self.project_hidden_action(hidden_action))
prior_params = self.prior(hidden_action)
prior_mean, prior_logvar = torch.chunk(prior_params, 2, dim=-1)
# 从先验分布采样
prior_dist = torch.distributions.Normal(prior_mean, torch.exp(F.softplus(prior_logvar)))
prior_state_t = prior_dist.rsample()
# 后验分布计算
if obs is None:
posterior_mean = prior_mean
posterior_logvar = prior_logvar
else:
hidden_obs = torch.cat([hidden_t, obs_t], dim=-1)
hidden_obs = self.act_fn(self.project_hidden_obs(hidden_obs))
posterior_params = self.posterior(hidden_obs)
posterior_mean, posterior_logvar = torch.chunk(posterior_params, 2, dim=-1)
# 从后验分布采样
posterior_dist = torch.distributions.Normal(posterior_mean, torch.exp(F.softplus(posterior_logvar)))
posterior_state_t = posterior_dist.rsample()
# 保存状态
posterior_means_list.append(posterior_mean.unsqueeze(1))
posterior_logvars_list.append(posterior_logvar.unsqueeze(1))
prior_means_list.append(prior_mean.unsqueeze(1))
prior_logvars_list.append(prior_logvar.unsqueeze(1))
prior_states_list.append(prior_state_t.unsqueeze(1))
posterior_states_list.append(posterior_state_t.unsqueeze(1))
hiddens_list.append(hidden_t.unsqueeze(1))
# 合并时序数据
hiddens = torch.cat(hiddens_list, dim=1)
prior_states = torch.cat(prior_states_list, dim=1)
posterior_states = torch.cat(posterior_states_list, dim=1)
prior_means = torch.cat(prior_means_list, dim=1)
prior_logvars = torch.cat(prior_logvars_list, dim=1)
posterior_means = torch.cat(posterior_means_list, dim=1)
posterior_logvars = torch.cat(posterior_logvars_list, dim=1)
return hiddens, prior_states, posterior_states, prior_means, prior_logvars, posterior_means, posterior_logvars需要特别注意的是，这里的观察输入并非原始观察数据，而是经过编码器处理后的嵌入表示。这种设计能够有效降低计算复杂度并提升模型的泛化能力。
RSSM 整体架构
将前述组件整合为完整的 RSSM 模型。其核心是 generate_rollout 方法，负责调用动态模型并生成环境动态的潜在表示序列。对于没有历史潜在状态的情况（通常发生在轨迹开始时），该方法会进行必要的初始化。下面是完整的实现代码：
class RSSM:
def __init__(self,
encoder: EncoderCNN,
decoder: DecoderCNN,
reward_model: RewardModel,
dynamics_model: nn.Module,
hidden_dim: int,
state_dim: int,
action_dim: int,
embedding_dim: int,
device: str = "mps"):
"""
循环状态空间模型（RSSM）实现
参数:
encoder: 确定性状态编码器
decoder: 观察重构解码器
reward_model: 奖励预测模型
dynamics_model: 状态动态模型
hidden_dim: RNN 隐藏层维度
state_dim: 随机状态维度
action_dim: 动作空间维度
embedding_dim: 观察嵌入维度
device: 计算设备
"""
super(RSSM, self).__init__()
# 模型组件初始化
self.dynamics = dynamics_model
self.encoder = encoder
self.decoder = decoder
self.reward_model = reward_model
# 维度参数存储
self.hidden_dim = hidden_dim
self.state_dim = state_dim
self.action_dim = action_dim
self.embedding_dim = embedding_dim
# 模型迁移至指定设备
self.dynamics.to(device)
self.encoder.to(device)
self.decoder.to(device)
self.reward_model.to(device)
def generate_rollout(self, actions: torch.Tensor, hiddens: torch.Tensor = None, states: torch.Tensor = None,
obs: torch.Tensor = None, dones: torch.Tensor = None):
"""
生成状态序列展开
参数:
actions: 动作序列
hiddens: 初始隐藏状态（可选）
states: 初始随机状态（可选）
obs: 观察序列（可选）
dones: 终止标志序列
返回:
完整的状态展开序列
"""
# 状态初始化
if hiddens is None:
hiddens = torch.zeros(actions.size(0), self.hidden_dim).to(actions.device)
if states is None:
