评价此页
★ ★ ★ ★ ★
intermediate/reinforcement_ppo

注意

跳转至页面底部 下载完整示例代码。

使用 TorchRL 进行强化学习 (PPO) 教程#

创建日期：2023年3月15日 | 最后更新：2025年9月17日 | 最后验证：2024年11月5日

作者：Vincent Moens

本教程演示了如何使用 PyTorch 和 torchrl 训练一个参数化策略网络，以解决来自 OpenAI-Gym/Farama-Gymnasium 控制库 的倒立摆（Inverted Pendulum）任务。

Inverted pendulum

倒立摆#

主要学习内容

  • 如何创建 TorchRL 环境、转换其输出并从该环境中收集数据；

  • 如何使用 TensorDict 实现类之间的交互；

  • 使用 TorchRL 构建训练循环的基础知识；

    • 如何计算策略梯度方法的优势（advantage）信号；

    • 如何使用概率神经网络创建随机策略；

    • 如何创建动态经验回放缓冲区并从中进行无重复采样。

我们将涵盖 TorchRL 的六个关键组件：

如果您在 Google Colab 中运行此代码，请确保安装以下依赖项

!pip3 install torchrl
!pip3 install gym[mujoco]
!pip3 install tqdm

近端策略优化（PPO）是一种策略梯度算法，它收集并直接消耗一批数据，以在存在某些近端约束的情况下最大化预期回报来训练策略。您可以将其视为 REINFORCE（基础策略优化算法）的复杂版本。有关更多信息，请参阅 近端策略优化算法论文。

PPO 通常被认为是一种快速高效的在线（online）、同策略（on-policy）强化学习算法。TorchRL 提供了一个可以为您处理所有工作的损失模块，因此您可以依赖此实现，并专注于解决您的问题，而不是每次想要训练策略时都重复造轮子。

为了完整起见，这里简要概述了该损失函数的计算内容，尽管这已由我们的 ClipPPOLoss 模块处理——该算法的工作原理如下：1. 我们通过在环境中运行策略一定步数来采样一批数据。2. 然后，我们使用该批数据的随机子样本，通过 REINFORCE 损失的裁剪版本进行一定次数的优化步骤。3. 裁剪将为我们的损失设置一个悲观边界：相比于较高的回报估计，较低的回报估计将受到青睐。该损失的精确公式为：

\[L(s,a,\theta_k,\theta) = \min\left( \frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)} A^{\pi_{\theta_k}}(s,a), \;\; g(\epsilon, A^{\pi_{\theta_k}}(s,a)) \right),\]

该损失包含两个部分：在最小运算符的第一部分，我们简单地计算了 REINFORCE 损失的权重重要性版本（例如，针对当前策略配置滞后于用于数据收集的配置这一事实进行了修正的 REINFORCE 损失）。最小运算符的第二部分是一个类似的损失，其中当比率超过或低于给定的一对阈值时，我们对其进行了裁剪。

此损失确保无论优势是正数还是负数，都会抑制可能导致与先前配置产生显著偏差的策略更新。

本教程结构如下

  1. 首先，我们将定义一组用于训练的超参数。

  2. 接下来，我们将重点使用 TorchRL 的包装器（wrappers）和转换（transforms）来创建我们的环境或模拟器。

  3. 接下来，我们将设计策略网络和价值模型，这对于损失函数是必不可少的。这些模块将用于配置我们的损失模块。

  4. 接下来，我们将创建经验回放缓冲区和数据加载器。

  5. 最后，我们将运行训练循环并分析结果。

在本教程中，我们将使用 tensordict 库。TensorDict 是 TorchRL 的通用语言：它帮助我们抽象模块的读写内容，让我们更关注算法本身，而不是特定的数据描述。

import warnings
warnings.filterwarnings("ignore")
from torch import multiprocessing


from collections import defaultdict

import matplotlib.pyplot as plt
import torch
from tensordict.nn import TensorDictModule
from tensordict.nn.distributions import NormalParamExtractor
from torch import nn
from torchrl.collectors import SyncDataCollector
from torchrl.data.replay_buffers import ReplayBuffer
from torchrl.data.replay_buffers.samplers import SamplerWithoutReplacement
from torchrl.data.replay_buffers.storages import LazyTensorStorage
from torchrl.envs import (Compose, DoubleToFloat, ObservationNorm, StepCounter,
                          TransformedEnv)
from torchrl.envs.libs.gym import GymEnv
from torchrl.envs.utils import check_env_specs, ExplorationType, set_exploration_type
from torchrl.modules import ProbabilisticActor, TanhNormal, ValueOperator
from torchrl.objectives import ClipPPOLoss
from torchrl.objectives.value import GAE
from tqdm import tqdm

定义超参数#

我们为算法设置超参数。根据可用的资源，可以选择在 GPU 或其他设备上执行策略。frame_skip 将控制单个动作执行的帧数。其余计算帧数的参数必须针对该值进行修正（因为一个环境步骤实际上会返回 frame_skip 帧）。

is_fork = multiprocessing.get_start_method() == "fork"
device = (
    torch.device(0)
    if torch.cuda.is_available() and not is_fork
    else torch.device("cpu")
)
num_cells = 256  # number of cells in each layer i.e. output dim.
lr = 3e-4
max_grad_norm = 1.0

