注意

转到末尾下载完整的示例代码。

TorchRL 简介¶

此演示在 ICML 2022 的行业演示日上进行了演示。

它很好地概述了 TorchRL 的功能。请随时联系我们发送至 vmoens@fb。com 或提交问题（如果您对它。

TorchRL 是 PyTorch 的开源强化学习（RL）库。

PyTorch 生态系统团队（Meta）已决定投资该库，以为在研究环境中开发 RL 解决方案提供领先的平台。

它为 RL 提供了 pytorch 和 python 优先、低级和高级抽象 # 旨在高效、文档化和经过适当测试。该代码旨在支持 RL 研究。其中大部分是用 Python 以高度模块化的方式进行，以便研究人员可以轻松地交换组件，轻松转换它们或编写新的组件。

此存储库尝试与现有的 pytorch 生态系统库保持一致因为它有一个数据集支柱（torchrl/envs）、转换、模型、数据实用程序（例如收集器和容器）等。TorchRL 的目标是尽可能少的依赖项（Python 标准库、NumPy 和 PyTorch）。常见的环境库（例如 OpenAI gym）是可选的。

内容：

与其他领域不同，RL 与其说是媒体，不如说是算法。因此，它更难制造真正独立的组件。

TorchRL 不是什么：

算法的集合：我们不打算提供 RL 算法的 SOTA 实现，但我们仅提供这些算法作为如何使用该库的示例。
研究框架：TorchRL 中的模块化有两种风格。首先，我们尝试构建可重用的组件，以便它们可以轻松地相互交换。其次，我们尽最大努力使组件可以独立于其他组件使用的库。

TorchRL 的核心依赖项非常少，主要是 PyTorch 和 numpy。都其他依赖项（gym、torchvision、wandb / Tensorboard）是可选的。

数据¶

张量字典¶

import torch
from tensordict import TensorDict

让我们创建一个 TensorDict。构造函数接受许多不同的格式，例如传递 dict 或使用关键字参数：

batch_size = 5
data = TensorDict(
    key1=torch.zeros(batch_size, 3),
    key2=torch.zeros(batch_size, 5, 6, dtype=torch.bool),
    batch_size=[batch_size],
)
print(data)

TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([5, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: Tensor(shape=torch.Size([5, 5, 6]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([5]),
    device=None,
    is_shared=False)

您可以沿 TensorDict 以及查询键为 TensorDict 编制索引。batch_size

print(data[2])
print(data["key1"] is data.get("key1"))

TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: Tensor(shape=torch.Size([5, 6]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)
True

下面展示了如何堆叠多个 TensorDict。这在编写转出循环时特别有用！

data1 = TensorDict(
    {
        "key1": torch.zeros(batch_size, 1),
        "key2": torch.zeros(batch_size, 5, 6, dtype=torch.bool),
    },
    batch_size=[batch_size],
)

data2 = TensorDict(
    {
        "key1": torch.ones(batch_size, 1),
        "key2": torch.ones(batch_size, 5, 6, dtype=torch.bool),
    },
    batch_size=[batch_size],
)

data = torch.stack([data1, data2], 0)
data.batch_size, data["key1"]

(torch.Size([2, 5]), tensor([[[0.],
         [0.],
         [0.],
         [0.],
         [0.]],

        [[1.],
         [1.],
         [1.],
         [1.],
         [1.]]]))