states = torch.zeros(actions.size(0), self.state_dim).to(actions.device)
# 执行动态模型展开
dynamics_result = self.dynamics(hiddens, states, actions, obs, dones)
hiddens, prior_states, posterior_states, prior_means, prior_logvars, posterior_means, posterior_logvars = dynamics_result
return hiddens, prior_states, posterior_states, prior_means, prior_logvars, posterior_means, posterior_logvars
def train(self):
"""启用训练模式"""
self.dynamics.train()
self.encoder.train()
self.decoder.train()
self.reward_model.train()
def eval(self):
"""启用评估模式"""
self.dynamics.eval()
self.encoder.eval()
self.decoder.eval()
self.reward_model.eval()
def encode(self, obs: torch.Tensor):
"""观察编码"""
return self.encoder(obs)
def decode(self, state: torch.Tensor):
"""状态解码为观察"""
return self.decoder(state)
def predict_reward(self, h: torch.Tensor, s: torch.Tensor):
"""奖励预测"""
return self.reward_model(h, s)
def parameters(self):
"""返回所有可训练参数"""
return list(self.dynamics.parameters()) + list(self.encoder.parameters()) + \
list(self.decoder.parameters()) + list(self.reward_model.parameters())
def save(self, path: str):
"""模型状态保存"""
torch.save({
"dynamics": self.dynamics.state_dict(),
"encoder": self.encoder.state_dict(),
"decoder": self.decoder.state_dict(),
"reward_model": self.reward_model.state_dict()
}, path)
def load(self, path: str):
"""模型状态加载"""
checkpoint = torch.load(path)
self.dynamics.load_state_dict(checkpoint["dynamics"])
self.encoder.load_state_dict(checkpoint["encoder"])
self.decoder.load_state_dict(checkpoint["decoder"])
self.reward_model.load_state_dict(checkpoint["reward_model"])这个实现提供了一个完整的 RSSM 框架，包含了模型的训练、评估、状态保存和加载等基本功能。该框架可以作为基础结构，根据具体应用场景进行扩展和优化。
训练系统设计
RSSM 的训练系统主要包含两个核心组件：经验回放缓冲区（Experience Replay Buffer）和智能体（Agent）。其中，缓冲区负责存储历史经验数据用于训练，而智能体则作为环境与 RSSM 之间的接口，实现数据收集策略。
经验回放缓冲区实现
缓冲区采用循环队列结构，用于存储和管理观察、动作、奖励和终止状态等数据。通过 sample 方法可以随机采样训练序列。
class Buffer:
def __init__(self, buffer_size: int, obs_shape: tuple, action_shape: tuple, device: torch.device):
"""
经验回放缓冲区初始化
参数:
buffer_size: 缓冲区容量
obs_shape: 观察数据维度
action_shape: 动作数据维度
device: 计算设备
"""
self.buffer_size = buffer_size
self.obs_buffer = np.zeros((buffer_size, *obs_shape), dtype=np.float32)
self.action_buffer = np.zeros((buffer_size, *action_shape), dtype=np.int32)
self.reward_buffer = np.zeros((buffer_size, 1), dtype=np.float32)
self.done_buffer = np.zeros((buffer_size, 1), dtype=np.bool_)
self.device = device
self.idx = 0
def add(self, obs: torch.Tensor, action: int, reward: float, done: bool):
"""
添加单步经验数据
"""
self.obs_buffer[self.idx] = obs
self.action_buffer[self.idx] = action
self.reward_buffer[self.idx] = reward
self.done_buffer[self.idx] = done
self.idx = (self.idx + 1) % self.buffer_size
def sample(self, batch_size: int, sequence_length: int):
"""
随机采样经验序列
参数:
batch_size: 批量大小
sequence_length: 序列长度
返回:
经验数据元组 (observations, actions, rewards, dones)
"""
# 随机选择序列起始位置
starting_idxs = np.random.randint(0, (self.idx % self.buffer_size) - sequence_length, (batch_size,))
# 构建完整序列索引
index_tensor = np.stack([np.arange(start, start + sequence_length) for start in starting_idxs])