数据收集参数#

在收集数据时，我们可以通过定义 frames_per_batch 参数来选择每个批次的大小。我们还将定义允许使用的帧数（例如与模拟器的交互次数）。通常，RL 算法的目标是尽可能快地通过环境交互来学习解决任务：total_frames 越低越好。

frames_per_batch = 1000
# For a complete training, bring the number of frames up to 1M
total_frames = 50_000

PPO 参数#

在每次数据收集（或批次收集）时，我们将在一定的 epochs 内进行优化，每次都会消耗我们在嵌套训练循环中刚刚获取的全部数据。在这里，sub_batch_size 与上面的 frames_per_batch 不同：请记住，我们正在处理来自收集器的一“批数据”，其大小由 frames_per_batch 定义，并且我们将在内部训练循环中将其进一步拆分为更小的子批次。这些子批次的大小由 sub_batch_size 控制。 

sub_batch_size = 64  # cardinality of the sub-samples gathered from the current data in the inner loop
num_epochs = 10  # optimization steps per batch of data collected
clip_epsilon = (
    0.2  # clip value for PPO loss: see the equation in the intro for more context.
)
gamma = 0.99
lmbda = 0.95
entropy_eps = 1e-4

定义环境#

在 RL 中，环境 通常是我们对模拟器或控制系统的称呼。各种库为强化学习提供了模拟环境，包括 Gymnasium（原 OpenAI Gym）、DeepMind control suite 等。作为一个通用库，TorchRL 的目标是为大量 RL 模拟器提供可互换的接口，使您可以轻松地更换环境。例如，创建包装好的 gym 环境只需很少的代码：

base_env = GymEnv("InvertedDoublePendulum-v4", device=device)

Gym has been unmaintained since 2022 and does not support NumPy 2.0 amongst other critical functionality.
Please upgrade to Gymnasium, the maintained drop-in replacement of Gym, or contact the authors of your software and request that they upgrade.
Users of this version of Gym should be able to simply replace 'import gym' with 'import gymnasium as gym' in the vast majority of cases.
See the migration guide at https://gymnasium.org.cn/introduction/migration_guide/ for additional information.

此代码中有几点需要注意：首先，我们通过调用 GymEnv 包装器创建了环境。如果传递了额外的关键字参数，它们将被传输到 gym.make 方法，从而涵盖最常见的环境构建命令。或者，也可以直接使用 gym.make(env_name, **kwargs) 创建 gym 环境，并将其包装在 GymWrapper 类中。

还有 device 参数：对于 gym，这仅控制存储输入动作和观测状态的设备，但执行始终会在 CPU 上完成。原因很简单，除非另有说明，否则 gym 不支持设备端执行。对于其他库，我们可以控制执行设备，并且在尽可能的情况下，我们力求在存储和执行后端方面保持一致。

转换#

我们将向环境添加一些转换（transforms），为策略准备数据。在 Gym 中，这通常通过包装器实现。TorchRL 采用了不同的方法，通过使用转换，这更类似于其他 PyTorch 领域库。要向环境添加转换，只需将其包装在 TransformedEnv 实例中，并将转换序列附加到其中即可。转换后的环境将继承包装环境的设备和元数据，并根据其包含的转换序列对这些数据进行转换。

归一化#

首先要编码的是归一化转换。根据经验，最好使数据大致符合单位高斯分布：为此，我们将执行一定数量的随机步骤，并计算这些观测值的统计摘要。

我们将添加另外两个转换：DoubleToFloat 转换会将双精度条目转换为单精度数字，以便策略读取。StepCounter 转换将用于计算环境终止前的步骤。我们将使用该指标作为性能的补充衡量标准。

正如我们稍后将看到的，许多 TorchRL 的类都依赖 TensorDict 进行通信。您可以将其视为一个带有额外张量功能的 Python 字典。在实践中，这意味着我们使用的许多模块需要被告知在接收到的 tensordict 中读取哪些键（in_keys）以及写入哪些键（out_keys）。通常，如果省略 out_keys，则假定会就地（in-place）更新 in_keys 条目。对于我们的转换，我们感兴趣的唯一条目被称为 "observation"，并且我们的转换层将被告知仅修改该条目。

env = TransformedEnv(
    base_env,
    Compose(
        # normalize observations
        ObservationNorm(in_keys=["observation"]),
        DoubleToFloat(),
        StepCounter(),
    ),
)

正如您可能注意到的，我们创建了一个归一化层，但没有设置其归一化参数。为此，ObservationNorm 可以自动收集我们环境的统计摘要。

env.transform[0].init_stats(num_iter=1000, reduce_dim=0, cat_dim=0)

ObservationNorm 转换现在已填充了将用于归一化数据的均值和标准差。

让我们对统计摘要的形状进行一点完整性检查。

print("normalization constant shape:", env.transform[0].loc.shape)

normalization constant shape: torch.Size([11])

环境不仅由其模拟器和转换定义，还由一系列描述执行过程中预期的元数据定义。出于效率考虑，TorchRL 在环境规格方面非常严格，但您可以轻松检查环境规格是否合适。在我们的示例中，继承自该规格的 GymWrapper 和 GymEnv 已经为您处理了设置适当规格的工作，因此您不必担心这一点。