以下是 TensorDict 的其他一些功能：查看、排列、共享内存或扩展。

print(
    "view(-1): ",
    data.view(-1).batch_size,
    data.view(-1).get("key1").shape,
)

print("to device: ", data.to("cpu"))

# print("pin_memory: ", data.pin_memory())

print("share memory: ", data.share_memory_())

print(
    "permute(1, 0): ",
    data.permute(1, 0).batch_size,
    data.permute(1, 0).get("key1").shape,
)

print(
    "expand: ",
    data.expand(3, *data.batch_size).batch_size,
    data.expand(3, *data.batch_size).get("key1").shape,
)

view(-1):  torch.Size([10]) torch.Size([10, 1])
to device:  TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([2, 5, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: Tensor(shape=torch.Size([2, 5, 5, 6]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([2, 5]),
    device=cpu,
    is_shared=False)
share memory:  TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([2, 5, 1]), device=cpu, dtype=torch.float32, is_shared=True),
        key2: Tensor(shape=torch.Size([2, 5, 5, 6]), device=cpu, dtype=torch.bool, is_shared=True)},
    batch_size=torch.Size([2, 5]),
    device=None,
    is_shared=True)
permute(1, 0):  torch.Size([5, 2]) torch.Size([5, 2, 1])
expand:  torch.Size([3, 2, 5]) torch.Size([3, 2, 5, 1])

您也可以创建嵌套数据。

data = TensorDict(
    source={
        "key1": torch.zeros(batch_size, 3),
        "key2": TensorDict(
            source={"sub_key1": torch.zeros(batch_size, 2, 1)},
            batch_size=[batch_size, 2],
        ),
    },
    batch_size=[batch_size],
)
data

TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([5, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: TensorDict(
            fields={
                sub_key1: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([5, 2]),
            device=None,
            is_shared=False)},
    batch_size=torch.Size([5]),
    device=None,
    is_shared=False)

重放缓冲区¶

重放缓冲区是许多 RL 算法中的关键组件。TorchRL 提供了一系列重放缓冲区实现。大多数基本功能适用于任何数据结构（list、tuples、dict），但要使用重放缓冲区来访问它们的完全扩展并且具有快速读写访问权限，应首选 TensorDict API。

from torchrl.data import PrioritizedReplayBuffer, ReplayBuffer

rb = ReplayBuffer(collate_fn=lambda x: x)

可以使用（n=1）进行加法或（n>1）。

rb.add(1)
rb.sample(1)
rb.extend([2, 3])
rb.sample(3)

[2, 1, 3]

还可以使用 Prioritized Replay Buffers：

rb = PrioritizedReplayBuffer(alpha=0.7, beta=1.1, collate_fn=lambda x: x)
rb.add(1)
rb.sample(1)
rb.update_priority(1, 0.5)

以下是将 replaybuffer 与 data_stack 一起使用的示例。使用它们可以轻松抽象出多个用例的重放缓冲区的行为。

collate_fn = torch.stack
rb = ReplayBuffer(collate_fn=collate_fn)
rb.add(TensorDict({"a": torch.randn(3)}, batch_size=[]))
len(rb)

rb.extend(TensorDict({"a": torch.randn(2, 3)}, batch_size=[2]))
print(len(rb))
print(rb.sample(10))
print(rb.sample(2).contiguous())

torch.manual_seed(0)
from torchrl.data import TensorDictPrioritizedReplayBuffer

rb = TensorDictPrioritizedReplayBuffer(alpha=0.7, beta=1.1, priority_key="td_error")
rb.extend(TensorDict({"a": torch.randn(2, 3)}, batch_size=[2]))
data_sample = rb.sample(2).contiguous()
print(data_sample)

print(data_sample["index"])

data_sample["td_error"] = torch.rand(2)
rb.update_tensordict_priority(data_sample)

for i, val in enumerate(rb._sampler._sum_tree):
    print(i, val)
    if i == len(rb):
        break

3
TensorDict(
    fields={
        a: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([10]),
    device=None,
    is_shared=False)
TensorDict(
    fields={
        a: Tensor(shape=torch.Size([2, 3]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False)
TensorDict(
    fields={
        _weight: Tensor(shape=torch.Size([2]), device=cpu, dtype=torch.float32, is_shared=False),
        a: Tensor(shape=torch.Size([2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        index: Tensor(shape=torch.Size([2]), device=cpu, dtype=torch.int64, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False)
tensor([1, 1])
0 1.0
1 0.28791671991348267
2 0.0