# 提取数据序列
obs_sequence = self.obs_buffer[index_tensor]
action_sequence = self.action_buffer[index_tensor]
reward_sequence = self.reward_buffer[index_tensor]
done_sequence = self.done_buffer[index_tensor]
return obs_sequence, action_sequence, reward_sequence, done_sequence
def save(self, path: str):
"""保存缓冲区数据"""
np.savez(path, obs_buffer=self.obs_buffer, action_buffer=self.action_buffer,
reward_buffer=self.reward_buffer, done_buffer=self.done_buffer, idx=self.idx)
def load(self, path: str):
"""加载缓冲区数据"""
data = np.load(path)
self.obs_buffer = data["obs_buffer"]
self.action_buffer = data["action_buffer"]
self.reward_buffer = data["reward_buffer"]
self.done_buffer = data["done_buffer"]
self.idx = data["idx"]智能体设计
智能体实现了数据收集和规划功能。当前实现采用了简单的随机策略进行数据收集，但该框架支持扩展更复杂的策略。
class Policy(ABC):
"""策略基类"""
@abstractmethod
def __call__(self, obs):
pass
class RandomPolicy(Policy):
"""随机采样策略"""
def __init__(self, env: Env):
self.env = env
def __call__(self, obs):
return self.env.action_space.sample()
class Agent:
def __init__(self, env: Env, rssm: RSSM, buffer_size: int = 100000,
collection_policy: str = "random", device="mps"):
"""
智能体初始化
参数:
env: 环境实例
rssm: RSSM模型实例
buffer_size: 经验缓冲区大小
collection_policy: 数据收集策略类型
device: 计算设备
"""
self.env = env
# 策略选择
match collection_policy:
case "random":
self.rollout_policy = RandomPolicy(env)
case _:
raise ValueError("Invalid rollout policy")
self.buffer = Buffer(buffer_size, env.observation_space.shape,
env.action_space.shape, device=device)
self.rssm = rssm
def data_collection_action(self, obs):
"""执行数据收集动作"""
return self.rollout_policy(obs)
def collect_data(self, num_steps: int):
"""
收集训练数据
参数:
num_steps: 收集步数
"""
obs = self.env.reset()
done = False
iterator = tqdm(range(num_steps), desc="Data Collection")
for _ in iterator:
action = self.data_collection_action(obs)
next_obs, reward, done, _, _ = self.env.step(action)
self.buffer.add(next_obs, action, reward, done)
obs = next_obs
if done:
obs = self.env.reset()
def imagine_rollout(self, prev_hidden: torch.Tensor, prev_state: torch.Tensor,
actions: torch.Tensor):
"""
执行想象展开
参数:
prev_hidden: 前一隐藏状态
prev_state: 前一随机状态
actions: 动作序列
返回:
完整的模型输出，包括隐藏状态、先验状态、后验状态等
"""
hiddens, prior_states, posterior_states, prior_means, prior_logvars, \
posterior_means, posterior_logvars = self.rssm.generate_rollout(
actions, prev_hidden, prev_state)
# 在想象阶段使用先验状态预测奖励
rewards = self.rssm.predict_reward(hiddens, prior_states)
return hiddens, prior_states, posterior_states, prior_means, \
prior_logvars, posterior_means, posterior_logvars, rewards
def plan(self, num_steps: int, prev_hidden: torch.Tensor,
prev_state: torch.Tensor, actions: torch.Tensor):
"""
执行规划
参数:
num_steps: 规划步数
prev_hidden: 初始隐藏状态
prev_state: 初始随机状态
actions: 动作序列
返回:
规划得到的隐藏状态和先验状态序列
"""
hidden_states = []
prior_states = []
hiddens = prev_hidden
states = prev_state
for _ in range(num_steps):
hiddens, states, _, _, _, _, _, _ = self.imagine_rollout(
hiddens, states, actions)
hidden_states.append(hiddens)
prior_states.append(states)
