多任務環境中的任務特定策略¶

本教學詳細說明如何使用多任務策略和批次環境。

在本教學結束時，您將能夠編寫可以使用一組不同的權重在各種設定中計算動作的策略。您也將能夠並行執行各種環境。

from tensordict.nn import TensorDictModule, TensorDictSequential
from torch import nn

from torchrl.envs import CatTensors, Compose, DoubleToFloat, ParallelEnv, TransformedEnv
from torchrl.envs.libs.dm_control import DMControlEnv
from torchrl.modules import MLP

我們設計了兩個環境，一個是必須完成站立任務的人形，另一個是必須學習行走的。

env1 = DMControlEnv("humanoid", "stand")
env1_obs_keys = list(env1.observation_spec.keys())
env1 = TransformedEnv(
    env1,
    Compose(
        CatTensors(env1_obs_keys, "observation_stand", del_keys=False),
        CatTensors(env1_obs_keys, "observation"),
        DoubleToFloat(
            in_keys=["observation_stand", "observation"],
            in_keys_inv=["action"],
        ),
    ),
)
env2 = DMControlEnv("humanoid", "walk")
env2_obs_keys = list(env2.observation_spec.keys())
env2 = TransformedEnv(
    env2,
    Compose(
        CatTensors(env2_obs_keys, "observation_walk", del_keys=False),
        CatTensors(env2_obs_keys, "observation"),
        DoubleToFloat(
            in_keys=["observation_walk", "observation"],
            in_keys_inv=["action"],
        ),
    ),
)

tdreset1 = env1.reset()
tdreset2 = env2.reset()

# With LazyStackedTensorDict, stacking is done in a lazy manner: the original tensordicts
# can still be recovered by indexing the main tensordict
tdreset = LazyStackedTensorDict.lazy_stack([tdreset1, tdreset2], 0)
assert tdreset[0] is tdreset1

print(tdreset[0])

TensorDict(
    fields={
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_stand: Tensor(shape=torch.Size([67]), 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)

策略¶

我們將設計一個策略，其中主幹讀取 "observation" 鍵。然後，特定的子元件將準備好堆疊的 tensordicts 的 "observation_stand" 和 "observation_walk" 鍵（如果存在），並將它們傳遞到專用的子網路。

action_dim = env1.action_spec.shape[-1]

policy_common = TensorDictModule(
    nn.Linear(67, 64), in_keys=["observation"], out_keys=["hidden"]
)
policy_stand = TensorDictModule(
    MLP(67 + 64, action_dim, depth=2),
    in_keys=["observation_stand", "hidden"],
    out_keys=["action"],
)
policy_walk = TensorDictModule(
    MLP(67 + 64, action_dim, depth=2),
    in_keys=["observation_walk", "hidden"],
    out_keys=["action"],
)
seq = TensorDictSequential(
    policy_common, policy_stand, policy_walk, partial_tolerant=True
)

讓我們檢查一下我們的序列是否為單個環境 (stand) 輸出動作。

seq(env1.reset())

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_stand: Tensor(shape=torch.Size([67]), 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)

讓我們檢查一下我們的序列是否為單個環境 (walk) 輸出動作。

seq(env2.reset())

TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_walk: Tensor(shape=torch.Size([67]), 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)

這也適用於堆疊：現在站立和行走鍵已消失，因為它們不是所有 tensordicts 共享的。但是 TensorDictSequential 仍然執行了操作。請注意，主幹是以向量化方式執行的 - 而不是在迴圈中 - 這更有效率。

seq(tdreset)

LazyStackedTensorDict(
    fields={
        action: Tensor(shape=torch.Size([2, 21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([2, 64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    exclusive_fields={
        0 ->
            observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        1 ->
            observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)