环境¶

TorchRL 提供了一系列环境包装器和实用程序。

健身房环境¶

try:
    import gymnasium as gym
except ModuleNotFoundError:
    import gym

from torchrl.envs.libs.gym import GymEnv, GymWrapper, set_gym_backend

gym_env = gym.make("Pendulum-v1")
env = GymWrapper(gym_env)
env = GymEnv("Pendulum-v1")

data = env.reset()
env.rand_step(data)

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, 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=None,
            is_shared=False),
        observation: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, 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=None,
    is_shared=False)

更改环境配置¶

env = GymEnv("Pendulum-v1", frame_skip=3, from_pixels=True, pixels_only=False)
env.reset()

env.close()
del env

from torchrl.envs import (
    Compose,
    NoopResetEnv,
    ObservationNorm,
    ToTensorImage,
    TransformedEnv,
)

base_env = GymEnv("Pendulum-v1", frame_skip=3, from_pixels=True, pixels_only=False)
env = TransformedEnv(base_env, Compose(NoopResetEnv(3), ToTensorImage()))
env.append_transform(ObservationNorm(in_keys=["pixels"], loc=2, scale=1))

TransformedEnv(
    env=GymEnv(env=Pendulum-v1, batch_size=torch.Size([]), device=None),
    transform=Compose(
            NoopResetEnv(noops=3, random=True),
            ToTensorImage(keys=['pixels']),
            ObservationNorm(loc=2.0000, scale=1.0000, keys=['pixels'])))

环境变换¶

转换的行为类似于 Gym 包装器，但具有更接近 torchvision 的 ' 转换的 API。有多种变换可供选择。torch.distributions

from torchrl.envs import (
    Compose,
    NoopResetEnv,
    ObservationNorm,
    StepCounter,
    ToTensorImage,
    TransformedEnv,
)

base_env = GymEnv("HalfCheetah-v4", frame_skip=3, from_pixels=True, pixels_only=False)
env = TransformedEnv(base_env, Compose(NoopResetEnv(3), ToTensorImage()))
env = env.append_transform(ObservationNorm(in_keys=["pixels"], loc=2, scale=1))

env.reset()

print("env: ", env)
print("last transform parent: ", env.transform[2].parent)

env:  TransformedEnv(
    env=GymEnv(env=HalfCheetah-v4, batch_size=torch.Size([]), device=None),
    transform=Compose(
            NoopResetEnv(noops=3, random=True),
            ToTensorImage(keys=['pixels']),
            ObservationNorm(loc=2.0000, scale=1.0000, keys=['pixels'])))
last transform parent:  TransformedEnv(
    env=GymEnv(env=HalfCheetah-v4, batch_size=torch.Size([]), device=None),
    transform=Compose(
            NoopResetEnv(noops=3, random=True),
            ToTensorImage(keys=['pixels'])))

矢量化环境¶

矢量化/并行环境可以提供一些显著的加速。

from torchrl.envs import ParallelEnv


def make_env():
    # You can control whether to use gym or gymnasium for your env
    with set_gym_backend("gym"):
        return GymEnv("Pendulum-v1", frame_skip=3, from_pixels=True, pixels_only=False)


base_env = ParallelEnv(
    4,
    make_env,
    mp_start_method="fork",  # This will break on Windows machines! Remove and decorate with if __name__ == "__main__"
)
env = TransformedEnv(
    base_env, Compose(StepCounter(), ToTensorImage())
)  # applies transforms on batch of envs
env.append_transform(ObservationNorm(in_keys=["pixels"], loc=2, scale=1))
env.reset()

print(env.action_spec)

env.close()
del env

BoundedContinuous(
    shape=torch.Size([4, 1]),
    space=ContinuousBox(
        low=Tensor(shape=torch.Size([4, 1]), device=cpu, dtype=torch.float32, contiguous=True),
        high=Tensor(shape=torch.Size([4, 1]), device=cpu, dtype=torch.float32, contiguous=True)),
    device=cpu,
    dtype=torch.float32,
    domain=continuous)