hidden_states = torch.stack(hidden_states)
prior_states = torch.stack(prior_states)
return hidden_states, prior_states这部分实现提供了完整的数据管理和智能体交互框架。通过经验回放缓冲区，可以高效地存储和重用历史数据；通过智能体的抽象策略接口，可以方便地扩展不同的数据收集策略。同时智能体还实现了基于模型的想象展开和规划功能，为后续的决策制定提供了基础。
训练器实现与实验
训练器设计
训练器是 RSSM 实现中的最后一个关键组件，负责协调模型训练过程。训练器接收 RSSM 模型、智能体、优化器等组件，并实现具体的训练逻辑。
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s - %(levelname)s - %(message)s",
handlers=[
logging.StreamHandler(), # 控制台输出
logging.FileHandler("training.log", mode="w") # 文件输出
]
)
logger = logging.getLogger(__name__)
class Trainer:
def __init__(self, rssm: RSSM, agent: Agent, optimizer: torch.optim.Optimizer,
device: torch.device):
"""
训练器初始化
参数:
rssm: RSSM 模型实例
agent: 智能体实例
optimizer: 优化器实例
device: 计算设备
"""
self.rssm = rssm
self.optimizer = optimizer
self.device = device
self.agent = agent
self.writer = SummaryWriter() # tensorboard 日志记录器
def train_batch(self, batch_size: int, seq_len: int, iteration: int,
save_images: bool = False):
"""
单批次训练
参数:
batch_size: 批量大小
seq_len: 序列长度
iteration: 当前迭代次数
save_images: 是否保存重建图像
"""
# 采样训练数据
obs, actions, rewards, dones = self.agent.buffer.sample(batch_size, seq_len)
# 数据预处理
actions = torch.tensor(actions).long().to(self.device)
actions = F.one_hot(actions, self.rssm.action_dim).float()
obs = torch.tensor(obs, requires_grad=True).float().to(self.device)
rewards = torch.tensor(rewards, requires_grad=True).float().to(self.device)
dones = torch.tensor(dones).float().to(self.device)
# 观察编码
encoded_obs = self.rssm.encoder(obs.reshape(-1, *obs.shape[2:]).permute(0, 3, 1, 2))
encoded_obs = encoded_obs.reshape(batch_size, seq_len, -1)
# 执行 RSSM 展开
rollout = self.rssm.generate_rollout(actions, obs=encoded_obs, dones=dones)
hiddens, prior_states, posterior_states, prior_means, prior_logvars, \
posterior_means, posterior_logvars = rollout
# 重构观察
hiddens_reshaped = hiddens.reshape(batch_size * seq_len, -1)
posterior_states_reshaped = posterior_states.reshape(batch_size * seq_len, -1)
decoded_obs = self.rssm.decoder(hiddens_reshaped, posterior_states_reshaped)
decoded_obs = decoded_obs.reshape(batch_size, seq_len, *obs.shape[-3:])
# 奖励预测
reward_params = self.rssm.reward_model(hiddens, posterior_states)
mean, logvar = torch.chunk(reward_params, 2, dim=-1)
logvar = F.softplus(logvar)
reward_dist = Normal(mean, torch.exp(logvar))
predicted_rewards = reward_dist.rsample()
# 可视化
if save_images:
batch_idx = np.random.randint(0, batch_size)
seq_idx = np.random.randint(0, seq_len - 3)
fig = self._visualize(obs, decoded_obs, rewards, predicted_rewards,
batch_idx, seq_idx, iteration, grayscale=True)
if not os.path.exists("reconstructions"):
os.makedirs("reconstructions")
fig.savefig(f"reconstructions/iteration_{iteration}.png")
self.writer.add_figure("Reconstructions", fig, iteration)
plt.close(fig)
# 计算损失
reconstruction_loss = self._reconstruction_loss(decoded_obs, obs)
kl_loss = self._kl_loss(prior_means, F.softplus(prior_logvars),
posterior_means, F.softplus(posterior_logvars))
reward_loss = self._reward_loss(rewards, predicted_rewards)