尽管如此，让我们通过查看其规格来看一个具体的示例。有三个规格需要查看：observation_spec 定义了在环境中执行动作时的预期内容，reward_spec 指示奖励域，最后是 input_spec（包含 action_spec），代表环境执行单个步骤所需的一切。

print("observation_spec:", env.observation_spec)
print("reward_spec:", env.reward_spec)
print("input_spec:", env.input_spec)
print("action_spec (as defined by input_spec):", env.action_spec)

observation_spec: Composite(
    observation: UnboundedContinuous(
        shape=torch.Size([11]),
        space=ContinuousBox(
            low=Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, contiguous=True),
            high=Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, contiguous=True)),
        device=cpu,
        dtype=torch.float32,
        domain=continuous),
    step_count: BoundedDiscrete(
        shape=torch.Size([1]),
        space=ContinuousBox(
            low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, contiguous=True),
            high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, contiguous=True)),
        device=cpu,
        dtype=torch.int64,
        domain=discrete),
    device=cpu,
    shape=torch.Size([]),
    data_cls=None)
reward_spec: UnboundedContinuous(
    shape=torch.Size([1]),
    space=ContinuousBox(
        low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True),
        high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True)),
    device=cpu,
    dtype=torch.float32,
    domain=continuous)
input_spec: Composite(
    full_state_spec: Composite(
        step_count: BoundedDiscrete(
            shape=torch.Size([1]),
            space=ContinuousBox(
                low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, contiguous=True),
                high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, contiguous=True)),
            device=cpu,
            dtype=torch.int64,
            domain=discrete),
        device=cpu,
        shape=torch.Size([]),
        data_cls=None),
    full_action_spec: Composite(
        action: BoundedContinuous(
            shape=torch.Size([1]),
            space=ContinuousBox(
                low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True),
                high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True)),
            device=cpu,
            dtype=torch.float32,
            domain=continuous),
        device=cpu,
        shape=torch.Size([]),
        data_cls=None),
    device=cpu,
    shape=torch.Size([]),
    data_cls=None)
action_spec (as defined by input_spec): BoundedContinuous(
    shape=torch.Size([1]),
    space=ContinuousBox(
        low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True),
        high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True)),
    device=cpu,
    dtype=torch.float32,
    domain=continuous)

该 check_env_specs() 函数运行一个小的回滚，并将其输出与环境规范进行比较。如果没有引发错误，我们可以确信规范已正确定义。

check_env_specs(env)

2026-07-08 22:28:00,565 [torchrl][INFO]    check_env_specs succeeded! [END]

有趣的是，让我们看看简单的随机展开（rollout）是什么样的。您可以调用 env.rollout(n_steps) 并概览环境输入和输出的样子。动作将自动从动作规格域中抽取，因此您无需考虑设计随机采样器。

通常，在每一步中，RL 环境都会接收一个动作作为输入，并输出观测值、奖励和完成状态。观测值可能是复合的，这意味着它可能由多个张量组成。这对 TorchRL 来说不是问题，因为整组观测值会自动打包在输出的 TensorDict 中。在执行一定步数的展开（例如，一系列环境步骤和随机动作生成）后，我们将检索到一个与该轨迹长度形状匹配的 TensorDict 实例。

rollout = env.rollout(3)
print("rollout of three steps:", rollout)
print("Shape of the rollout TensorDict:", rollout.batch_size)

rollout of three steps: TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([3, 11]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                step_count: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.int64, is_shared=False),
                terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([3]),
            device=cpu,
            is_shared=False),
        observation: Tensor(shape=torch.Size([3, 11]), device=cpu, dtype=torch.float32, is_shared=False),
        step_count: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.int64, is_shared=False),
        terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([3]),
    device=cpu,
    is_shared=False)
Shape of the rollout TensorDict: torch.Size([3])

我们的展开数据形状为 torch.Size([3])，这与我们运行的步数相匹配。"next" 条目指向当前步骤之后的数据。在大多数情况下，时间 t 的 "next" 数据与 t+1 的数据相匹配，但如果我们使用某些特定的转换（例如多步转换），情况可能并非如此。

策略 (Policy)#

PPO 利用随机策略来处理探索。这意味着我们的神经网络必须输出分布的参数，而不是对应于所采取动作的单个值。

由于数据是连续的，我们使用 Tanh-Normal 分布来遵守动作空间边界。TorchRL 提供了这样的分布，我们唯一需要关心的是构建一个输出正确参数数量（位置或均值，以及尺度或标准差）的神经网络。

\[f_{\theta}(\text{observation}) = \mu_{\theta}(\text{observation}), \sigma^{+}_{\theta}(\text{observation})\]

这里带来的唯一额外困难是将输出分成两等份，并将第二部分映射到严格正的空间。

我们分三步设计策略：

  1. 定义神经网络 D_obs -> 2 * D_action。实际上，我们的 loc (mu) 和 scale (sigma) 都具有维度 D_action。

  2. 附加一个 NormalParamExtractor 来提取位置和尺度（例如，将输入分成两等份，并对尺度参数应用正变换）。

  3. 创建一个可以生成该分布并从中采样的概率 TensorDictModule。

actor_net = nn.Sequential(
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(2 * env.action_spec.shape[-1], device=device),
    NormalParamExtractor(),
)

为了使策略能够通过 tensordict 数据载体与环境“交流”，我们将 nn.Module 包装在 TensorDictModule 中。该类将简单地读取提供给它的 in_keys，并将输出原位写入注册的 out_keys。

policy_module = TensorDictModule(
    actor_net, in_keys=["observation"], out_keys=["loc", "scale"]
)

我们现在需要基于正态分布的位置和尺度构建一个分布。为此，我们指示 ProbabilisticActor 类基于位置和尺度参数构建一个 TanhNormal 分布。我们还提供从环境规格中收集的分布最小值和最大值。