模块¶

可以在库中找到多个模块（utils、models 和 wrappers）。

模型¶

MLP 模型示例：

from torch import nn
from torchrl.modules import ConvNet, MLP
from torchrl.modules.models.utils import SquashDims

net = MLP(num_cells=[32, 64], out_features=4, activation_class=nn.ELU)
print(net)
print(net(torch.randn(10, 3)).shape)

MLP(
  (0): LazyLinear(in_features=0, out_features=32, bias=True)
  (1): ELU(alpha=1.0)
  (2): Linear(in_features=32, out_features=64, bias=True)
  (3): ELU(alpha=1.0)
  (4): Linear(in_features=64, out_features=4, bias=True)
)
torch.Size([10, 4])

CNN 模型示例：

cnn = ConvNet(
    num_cells=[32, 64],
    kernel_sizes=[8, 4],
    strides=[2, 1],
    aggregator_class=SquashDims,
)
print(cnn)
print(cnn(torch.randn(10, 3, 32, 32)).shape)  # last tensor is squashed

ConvNet(
  (0): LazyConv2d(0, 32, kernel_size=(8, 8), stride=(2, 2))
  (1): ELU(alpha=1.0)
  (2): Conv2d(32, 64, kernel_size=(4, 4), stride=(1, 1))
  (3): ELU(alpha=1.0)
  (4): SquashDims()
)
torch.Size([10, 6400])

TensorDict 模块¶

一些模块是专门为使用 tensordict 输入而设计的。

from tensordict.nn import TensorDictModule

data = TensorDict({"key1": torch.randn(10, 3)}, batch_size=[10])
module = nn.Linear(3, 4)
td_module = TensorDictModule(module, in_keys=["key1"], out_keys=["key2"])
td_module(data)
print(data)

TensorDict(
    fields={
        key1: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: Tensor(shape=torch.Size([10, 4]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([10]),
    device=None,
    is_shared=False)

模块序列¶

通过以下方式可以轻松创建模块序列：TensorDictSequential

from tensordict.nn import TensorDictSequential

backbone_module = nn.Linear(5, 3)
backbone = TensorDictModule(
    backbone_module, in_keys=["observation"], out_keys=["hidden"]
)
actor_module = nn.Linear(3, 4)
actor = TensorDictModule(actor_module, in_keys=["hidden"], out_keys=["action"])
value_module = MLP(out_features=1, num_cells=[4, 5])
value = TensorDictModule(value_module, in_keys=["hidden", "action"], out_keys=["value"])

sequence = TensorDictSequential(backbone, actor, value)
print(sequence)

print(sequence.in_keys, sequence.out_keys)

data = TensorDict(
    {"observation": torch.randn(3, 5)},
    [3],
)
backbone(data)
actor(data)
value(data)

data = TensorDict(
    {"observation": torch.randn(3, 5)},
    [3],
)
sequence(data)
print(data)

TensorDictSequential(
    module=ModuleList(
      (0): TensorDictModule(
          module=Linear(in_features=5, out_features=3, bias=True),
          device=cpu,
          in_keys=['observation'],
          out_keys=['hidden'])
      (1): TensorDictModule(
          module=Linear(in_features=3, out_features=4, bias=True),
          device=cpu,
          in_keys=['hidden'],
          out_keys=['action'])
      (2): TensorDictModule(
          module=MLP(
            (0): LazyLinear(in_features=0, out_features=4, bias=True)
            (1): Tanh()
            (2): Linear(in_features=4, out_features=5, bias=True)
            (3): Tanh()
            (4): Linear(in_features=5, out_features=1, bias=True)
          ),
          device=cpu,
          in_keys=['hidden', 'action'],
          out_keys=['value'])
    ),
    device=cpu,
    in_keys=['observation'],
    out_keys=['hidden', 'action', 'value'])
['observation'] ['hidden', 'action', 'value']
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        hidden: Tensor(shape=torch.Size([3, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
        value: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

