使用预训练模型

本教程介绍如何在 TorchRL 中使用预训练模型。

本教程结束时，您将能够使用预训练模型进行高效的图像表征，并对它们进行微调。

TorchRL 提供预训练模型，这些模型可用作变换或策略的组件。由于语义相同，它们可以在一种或另一种上下文中互换使用。在本教程中，我们将使用 R3M (https://arxiv.org/abs/2203.12601)，但其他模型（例如 VIP）同样适用。

import torch.cuda
from tensordict.nn import TensorDictSequential
from torch import nn
from torchrl.envs import R3MTransform, TransformedEnv
from torchrl.envs.libs.gym import GymEnv
from torchrl.modules import Actor

is_fork = multiprocessing.get_start_method() == "fork"
device = (
    torch.device(0)
    if torch.cuda.is_available() and not is_fork
    else torch.device("cpu")
)

首先，我们来创建一个环境。为简化起见，我们将使用一个常见的 Gym 环境。在实际应用中，该方法同样适用于更具挑战性的具身 AI 场景（例如，可参考我们的 Habitat 封装器）。

base_env = GymEnv("Ant-v4", from_pixels=True, device=device)

让我们获取预训练模型。我们通过设置 `download=True` 标志来请求该模型的预训练版本；默认情况下，此标志处于关闭状态。 接下来，我们将把变换（transform）添加到环境中。在实际运行中，每一批采集到的数据都会经过该变换，并在输出的 `tensordict` 中映射为一个名为 `r3m_vec` 的条目。我们的策略（policy）由单层多层感知机（MLP）构成，它将读取该向量并计算出对应的动作。

r3m = R3MTransform(
    "resnet50",
    in_keys=["pixels"],
    download=True,
)
env_transformed = TransformedEnv(base_env, r3m)
net = nn.Sequential(
    nn.LazyLinear(128, device=device),
    nn.Tanh(),
    nn.Linear(128, base_env.action_spec.shape[-1], device=device),
)
policy = Actor(net, in_keys=["r3m_vec"])

Downloading: "https://pytorch.s3.amazonaws.com/models/rl/r3m/r3m_50.pt" to /root/.cache/torch/hub/checkpoints/r3m_50.pt

  0%|          | 0.00/374M [00:00<?, ?B/s]
  4%|▍         | 16.2M/374M [00:00<00:02, 170MB/s]
  9%|▊         | 32.5M/374M [00:00<00:03, 98.9MB/s]
 12%|█▏        | 43.5M/374M [00:00<00:03, 87.6MB/s]
 14%|█▍        | 52.8M/374M [00:00<00:05, 65.4MB/s]
 17%|█▋        | 64.0M/374M [00:00<00:04, 70.1MB/s]
 19%|█▉        | 71.4M/374M [00:00<00:04, 71.4MB/s]
 22%|██▏       | 82.0M/374M [00:01<00:04, 68.7MB/s]
 26%|██▌       | 97.8M/374M [00:01<00:03, 76.9MB/s]
 28%|██▊       | 105M/374M [00:01<00:03, 73.2MB/s]
 31%|███       | 115M/374M [00:01<00:03, 71.1MB/s]
 35%|███▌      | 131M/374M [00:01<00:03, 79.1MB/s]
 39%|███▉      | 148M/374M [00:01<00:02, 82.9MB/s]
 43%|████▎     | 162M/374M [00:02<00:02, 92.9MB/s]
 46%|████▌     | 172M/374M [00:02<00:02, 84.4MB/s]
 48%|████▊     | 180M/374M [00:02<00:02, 81.6MB/s]
 50%|█████     | 188M/374M [00:02<00:02, 69.9MB/s]
 52%|█████▏    | 195M/374M [00:02<00:03, 58.3MB/s]
 54%|█████▎    | 201M/374M [00:02<00:03, 52.6MB/s]
 57%|█████▋    | 213M/374M [00:03<00:02, 58.5MB/s]
 61%|██████▏   | 229M/374M [00:03<00:02, 65.8MB/s]
 66%|██████▌   | 246M/374M [00:03<00:01, 72.7MB/s]
 70%|██████▉   | 262M/374M [00:03<00:01, 63.4MB/s]
 72%|███████▏  | 268M/374M [00:04<00:02, 53.6MB/s]
 74%|███████▍  | 277M/374M [00:04<00:01, 55.5MB/s]
 75%|███████▌  | 282M/374M [00:04<00:01, 52.0MB/s]
 78%|███████▊  | 293M/374M [00:04<00:01, 58.9MB/s]
 80%|███████▉  | 299M/374M [00:04<00:01, 54.9MB/s]
 83%|████████▎ | 310M/374M [00:04<00:01, 63.3MB/s]
 84%|████████▍ | 316M/374M [00:04<00:01, 55.4MB/s]
 87%|████████▋ | 326M/374M [00:05<00:01, 50.1MB/s]
 89%|████████▊ | 331M/374M [00:05<00:00, 45.4MB/s]
 92%|█████████▏| 342M/374M [00:05<00:00, 45.7MB/s]
 93%|█████████▎| 347M/374M [00:05<00:00, 42.6MB/s]
 96%|█████████▌| 359M/374M [00:05<00:00, 45.3MB/s]
 97%|█████████▋| 363M/374M [00:06<00:00, 41.3MB/s]
100%|██████████| 374M/374M [00:06<00:00, 63.4MB/s]