上述 TensorDictModule 的 in_keys（以及因此的 out_keys）名称不能随意设置，因为 TanhNormal 分布构造函数将需要 loc 和 scale 关键字参数。话虽如此，ProbabilisticActor 也接受 Dict[str, str] 类型的 in_keys，其中键值对指示应为每个要使用的关键字参数使用哪个 in_key 字符串。

policy_module = ProbabilisticActor(
    module=policy_module,
    spec=env.action_spec,
    in_keys=["loc", "scale"],
    distribution_class=TanhNormal,
    distribution_kwargs={
        "low": env.action_spec.space.low,
        "high": env.action_spec.space.high,
    },
    return_log_prob=True,
    # we'll need the log-prob for the numerator of the importance weights
)

价值网络#

价值网络是 PPO 算法的关键组件，尽管它在推理时不会使用。该模块将读取观测值并返回后续轨迹的折现回报估计。这允许我们通过依赖在训练期间动态学习的效用估计来分摊学习成本。我们的价值网络与策略具有相同的结构，但为了简单起见，我们为其分配了自己的一组参数。

value_net = nn.Sequential(
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(num_cells, device=device),
    nn.Tanh(),
    nn.LazyLinear(1, device=device),
)

value_module = ValueOperator(
    module=value_net,
    in_keys=["observation"],
)

让我们尝试一下我们的策略和价值模块。正如我们前面所说，TensorDictModule 的使用使得直接读取环境输出以运行这些模块成为可能，因为它们知道要读取什么信息以及在哪里写入它。

print("Running policy:", policy_module(env.reset()))
print("Running value:", value_module(env.reset()))

Running policy: TensorDict(
    fields={
        action: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
        action_log_prob: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        loc: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, is_shared=False),
        scale: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
        step_count: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, is_shared=False),
        terminated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([]),
    device=cpu,
    is_shared=False)
Running value: TensorDict(
    fields={
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        observation: Tensor(shape=torch.Size([11]), device=cpu, dtype=torch.float32, is_shared=False),
        state_value: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
        step_count: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, is_shared=False),
        terminated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([]),
    device=cpu,
    is_shared=False)

数据收集器#

TorchRL 提供了一组 数据收集器类。简而言之，这些类执行三个操作：重置环境、根据最新观测值计算动作、在环境中执行一步，并重复后两个步骤，直到环境发出停止信号（或达到完成状态）。

它们允许您控制每次迭代收集多少帧（通过 frames_per_batch 参数）、何时重置环境（通过 max_frames_per_traj 参数）、策略应在哪个 device 上执行等。它们还旨在与批处理和多进程环境高效协同工作。

最简单的数据收集器是 SyncDataCollector：它是一个迭代器，您可以使用它获取给定长度的数据批次，并在收集到总帧数（total_frames）后停止。其他数据收集器（MultiSyncDataCollector 和 MultiaSyncDataCollector）将在多进程工作器集上以同步和异步方式执行相同的操作。

与之前的策略和环境一样，数据收集器将返回 TensorDict 实例，其元素总数将与 frames_per_batch 匹配。使用 TensorDict 将数据传递到训练循环中，使您能够编写完全不关心展开内容具体细节的数据加载管道。 

collector = SyncDataCollector(
    env,
    policy_module,
    frames_per_batch=frames_per_batch,
    total_frames=total_frames,
    split_trajs=False,
    device=device,
)

回放缓冲区#

回放缓冲区是离策略 RL 算法的常见构建模块。在策略环境中，每当收集一批数据时，回放缓冲区就会被重新填充，并且其数据会在一定数量的 epoch 中被重复消耗。

TorchRL 的经验回放缓冲区是使用通用容器 ReplayBuffer 构建的，该容器将缓冲区的组件作为参数：存储、写入器、采样器以及可能的转换。只有存储（表示回放缓冲区的容量）是强制性的。我们还指定了一个无重复采样器，以避免在一个 epoch 中多次采样同一个项目。将经验回放缓冲区用于 PPO 不是强制性的，我们可以简单地从收集到的批次中采样子批次，但使用这些类使我们能够以可重复的方式轻松构建内部训练循环。

replay_buffer = ReplayBuffer(
    storage=LazyTensorStorage(max_size=frames_per_batch),
    sampler=SamplerWithoutReplacement(),
)

损失函数#

PPO 损失可以使用 ClipPPOLoss 类直接从 TorchRL 导入。这是使用 PPO 最简单的方法：它隐藏了 PPO 的数学运算及其附带的控制流。

PPO 需要计算一些“优势估计”。简而言之，优势是一个值，它在处理偏差/方差权衡时反映了回报值的期望。要计算优势，只需（1）构建利用我们价值运算符的优势模块，以及（2）在每个 epoch 之前将每批数据通过它。GAE 模块将使用新的 "advantage" 和 "value_target" 条目更新输入 tensordict。"value_target" 是一个无梯度的张量，表示价值网络应使用输入观测值表示的经验价值。这两者都将由 ClipPPOLoss 使用，以返回策略和价值损失。

advantage_module = GAE(
    gamma=gamma, lmbda=lmbda, value_network=value_module, average_gae=True, device=device,
)

loss_module = ClipPPOLoss(
    actor_network=policy_module,
    critic_network=value_module,
    clip_epsilon=clip_epsilon,
    entropy_bonus=bool(entropy_eps),
    entropy_coef=entropy_eps,
    # these keys match by default but we set this for completeness
    critic_coef=1.0,
    loss_critic_type="smooth_l1",
)

optim = torch.optim.Adam(loss_module.parameters(), lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
    optim, total_frames // frames_per_batch, 0.0
)