函数式编程（Ensembling / Meta-RL）¶

函数调用从未如此简单。使用、和提取参数将它们替换为：from_module()to_module()

from tensordict import from_module

params = from_module(sequence)
print("extracted params", params)

extracted params TensorDict(
    fields={
        module: TensorDict(
            fields={
                0: TensorDict(
                    fields={
                        module: TensorDict(
                            fields={
                                bias: Parameter(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
                                weight: Parameter(shape=torch.Size([3, 5]), device=cpu, dtype=torch.float32, is_shared=False)},
                            batch_size=torch.Size([]),
                            device=None,
                            is_shared=False)},
                    batch_size=torch.Size([]),
                    device=None,
                    is_shared=False),
                1: TensorDict(
                    fields={
                        module: TensorDict(
                            fields={
                                bias: Parameter(shape=torch.Size([4]), device=cpu, dtype=torch.float32, is_shared=False),
                                weight: Parameter(shape=torch.Size([4, 3]), device=cpu, dtype=torch.float32, is_shared=False)},
                            batch_size=torch.Size([]),
                            device=None,
                            is_shared=False)},
                    batch_size=torch.Size([]),
                    device=None,
                    is_shared=False),
                2: TensorDict(
                    fields={
                        module: TensorDict(
                            fields={
                                0: TensorDict(
                                    fields={
                                        bias: Parameter(shape=torch.Size([4]), device=cpu, dtype=torch.float32, is_shared=False),
                                        weight: Parameter(shape=torch.Size([4, 7]), device=cpu, dtype=torch.float32, is_shared=False)},
                                    batch_size=torch.Size([]),
                                    device=None,
                                    is_shared=False),
                                2: TensorDict(
                                    fields={
                                        bias: Parameter(shape=torch.Size([5]), device=cpu, dtype=torch.float32, is_shared=False),
                                        weight: Parameter(shape=torch.Size([5, 4]), device=cpu, dtype=torch.float32, is_shared=False)},
                                    batch_size=torch.Size([]),
                                    device=None,
                                    is_shared=False),
                                4: TensorDict(
                                    fields={
                                        bias: Parameter(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
                                        weight: Parameter(shape=torch.Size([1, 5]), device=cpu, dtype=torch.float32, is_shared=False)},
                                    batch_size=torch.Size([]),
                                    device=None,
                                    is_shared=False)},
                            batch_size=torch.Size([]),
                            device=None,
                            is_shared=False)},
                    batch_size=torch.Size([]),
                    device=None,
                    is_shared=False)},
            batch_size=torch.Size([]),
            device=None,
            is_shared=False)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)

使用 tensordict 的函数调用：

with params.to_module(sequence):
    data = sequence(data)

VMAP¶

快速执行类似架构的多个副本是快速训练模型的关键。专为实现此目的而量身定制：

from torch import vmap

params_expand = params.expand(4)

def exec_sequence(params, data):
    with params.to_module(sequence):
        return sequence(data)

tensordict_exp = vmap(exec_sequence, (0, None))(params_expand, data)
print(tensordict_exp)

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([4, 3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        hidden: Tensor(shape=torch.Size([4, 3, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([4, 3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
        value: Tensor(shape=torch.Size([4, 3, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([4, 3]),
    device=None,
    is_shared=False)

专业类¶

TorchRL 还提供了一些专门的模块，用于对输出值运行检查。

torch.manual_seed(0)
from torchrl.data import Bounded
from torchrl.modules import SafeModule

spec = Bounded(-torch.ones(3), torch.ones(3))
base_module = nn.Linear(5, 3)
module = SafeModule(
    module=base_module, spec=spec, in_keys=["obs"], out_keys=["action"], safe=True
)
data = TensorDict({"obs": torch.randn(5)}, batch_size=[])
module(data)["action"]

data = TensorDict({"obs": torch.randn(5) * 100}, batch_size=[])
module(data)["action"]  # safe=True projects the result within the set

tensor([-1.,  1., -1.], grad_fn=<AsStridedBackward0>)