loss = reconstruction_loss + kl_loss + reward_loss
# 反向传播和优化
self.optimizer.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(self.rssm.parameters(), 1, norm_type=2)
self.optimizer.step()
return loss.item(), reconstruction_loss.item(), kl_loss.item(), reward_loss.item()
def train(self, iterations: int, batch_size: int, seq_len: int):
"""
执行完整训练过程
参数:
iterations: 迭代总次数
batch_size: 批量大小
seq_len: 序列长度
"""
self.rssm.train()
iterator = tqdm(range(iterations), desc="Training", total=iterations)
losses = []
infos = []
last_loss = float("inf")
for i in iterator:
# 执行单批次训练
loss, reconstruction_loss, kl_loss, reward_loss = self.train_batch(
batch_size, seq_len, i, save_images=i % 100 == 0)
# 记录训练指标
self.writer.add_scalar("Loss", loss, i)
self.writer.add_scalar("Reconstruction Loss", reconstruction_loss, i)
self.writer.add_scalar("KL Loss", kl_loss, i)
self.writer.add_scalar("Reward Loss", reward_loss, i)
# 保存最佳模型
if loss < last_loss:
self.rssm.save("rssm.pth")
last_loss = loss
# 记录详细信息
info = {
"Loss": loss,
"Reconstruction Loss": reconstruction_loss,
"KL Loss": kl_loss,
"Reward Loss": reward_loss
}
losses.append(loss)
infos.append(info)
# 定期输出训练状态
if i % 10 == 0:
logger.info("\n----------------------------")
logger.info(f"Iteration: {i}")
logger.info(f"Loss: {loss:.4f}")
logger.info(f"Running average last 20 losses: {sum(losses[-20:]) / 20: .4f}")
logger.info(f"Reconstruction Loss: {reconstruction_loss:.4f}")
logger.info(f"KL Loss: {kl_loss:.4f}")
logger.info(f"Reward Loss: {reward_loss:.4f}")
### 实验示例
以下是一个在 CarRacing 环境中训练 RSSM 的完整示例：
```python
# 环境初始化
env = make_env("CarRacing-v2", render_mode="rgb_array", continuous=False, grayscale=True)
# 模型参数设置
hidden_size = 1024
embedding_dim = 1024
state_dim = 512
# 模型组件实例化
encoder = EncoderCNN(in_channels=1, embedding_dim=embedding_dim)
decoder = DecoderCNN(hidden_size=hidden_size, state_size=state_dim,
embedding_size=embedding_dim, output_shape=(1,128,128))
reward_model = RewardModel(hidden_dim=hidden_size, state_dim=state_dim)
dynamics_model = DynamicsModel(hidden_dim=hidden_size, state_dim=state_dim,
action_dim=5, embedding_dim=embedding_dim)
# RSSM 模型构建
rssm = RSSM(dynamics_model=dynamics_model,
encoder=encoder,
decoder=decoder,
reward_model=reward_model,
hidden_dim=hidden_size,
state_dim=state_dim,
action_dim=5,
embedding_dim=embedding_dim)
# 训练设置
optimizer = torch.optim.Adam(rssm.parameters(), lr=1e-3)
agent = Agent(env, rssm)
trainer = Trainer(rssm, agent, optimizer=optimizer, device="cuda")
# 数据收集和训练
trainer.collect_data(20000) # 收集 20000 步经验数据
trainer.save_buffer("buffer.npz") # 保存经验缓冲区
trainer.train(10000, 32, 20) # 执行 10000 次迭代训练总结
本文详细介绍了基于 PyTorch 实现 RSSM 的完整过程。RSSM 的架构相比传统的 VAE 或 RNN 更为复杂，这主要源于其混合了随机和确定性状态的特性。通过手动实现这一架构，我们可以深入理解其背后的理论基础及其强大之处。RSSM 能够递归地生成未来潜在状态轨迹，这为智能体的行为规划提供了基础。
实现的优点在于其计算负载适中，可以在单个消费级 GPU 上进行训练，在有充足时间的情况下甚至可以在 CPU 上运行。这一工作基于论文《Learning Latent Dynamics for Planning from Pixels》，该论文为 RSSM 类动态模型奠定了基础。后续的研究工作如《Dream to Control: Learning Behaviors by Latent Imagination》进一步发展了这一架构。这些改进的架构将在未来的研究中深入探讨，因为它们对理解 MBRL 方法提供了重要的见解。
作者：Lukas Bierling