/usr/local/lib/python3.10/dist-packages/torchrl/objectives/ppo.py:445: DeprecationWarning: 'critic_coef' is deprecated and will be removed in torchrl v0.11. Please use 'critic_coeff' instead.
  warnings.warn(
/usr/local/lib/python3.10/dist-packages/torchrl/objectives/ppo.py:511: DeprecationWarning: 'entropy_coef' is deprecated and will be removed in torchrl v0.11. Please use 'entropy_coeff' instead.
  warnings.warn(

训练循环#

现在我们有了编写训练循环所需的所有组件。步骤包括

  • 收集数据

    • 计算优势

      • 循环收集的数据以计算损失值

      • 反向传播

      • 优化

      • 重复

    • 重复

  • 重复

logs = defaultdict(list)
pbar = tqdm(total=total_frames)
eval_str = ""

# We iterate over the collector until it reaches the total number of frames it was
# designed to collect:
for i, tensordict_data in enumerate(collector):
    # we now have a batch of data to work with. Let's learn something from it.
    for _ in range(num_epochs):
        # We'll need an "advantage" signal to make PPO work.
        # We re-compute it at each epoch as its value depends on the value
        # network which is updated in the inner loop.
        advantage_module(tensordict_data)
        data_view = tensordict_data.reshape(-1)
        replay_buffer.extend(data_view.cpu())
        for _ in range(frames_per_batch // sub_batch_size):
            subdata = replay_buffer.sample(sub_batch_size)
            loss_vals = loss_module(subdata.to(device))
            loss_value = (
                loss_vals["loss_objective"]
                + loss_vals["loss_critic"]
                + loss_vals["loss_entropy"]
            )

            # Optimization: backward, grad clipping and optimization step
            loss_value.backward()
            # this is not strictly mandatory but it's good practice to keep
            # your gradient norm bounded
            torch.nn.utils.clip_grad_norm_(loss_module.parameters(), max_grad_norm)
            optim.step()
            optim.zero_grad()

    logs["reward"].append(tensordict_data["next", "reward"].mean().item())
    pbar.update(tensordict_data.numel())
    cum_reward_str = (
        f"average reward={logs['reward'][-1]: 4.4f} (init={logs['reward'][0]: 4.4f})"
    )
    logs["step_count"].append(tensordict_data["step_count"].max().item())
    stepcount_str = f"step count (max): {logs['step_count'][-1]}"
    logs["lr"].append(optim.param_groups[0]["lr"])
    lr_str = f"lr policy: {logs['lr'][-1]: 4.4f}"
    if i % 10 == 0:
        # We evaluate the policy once every 10 batches of data.
        # Evaluation is rather simple: execute the policy without exploration
        # (take the expected value of the action distribution) for a given
        # number of steps (1000, which is our ``env`` horizon).
        # The ``rollout`` method of the ``env`` can take a policy as argument:
        # it will then execute this policy at each step.
        with set_exploration_type(ExplorationType.DETERMINISTIC), torch.no_grad():
            # execute a rollout with the trained policy
            eval_rollout = env.rollout(1000, policy_module)
            logs["eval reward"].append(eval_rollout["next", "reward"].mean().item())
            logs["eval reward (sum)"].append(
                eval_rollout["next", "reward"].sum().item()
            )
            logs["eval step_count"].append(eval_rollout["step_count"].max().item())
            eval_str = (
                f"eval cumulative reward: {logs['eval reward (sum)'][-1]: 4.4f} "
                f"(init: {logs['eval reward (sum)'][0]: 4.4f}), "
                f"eval step-count: {logs['eval step_count'][-1]}"
            )
            del eval_rollout
    pbar.set_description(", ".join([eval_str, cum_reward_str, stepcount_str, lr_str]))

    # We're also using a learning rate scheduler. Like the gradient clipping,
    # this is a nice-to-have but nothing necessary for PPO to work.
    scheduler.step()