该类具有预定义的输出键（）：Actor"action"

from torchrl.modules import Actor

base_module = nn.Linear(5, 3)
actor = Actor(base_module, in_keys=["obs"])
data = TensorDict({"obs": torch.randn(5)}, batch_size=[])
actor(data)  # action is the default value

from tensordict.nn import (
    ProbabilisticTensorDictModule,
    ProbabilisticTensorDictSequential,
)

多亏了 API，使用概率模型也变得容易：tensordict.nn

from torchrl.modules import NormalParamExtractor, TanhNormal

td = TensorDict({"input": torch.randn(3, 5)}, [3])
net = nn.Sequential(
    nn.Linear(5, 4), NormalParamExtractor()
)  # splits the output in loc and scale
module = TensorDictModule(net, in_keys=["input"], out_keys=["loc", "scale"])
td_module = ProbabilisticTensorDictSequential(
    module,
    ProbabilisticTensorDictModule(
        in_keys=["loc", "scale"],
        out_keys=["action"],
        distribution_class=TanhNormal,
        return_log_prob=False,
    ),
)
td_module(td)
print(td)

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False),
        input: Tensor(shape=torch.Size([3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
        loc: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False),
        scale: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

# returning the log-probability
td = TensorDict({"input": torch.randn(3, 5)}, [3])
td_module = ProbabilisticTensorDictSequential(
    module,
    ProbabilisticTensorDictModule(
        in_keys=["loc", "scale"],
        out_keys=["action"],
        distribution_class=TanhNormal,
        return_log_prob=True,
    ),
)
td_module(td)
print(td)

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False),
        input: Tensor(shape=torch.Size([3, 5]), device=cpu, dtype=torch.float32, is_shared=False),
        loc: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False),
        sample_log_prob: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        scale: Tensor(shape=torch.Size([3, 2]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

控制随机性和采样策略是通过上下文管理器实现的，：set_exploration_type

from torchrl.envs.utils import ExplorationType, set_exploration_type

td = TensorDict({"input": torch.randn(3, 5)}, [3])

torch.manual_seed(0)
with set_exploration_type(ExplorationType.RANDOM):
    td_module(td)
    print("random:", td["action"])

with set_exploration_type(ExplorationType.DETERMINISTIC):
    td_module(td)
    print("mode:", td["action"])

random: tensor([[ 0.8728, -0.1334],
        [-0.9833,  0.3494],
        [-0.6887, -0.6402]], grad_fn=<_SafeTanhNoEpsBackward>)
mode: tensor([[-0.1132,  0.1762],
        [-0.3430, -0.2668],
        [ 0.2918,  0.6239]], grad_fn=<_SafeTanhNoEpsBackward>)

使用环境和模块¶

让我们看看如何组合环境和模块：

from torchrl.envs.utils import step_mdp

env = GymEnv("Pendulum-v1")

action_spec = env.action_spec
actor_module = nn.Linear(3, 1)
actor = SafeModule(
    actor_module, spec=action_spec, in_keys=["observation"], out_keys=["action"]
)

torch.manual_seed(0)
env.set_seed(0)

max_steps = 100
data = env.reset()
data_stack = TensorDict(batch_size=[max_steps])
for i in range(max_steps):
    actor(data)
    data_stack[i] = env.step(data)
    if data["done"].any():
        break
    data = step_mdp(data)  # roughly equivalent to obs = next_obs

tensordicts_prealloc = data_stack.clone()
print("total steps:", i)
print(data_stack)

total steps: 99
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([100, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([100]),
            device=None,
            is_shared=False),
        observation: Tensor(shape=torch.Size([100, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([100]),
    device=None,
    is_shared=False)