並行執行各種任務¶

如果常見的鍵值對共享相同的規格（尤其是它們的形狀和 dtype 必須匹配），我們可以並行化操作：如果觀察形狀不同但指向相同的鍵，則無法執行以下操作。

如果 ParallelEnv 收到單個環境建立函數，它將假設必須執行單個任務。如果提供了函數列表，則它將假設我們處於多任務設定中。

def env1_maker():
    return TransformedEnv(
        DMControlEnv("humanoid", "stand"),
        Compose(
            CatTensors(env1_obs_keys, "observation_stand", del_keys=False),
            CatTensors(env1_obs_keys, "observation"),
            DoubleToFloat(
                in_keys=["observation_stand", "observation"],
                in_keys_inv=["action"],
            ),
        ),
    )


def env2_maker():
    return TransformedEnv(
        DMControlEnv("humanoid", "walk"),
        Compose(
            CatTensors(env2_obs_keys, "observation_walk", del_keys=False),
            CatTensors(env2_obs_keys, "observation"),
            DoubleToFloat(
                in_keys=["observation_walk", "observation"],
                in_keys_inv=["action"],
            ),
        ),
    )


env = ParallelEnv(2, [env1_maker, env2_maker])
assert not env._single_task

tdreset = env.reset()
print(tdreset)
print(tdreset[0])
print(tdreset[1])  # should be different

LazyStackedTensorDict(
    fields={
        done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    exclusive_fields={
        0 ->
            observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        1 ->
            observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)
TensorDict(
    fields={
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_stand: Tensor(shape=torch.Size([67]), 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)
TensorDict(
    fields={
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_walk: Tensor(shape=torch.Size([67]), 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)

讓我們將輸出傳遞到我們的網路中。

tdreset = seq(tdreset)
print(tdreset)
print(tdreset[0])
print(tdreset[1])  # should be different but all have an "action" key


env.step(tdreset)  # computes actions and execute steps in parallel
print(tdreset)
print(tdreset[0])
print(tdreset[1])  # next_observation has now been written

LazyStackedTensorDict(
    fields={
        action: Tensor(shape=torch.Size([2, 21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([2, 64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    exclusive_fields={
        0 ->
            observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        1 ->
            observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_stand: Tensor(shape=torch.Size([67]), 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)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_walk: Tensor(shape=torch.Size([67]), 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)
LazyStackedTensorDict(
    fields={
        action: Tensor(shape=torch.Size([2, 21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([2, 64]), device=cpu, dtype=torch.float32, is_shared=False),
        next: LazyStackedTensorDict(
            fields={
                done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.float64, is_shared=False),
                terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            exclusive_fields={
                0 ->
                    observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
                1 ->
                    observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([2]),
            device=None,
            is_shared=False,
            stack_dim=0),
        observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    exclusive_fields={
        0 ->
            observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        1 ->
            observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, 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([67]), device=cpu, dtype=torch.float32, is_shared=False),
                observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float64, 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([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_stand: Tensor(shape=torch.Size([67]), 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)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.float32, 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([67]), device=cpu, dtype=torch.float32, is_shared=False),
                observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float64, 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([67]), device=cpu, dtype=torch.float32, is_shared=False),
        observation_walk: Tensor(shape=torch.Size([67]), 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)

Rollout¶

td_rollout = env.rollout(100, policy=seq, return_contiguous=False)

td_rollout[:, 0]  # tensordict of the first step: only the common keys are shown

LazyStackedTensorDict(
    fields={
        action: Tensor(shape=torch.Size([2, 21]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        hidden: Tensor(shape=torch.Size([2, 64]), device=cpu, dtype=torch.float32, is_shared=False),
        next: LazyStackedTensorDict(
            fields={
                done: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.float64, is_shared=False),
                terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            exclusive_fields={
                0 ->
                    observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
                1 ->
                    observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([2]),
            device=None,
            is_shared=False,
            stack_dim=0),
        observation: Tensor(shape=torch.Size([2, 67]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        truncated: Tensor(shape=torch.Size([2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    exclusive_fields={
        0 ->
            observation_stand: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False),
        1 ->
            observation_walk: Tensor(shape=torch.Size([67]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)

td_rollout[0]  # tensordict of the first env: the stand obs is present

env.close()
del env

腳本的總執行時間：（0 分鐘 51.107 秒）

預估記憶體用量： 322 MB

由 Sphinx-Gallery 產生的圖庫

多任務環境中的任務特定策略¶

策略¶

並行執行各種任務¶

Rollout¶

文件

教學

資源