  0%|          | 0/50000 [00:00<?, ?it/s]
  2%|▏         | 1000/50000 [00:03<02:29, 327.70it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.0952 (init= 9.0952), step count (max): 13, lr policy:  0.0003:   2%|▏         | 1000/50000 [00:03<02:29, 327.70it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.0952 (init= 9.0952), step count (max): 13, lr policy:  0.0003:   4%|▍         | 2000/50000 [00:06<02:26, 326.55it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1174 (init= 9.0952), step count (max): 12, lr policy:  0.0003:   4%|▍         | 2000/50000 [00:06<02:26, 326.55it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1174 (init= 9.0952), step count (max): 12, lr policy:  0.0003:   6%|▌         | 3000/50000 [00:08<02:19, 336.36it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1532 (init= 9.0952), step count (max): 16, lr policy:  0.0003:   6%|▌         | 3000/50000 [00:08<02:19, 336.36it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1532 (init= 9.0952), step count (max): 16, lr policy:  0.0003:   8%|▊         | 4000/50000 [00:11<02:14, 342.01it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1788 (init= 9.0952), step count (max): 24, lr policy:  0.0003:   8%|▊         | 4000/50000 [00:11<02:14, 342.01it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.1788 (init= 9.0952), step count (max): 24, lr policy:  0.0003:  10%|█         | 5000/50000 [00:14<02:10, 345.67it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2062 (init= 9.0952), step count (max): 21, lr policy:  0.0003:  10%|█         | 5000/50000 [00:14<02:10, 345.67it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2062 (init= 9.0952), step count (max): 21, lr policy:  0.0003:  12%|█▏        | 6000/50000 [00:17<02:06, 348.69it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2284 (init= 9.0952), step count (max): 39, lr policy:  0.0003:  12%|█▏        | 6000/50000 [00:17<02:06, 348.69it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2284 (init= 9.0952), step count (max): 39, lr policy:  0.0003:  14%|█▍        | 7000/50000 [00:20<02:02, 350.70it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2333 (init= 9.0952), step count (max): 30, lr policy:  0.0003:  14%|█▍        | 7000/50000 [00:20<02:02, 350.70it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2333 (init= 9.0952), step count (max): 30, lr policy:  0.0003:  16%|█▌        | 8000/50000 [00:23<01:59, 352.24it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2317 (init= 9.0952), step count (max): 29, lr policy:  0.0003:  16%|█▌        | 8000/50000 [00:23<01:59, 352.24it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2317 (init= 9.0952), step count (max): 29, lr policy:  0.0003:  18%|█▊        | 9000/50000 [00:26<01:57, 347.55it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2344 (init= 9.0952), step count (max): 40, lr policy:  0.0003:  18%|█▊        | 9000/50000 [00:26<01:57, 347.55it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2344 (init= 9.0952), step count (max): 40, lr policy:  0.0003:  20%|██        | 10000/50000 [00:28<01:54, 350.23it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2310 (init= 9.0952), step count (max): 48, lr policy:  0.0003:  20%|██        | 10000/50000 [00:28<01:54, 350.23it/s]
eval cumulative reward:  91.8148 (init:  91.8148), eval step-count: 9, average reward= 9.2310 (init= 9.0952), step count (max): 48, lr policy:  0.0003:  22%|██▏       | 11000/50000 [00:31<01:50, 352.40it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2550 (init= 9.0952), step count (max): 54, lr policy:  0.0003:  22%|██▏       | 11000/50000 [00:31<01:50, 352.40it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2550 (init= 9.0952), step count (max): 54, lr policy:  0.0003:  24%|██▍       | 12000/50000 [00:34<01:47, 352.37it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2447 (init= 9.0952), step count (max): 41, lr policy:  0.0003:  24%|██▍       | 12000/50000 [00:34<01:47, 352.37it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2447 (init= 9.0952), step count (max): 41, lr policy:  0.0003:  26%|██▌       | 13000/50000 [00:37<01:44, 354.27it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2631 (init= 9.0952), step count (max): 54, lr policy:  0.0003:  26%|██▌       | 13000/50000 [00:37<01:44, 354.27it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2631 (init= 9.0952), step count (max): 54, lr policy:  0.0003:  28%|██▊       | 14000/50000 [00:40<01:41, 355.48it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2747 (init= 9.0952), step count (max): 69, lr policy:  0.0003:  28%|██▊       | 14000/50000 [00:40<01:41, 355.48it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2747 (init= 9.0952), step count (max): 69, lr policy:  0.0003:  30%|███       | 15000/50000 [00:42<01:38, 356.29it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2754 (init= 9.0952), step count (max): 81, lr policy:  0.0002:  30%|███       | 15000/50000 [00:42<01:38, 356.29it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2754 (init= 9.0952), step count (max): 81, lr policy:  0.0002:  32%|███▏      | 16000/50000 [00:45<01:35, 356.99it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2679 (init= 9.0952), step count (max): 45, lr policy:  0.0002:  32%|███▏      | 16000/50000 [00:45<01:35, 356.99it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2679 (init= 9.0952), step count (max): 45, lr policy:  0.0002:  34%|███▍      | 17000/50000 [00:48<01:33, 351.85it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2775 (init= 9.0952), step count (max): 58, lr policy:  0.0002:  34%|███▍      | 17000/50000 [00:48<01:33, 351.85it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2775 (init= 9.0952), step count (max): 58, lr policy:  0.0002:  36%|███▌      | 18000/50000 [00:51<01:30, 353.66it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2641 (init= 9.0952), step count (max): 51, lr policy:  0.0002:  36%|███▌      | 18000/50000 [00:51<01:30, 353.66it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2641 (init= 9.0952), step count (max): 51, lr policy:  0.0002:  38%|███▊      | 19000/50000 [00:54<01:27, 355.21it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2781 (init= 9.0952), step count (max): 73, lr policy:  0.0002:  38%|███▊      | 19000/50000 [00:54<01:27, 355.21it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2781 (init= 9.0952), step count (max): 73, lr policy:  0.0002:  40%|████      | 20000/50000 [00:56<01:24, 356.34it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2817 (init= 9.0952), step count (max): 58, lr policy:  0.0002:  40%|████      | 20000/50000 [00:56<01:24, 356.34it/s]
eval cumulative reward:  251.3744 (init:  91.8148), eval step-count: 26, average reward= 9.2817 (init= 9.0952), step count (max): 58, lr policy:  0.0002:  42%|████▏     | 21000/50000 [00:59<01:21, 357.28it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2867 (init= 9.0952), step count (max): 63, lr policy:  0.0002:  42%|████▏     | 21000/50000 [00:59<01:21, 357.28it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2867 (init= 9.0952), step count (max): 63, lr policy:  0.0002:  44%|████▍     | 22000/50000 [01:02<01:18, 356.06it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2745 (init= 9.0952), step count (max): 41, lr policy:  0.0002:  44%|████▍     | 22000/50000 [01:02<01:18, 356.06it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2745 (init= 9.0952), step count (max): 41, lr policy:  0.0002:  46%|████▌     | 23000/50000 [01:05<01:15, 356.79it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2752 (init= 9.0952), step count (max): 50, lr policy:  0.0002:  46%|████▌     | 23000/50000 [01:05<01:15, 356.79it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2752 (init= 9.0952), step count (max): 50, lr policy:  0.0002:  48%|████▊     | 24000/50000 [01:08<01:14, 351.15it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2793 (init= 9.0952), step count (max): 62, lr policy:  0.0002:  48%|████▊     | 24000/50000 [01:08<01:14, 351.15it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2793 (init= 9.0952), step count (max): 62, lr policy:  0.0002:  50%|█████     | 25000/50000 [01:11<01:10, 353.24it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2772 (init= 9.0952), step count (max): 63, lr policy:  0.0002:  50%|█████     | 25000/50000 [01:11<01:10, 353.24it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2772 (init= 