# equivalent
torch.manual_seed(0)
env.set_seed(0)

max_steps = 100
data = env.reset()
data_stack = []
for _ in range(max_steps):
    actor(data)
    data_stack.append(env.step(data))
    if data["done"].any():
        break
    data = step_mdp(data)  # roughly equivalent to obs = next_obs
tensordicts_stack = torch.stack(data_stack, 0)
print("total steps:", i)
print(tensordicts_stack)

total steps: 99
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([100, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([100]),
            device=None,
            is_shared=False),
        observation: Tensor(shape=torch.Size([100, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([100, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([100]),
    device=None,
    is_shared=False)

(tensordicts_stack == tensordicts_prealloc).all()

True

torch.manual_seed(0)
env.set_seed(0)
tensordict_rollout = env.rollout(policy=actor, max_steps=max_steps)
tensordict_rollout


(tensordict_rollout == tensordicts_prealloc).all()

from tensordict.nn import TensorDictModule

收藏家¶

我们还提供了一组数据收集器，它们根据需要自动收集每批任意数量的帧。它们从单节点、单个工作线程到多节点、多工作线程设置。

from torchrl.collectors import MultiaSyncDataCollector, MultiSyncDataCollector

from torchrl.envs import EnvCreator, SerialEnv
from torchrl.envs.libs.gym import GymEnv

EnvCreator 确保我们可以从一个进程发送到另一个进程 lambda 函数为了简单起见（单个 worker），我们使用 a （多个 worker）会更适合。

注意

Multiprocessed envs 和 Multiprocessed collector 可以组合使用！

parallel_env = SerialEnv(
    3,
    EnvCreator(lambda: GymEnv("Pendulum-v1")),
)
create_env_fn = [parallel_env, parallel_env]

actor_module = nn.Linear(3, 1)
actor = TensorDictModule(actor_module, in_keys=["observation"], out_keys=["action"])

同步多处理数据收集器¶

devices = ["cpu", "cpu"]

collector = MultiSyncDataCollector(
    create_env_fn=create_env_fn,  # either a list of functions or a ParallelEnv
    policy=actor,
    total_frames=240,
    max_frames_per_traj=-1,  # envs are terminating, we don't need to stop them early
    frames_per_batch=60,  # we want 60 frames at a time (we have 3 envs per sub-collector)
    device=devices,
)

for i, d in enumerate(collector):
    if i == 0:
        print(d)  # trajectories are split automatically in [6 workers x 10 steps]
    collector.update_policy_weights_()  # make sure that our policies have the latest weights if working on multiple devices
print(i)
collector.shutdown()
del collector

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        collector: TensorDict(
            fields={
                traj_ids: Tensor(shape=torch.Size([2, 3, 10]), device=cpu, dtype=torch.int64, is_shared=False)},
            batch_size=torch.Size([2, 3, 10]),
            device=cpu,
            is_shared=False),
        done: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([2, 3, 10, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([2, 3, 10]),
            device=cpu,
            is_shared=False),
        observation: Tensor(shape=torch.Size([2, 3, 10, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 3, 10, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([2, 3, 10]),
    device=cpu,
    is_shared=False)
3

异步多处理数据收集器¶

此类允许您在模型训练时收集数据。这在 off-policy 设置中特别有用因为它将推理和模型训练解耦。数据以先就绪先得的方式（工作人员将他们的结果排队）：

collector = MultiaSyncDataCollector(
    create_env_fn=create_env_fn,  # either a list of functions or a ParallelEnv
    policy=actor,
    total_frames=240,
    max_frames_per_traj=-1,  # envs are terminating, we don't need to stop them early
    frames_per_batch=60,  # we want 60 frames at a time (we have 3 envs per sub-collector)
    device=devices,
)

for i, d in enumerate(collector):
    if i == 0:
        print(d)  # trajectories are split automatically in [6 workers x 10 steps]
    collector.update_policy_weights_()  # make sure that our policies have the latest weights if working on multiple devices
print(i)
collector.shutdown()
del collector
del create_env_fn
del parallel_env