让我们检查策略的参数数量：

print("number of params:", len(list(policy.parameters())))

number of params: 4

我们收集 32 步的 rollout 并打印其输出：

rollout = env_transformed.rollout(32, policy)
print("rollout with transform:", rollout)

rollout with transform: TensorDict(
    fields={
        action: Tensor(shape=torch.Size([32, 8]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                r3m_vec: Tensor(shape=torch.Size([32, 2048]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([32]),
            device=cpu,
            is_shared=False),
        r3m_vec: Tensor(shape=torch.Size([32, 2048]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([32]),
    device=cpu,
    is_shared=False)

对于微调，我们在使参数变为可训练状态后，将变换操作集成到策略中。在实践中，限制该操作仅作用于参数的一个子集（例如多层感知机的最后一层）可能更为明智。

r3m.train()
policy = TensorDictSequential(r3m, policy)
print("number of params after r3m is integrated:", len(list(policy.parameters())))

number of params after r3m is integrated: 163

同样，我们再次使用 R3M 进行 rollout。输出结构发生了轻微变化，因为此时环境返回的是像素值（而非嵌入向量）。“r3m_vec” 嵌入是我们的策略所生成的一个中间结果。

rollout = base_env.rollout(32, policy)
print("rollout, fine tuning:", rollout)

rollout, fine tuning: TensorDict(
    fields={
        action: Tensor(shape=torch.Size([32, 8]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                pixels: Tensor(shape=torch.Size([32, 480, 480, 3]), device=cpu, dtype=torch.uint8, is_shared=False),
                reward: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([32]),
            device=cpu,
            is_shared=False),
        r3m_vec: Tensor(shape=torch.Size([32, 2048]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([32]),
    device=cpu,
    is_shared=False)

我们将变换从环境交换到策略的简便性，是因为两者都像 TensorDictModule 一样：它们拥有一组 “in_keys” 和 “out_keys”，使得在不同上下文中读写输出变得容易。

本教程最后，让我们看看如何使用 R3M 读取存储在回放缓冲区（例如，在离线强化学习场景中）中的图像。首先，我们来构建数据集：

from torchrl.data import LazyMemmapStorage, ReplayBuffer

storage = LazyMemmapStorage(1000)
rb = ReplayBuffer(storage=storage, transform=r3m)

现在，我们可以收集数据（对我们的目的而言是随机 rollout）并用其填充经验回放缓冲区：

total = 0
while total < 1000:
    tensordict = base_env.rollout(1000)
    rb.extend(tensordict)
    total += tensordict.numel()

让我们检查一下重放缓冲区存储的结构。由于我们尚未使用它，因此其中不应包含“r3m_vec”条目：

print("stored data:", storage._storage)

stored data: TensorDict(
    fields={
        action: MemoryMappedTensor(shape=torch.Size([1000, 8]), device=cpu, dtype=torch.float32, is_shared=False),
        done: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                pixels: MemoryMappedTensor(shape=torch.Size([1000, 480, 480, 3]), device=cpu, dtype=torch.uint8, is_shared=False),
                reward: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([1000]),
            device=cpu,
            is_shared=False),
        pixels: MemoryMappedTensor(shape=torch.Size([1000, 480, 480, 3]), device=cpu, dtype=torch.uint8, is_shared=False),
        terminated: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: MemoryMappedTensor(shape=torch.Size([1000, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([1000]),
    device=cpu,
    is_shared=False)

采样时，数据将经过 R3M 变换，从而得到我们所需的处理后数据。 通过这种方式，我们可以在由图像构成的数据集上离线训练算法：

batch = rb.sample(32)
print("data after sampling:", batch)

data after sampling: TensorDict(
    fields={
        action: Tensor(shape=torch.Size([32, 8]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                pixels: Tensor(shape=torch.Size([32, 480, 480, 3]), device=cpu, dtype=torch.uint8, is_shared=False),
                reward: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([32]),
            device=cpu,
            is_shared=False),
        r3m_vec: Tensor(shape=torch.Size([32, 2048]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([32, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([32]),
    device=cpu,
    is_shared=False)

估计内存使用量: 4016 MB