9.0952), step count (max): 63, lr policy:  0.0002:  52%|█████▏    | 26000/50000 [01:13<01:07, 354.83it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2732 (init= 9.0952), step count (max): 58, lr policy:  0.0001:  52%|█████▏    | 26000/50000 [01:13<01:07, 354.83it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2732 (init= 9.0952), step count (max): 58, lr policy:  0.0001:  54%|█████▍    | 27000/50000 [01:16<01:04, 355.90it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2794 (init= 9.0952), step count (max): 57, lr policy:  0.0001:  54%|█████▍    | 27000/50000 [01:16<01:04, 355.90it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2794 (init= 9.0952), step count (max): 57, lr policy:  0.0001:  56%|█████▌    | 28000/50000 [01:19<01:01, 356.62it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2685 (init= 9.0952), step count (max): 53, lr policy:  0.0001:  56%|█████▌    | 28000/50000 [01:19<01:01, 356.62it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2685 (init= 9.0952), step count (max): 53, lr policy:  0.0001:  58%|█████▊    | 29000/50000 [01:22<00:58, 357.37it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2667 (init= 9.0952), step count (max): 48, lr policy:  0.0001:  58%|█████▊    | 29000/50000 [01:22<00:58, 357.37it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2667 (init= 9.0952), step count (max): 48, lr policy:  0.0001:  60%|██████    | 30000/50000 [01:25<00:55, 358.14it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2666 (init= 9.0952), step count (max): 48, lr policy:  0.0001:  60%|██████    | 30000/50000 [01:25<00:55, 358.14it/s]
eval cumulative reward:  249.6714 (init:  91.8148), eval step-count: 26, average reward= 9.2666 (init= 9.0952), step count (max): 48, lr policy:  0.0001:  62%|██████▏   | 31000/50000 [01:27<00:52, 358.91it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2815 (init= 9.0952), step count (max): 67, lr policy:  0.0001:  62%|██████▏   | 31000/50000 [01:27<00:52, 358.91it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2815 (init= 9.0952), step count (max): 67, lr policy:  0.0001:  64%|██████▍   | 32000/50000 [01:30<00:51, 350.52it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2840 (init= 9.0952), step count (max): 80, lr policy:  0.0001:  64%|██████▍   | 32000/50000 [01:30<00:51, 350.52it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2840 (init= 9.0952), step count (max): 80, lr policy:  0.0001:  66%|██████▌   | 33000/50000 [01:33<00:48, 353.54it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2886 (init= 9.0952), step count (max): 93, lr policy:  0.0001:  66%|██████▌   | 33000/50000 [01:33<00:48, 353.54it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2886 (init= 9.0952), step count (max): 93, lr policy:  0.0001:  68%|██████▊   | 34000/50000 [01:36<00:45, 355.07it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2731 (init= 9.0952), step count (max): 50, lr policy:  0.0001:  68%|██████▊   | 34000/50000 [01:36<00:45, 355.07it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2731 (init= 9.0952), step count (max): 50, lr policy:  0.0001:  70%|███████   | 35000/50000 [01:39<00:42, 356.35it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2815 (init= 9.0952), step count (max): 60, lr policy:  0.0001:  70%|███████   | 35000/50000 [01:39<00:42, 356.35it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2815 (init= 9.0952), step count (max): 60, lr policy:  0.0001:  72%|███████▏  | 36000/50000 [01:41<00:39, 357.32it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2872 (init= 9.0952), step count (max): 57, lr policy:  0.0001:  72%|███████▏  | 36000/50000 [01:41<00:39, 357.32it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2872 (init= 9.0952), step count (max): 57, lr policy:  0.0001:  74%|███████▍  | 37000/50000 [01:44<00:36, 357.61it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2843 (init= 9.0952), step count (max): 66, lr policy:  0.0001:  74%|███████▍  | 37000/50000 [01:44<00:36, 357.61it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2843 (init= 9.0952), step count (max): 66, lr policy:  0.0001:  76%|███████▌  | 38000/50000 [01:47<00:33, 357.90it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2921 (init= 9.0952), step count (max): 55, lr policy:  0.0000:  76%|███████▌  | 38000/50000 [01:47<00:33, 357.90it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2921 (init= 9.0952), step count (max): 55, lr policy:  0.0000:  78%|███████▊  | 39000/50000 [01:50<00:30, 358.80it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2943 (init= 9.0952), step count (max): 59, lr policy:  0.0000:  78%|███████▊  | 39000/50000 [01:50<00:30, 358.80it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2943 (init= 9.0952), step count (max): 59, lr policy:  0.0000:  80%|████████  | 40000/50000 [01:53<00:28, 353.70it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2989 (init= 9.0952), step count (max): 58, lr policy:  0.0000:  80%|████████  | 40000/50000 [01:53<00:28, 353.70it/s]
eval cumulative reward:  381.3834 (init:  91.8148), eval step-count: 40, average reward= 9.2989 (init= 9.0952), step count (max): 58, lr policy:  0.0000:  82%|████████▏ | 41000/50000 [01:56<00:25, 355.14it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2873 (init= 9.0952), step count (max): 60, lr policy:  0.0000:  82%|████████▏ | 41000/50000 [01:56<00:25, 355.14it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2873 (init= 9.0952), step count (max): 60, lr policy:  0.0000:  84%|████████▍ | 42000/50000 [01:58<00:22, 354.30it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2945 (init= 9.0952), step count (max): 65, lr policy:  0.0000:  84%|████████▍ | 42000/50000 [01:58<00:22, 354.30it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2945 (init= 9.0952), step count (max): 65, lr policy:  0.0000:  86%|████████▌ | 43000/50000 [02:01<00:19, 355.85it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3009 (init= 9.0952), step count (max): 61, lr policy:  0.0000:  86%|████████▌ | 43000/50000 [02:01<00:19, 355.85it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3009 (init= 9.0952), step count (max): 61, lr policy:  0.0000:  88%|████████▊ | 44000/50000 [02:04<00:16, 356.98it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2961 (init= 9.0952), step count (max): 60, lr policy:  0.0000:  88%|████████▊ | 44000/50000 [02:04<00:16, 356.98it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2961 (init= 9.0952), step count (max): 60, lr policy:  0.0000:  90%|█████████ | 45000/50000 [02:07<00:13, 357.18it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2978 (init= 9.0952), step count (max): 73, lr policy:  0.0000:  90%|█████████ | 45000/50000 [02:07<00:13, 357.18it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2978 (init= 9.0952), step count (max): 73, lr policy:  0.0000:  92%|█████████▏| 46000/50000 [02:10<00:11, 356.44it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3030 (init= 9.0952), step count (max): 84, lr policy:  0.0000:  92%|█████████▏| 46000/50000 [02:10<00:11, 356.44it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3030 (init= 9.0952), step count (max): 84, lr policy:  0.0000:  94%|█████████▍| 47000/50000 [02:12<00:08, 357.02it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3030 (init= 9.0952), step count (max): 58, lr policy:  0.0000:  94%|█████████▍| 47000/50000 [02:12<00:08, 357.02it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3030 (init= 9.0952), step count (max): 58, lr policy:  0.0000:  96%|█████████▌| 48000/50000 [02:15<00:05, 351.31it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2982 (init= 9.0952), step count (max): 79, lr policy:  0.0000:  96%|█████████▌| 48000/50000 [02:15<00:05, 351.31it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.2982 (init= 9.0952), step count (max): 79, lr policy:  0.0000:  98%|█████████▊| 49000/50000 [02:18<00:02, 353.47it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3050 (init= 9.0952), step count (max): 59, lr policy:  0.0000:  98%|█████████▊| 49000/50000 [02:18<00:02, 353.47it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3050 (init= 9.0952), step count (max): 59, lr policy:  0.0000: 100%|██████████| 50000/50000 [02:21<00:00, 355.35it/s]
eval cumulative reward:  325.3639 (init:  91.8148), eval step-count: 34, average reward= 9.3011 (init= 9.0952), step count (max): 74, lr policy:  0.0000: 100%|██████████| 50000/50000 [02:21<00:00, 355.35it/s]