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        collector: TensorDict(
            fields={
                traj_ids: Tensor(shape=torch.Size([3, 20]), device=cpu, dtype=torch.int64, is_shared=False)},
            batch_size=torch.Size([3, 20]),
            device=cpu,
            is_shared=False),
        done: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([3, 20, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([3, 20]),
            device=cpu,
            is_shared=False),
        observation: Tensor(shape=torch.Size([3, 20, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([3, 20]),
    device=cpu,
    is_shared=False)
3

目标¶

Objectives 是编写新算法时的主要切入点。

from torchrl.objectives import DDPGLoss

actor_module = nn.Linear(3, 1)
actor = TensorDictModule(actor_module, in_keys=["observation"], out_keys=["action"])


class ConcatModule(nn.Linear):
    def forward(self, obs, action):
        return super().forward(torch.cat([obs, action], -1))


value_module = ConcatModule(4, 1)
value = TensorDictModule(
    value_module, in_keys=["observation", "action"], out_keys=["state_action_value"]
)

loss_fn = DDPGLoss(actor, value)
loss_fn.make_value_estimator(loss_fn.default_value_estimator, gamma=0.99)

data = TensorDict(
    {
        "observation": torch.randn(10, 3),
        "next": {
            "observation": torch.randn(10, 3),
            "reward": torch.randn(10, 1),
            "done": torch.zeros(10, 1, dtype=torch.bool),
        },
        "action": torch.randn(10, 1),
    },
    batch_size=[10],
    device="cpu",
)
loss_td = loss_fn(data)

print(loss_td)

print(data)

TensorDict(
    fields={
        loss_actor: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        loss_value: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        pred_value: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float32, is_shared=False),
        pred_value_max: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        target_value: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float32, is_shared=False),
        target_value_max: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        td_error: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([10]),
            device=cpu,
            is_shared=False),
        observation: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        td_error: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([10]),
    device=cpu,
    is_shared=False)

安装库¶

该库位于 PyPI 上：pip install torchrl 有关更多信息，请参阅 README。

贡献¶

我们正在积极寻找贡献者和早期用户。如果你在 RL（或者只是好奇），试试吧！给我们反馈：什么会让 TorchRL 是它满足研究人员需求的程度。要做到这一点，我们需要他们的意见！由于库刚刚起步，现在是你按照自己的方式塑造它的好时机！

有关详细信息，请参阅贡献指南。

脚本总运行时间：（2 分 32.651 秒）

估计内存使用量：322 MB

由 Sphinx-Gallery 生成的图库

TorchRL 简介¶

数据¶

张量字典¶

重放缓冲区¶

环境¶

健身房环境¶

更改环境配置¶

环境变换¶

矢量化环境¶

模块¶

模型¶

TensorDict 模块¶

模块序列¶

函数式编程（Ensembling / Meta-RL）¶

VMAP¶

专业类¶

使用环境和模块¶

收藏家¶

同步多处理数据收集器¶

异步多处理数据收集器¶

目标¶

安装库¶

贡献¶

文档

教程

资源

TorchRL 简介¶

数据¶

张量字典¶

重放缓冲区¶

环境¶

健身房环境¶

更改环境配置¶

环境变换¶

矢量化环境¶

模块¶

模型¶

TensorDict 模块¶

模块序列¶

函数式编程 （Ensembling / Meta-RL）¶

VMAP¶

专业类¶

使用环境和模块¶

收藏家¶

同步多处理数据收集器¶

异步多处理数据收集器¶

目标¶

安装库¶

贡献¶

文档

教程

资源

函数式编程（Ensembling / Meta-RL）¶