结果#

在达到 100 万步上限之前，算法应该已经达到了 1000 步的最大步数，这是轨迹被截断之前的最大步数。 

plt.figure(figsize=(10, 10))
plt.subplot(2, 2, 1)
plt.plot(logs["reward"])
plt.title("training rewards (average)")
plt.subplot(2, 2, 2)
plt.plot(logs["step_count"])
plt.title("Max step count (training)")
plt.subplot(2, 2, 3)
plt.plot(logs["eval reward (sum)"])
plt.title("Return (test)")
plt.subplot(2, 2, 4)
plt.plot(logs["eval step_count"])
plt.title("Max step count (test)")
plt.show()
training rewards (average), Max step count (training), Return (test), Max step count (test)

结论与后续步骤#

在本教程中，我们学习了：

  1. 如何使用 torchrl 创建和自定义环境；

  2. 如何编写模型和损失函数；

  3. 如何设置典型的训练循环。

如果您想对本教程进行更多实验，可以应用以下修改：

  • 从效率的角度来看，我们可以并行运行多个模拟来加速数据收集。有关更多信息，请查看 ParallelEnv。

  • 从日志记录的角度来看，可以在请求渲染后将 torchrl.record.VideoRecorder 转换添加到环境中，以获得倒立摆动作的视觉呈现。查看 torchrl.record 以了解更多信息。

脚本总运行时间：（2 分 23.200 秒）

在此页面上
PyTorch 库