DiscreteIQLLoss

class torchrl.objectives.DiscreteIQLLoss(*args, **kwargs)

TorchRL 对离散 IQL 损失函数的实现。

选自论文“Offline Reinforcement Learning with Implicit Q-Learning” https://arxiv.org/abs/2110.06169

参数:

• actor_network (ProbabilisticActor) – 随机策略网络

• qvalue_network (TensorDictModule) – Q(s, a) 参数化模型。

• value_network (TensorDictModule, optional) – V(s) 参数化模型。

关键字参数:

• action_space (str or TensorSpec) – 动作空间。必须是 "one-hot", "mult_one_hot", "binary" 或 "categorical" 中的一个，或是相应 spec (torchrl.data.OneHot, torchrl.data.MultiOneHot, torchrl.data.Binary 或 torchrl.data.Categorical) 的实例。

• num_qvalue_nets (integer, optional) – 使用的 Q-Value 网络数量。默认为 2。

• loss_function (str, optional) – 用于值函数损失的损失函数。默认为 “smooth_l1”。

• temperature (float, optional) – 逆温度 (beta)。对于较小的超参数值，目标函数的行为类似于行为克隆；而对于较大的值，它试图恢复 Q 函数的最大值。

• expectile (float, optional) – expectile \(\tau\)。较大的 \(\tau\) 值对于需要动态规划（“stichting”）的 antmaze 任务至关重要。

• priority_key (str, optional) – [已弃用，请改用 .set_keys(priority_key=priority_key) 代替] 用于写入优先级的 tensordict 键（用于优先经验回放缓冲区）。默认为 “td_error”。

• separate_losses (bool, optional) – 如果为 True，策略和 critic 之间的共享参数将仅通过策略损失进行训练。默认为 False，即梯度会传播到策略和 critic 损失的共享参数。

• reduction (str, optional) – 指定应用于输出的归约方式："none" | "mean" | "sum"。"none"：不应用归约；"mean"：输出的总和将除以输出中的元素数量；"sum"：输出将被求和。默认值："mean"。

示例

>>> import torch
>>> from torch import nn
>>> from torchrl.data.tensor_specs import OneHot
>>> from torchrl.modules.distributions.discrete import OneHotCategorical
>>> from torchrl.modules.tensordict_module.actors import ProbabilisticActor
>>> from torchrl.modules.tensordict_module.common import SafeModule
>>> from torchrl.objectives.iql import DiscreteIQLLoss
>>> from tensordict import TensorDict
>>> n_act, n_obs = 4, 3
>>> spec = OneHot(n_act)
>>> module = SafeModule(nn.Linear(n_obs, n_act), in_keys=["observation"], out_keys=["logits"])
>>> actor = ProbabilisticActor(
...     module=module,
...     in_keys=["logits"],
...     out_keys=["action"],
...     spec=spec,
...     distribution_class=OneHotCategorical)
>>> qvalue = SafeModule(
...     nn.Linear(n_obs, n_act),
...     in_keys=["observation"],
...     out_keys=["state_action_value"],
... )
>>> value = SafeModule(
...     nn.Linear(n_obs, 1),
...     in_keys=["observation"],
...     out_keys=["state_value"],
... )
>>> loss = DiscreteIQLLoss(actor, qvalue, value)
>>> batch = [2, ]
>>> action = spec.rand(batch).long()
>>> data = TensorDict({
...         "observation": torch.randn(*batch, n_obs),
...         "action": action,
...         ("next", "done"): torch.zeros(*batch, 1, dtype=torch.bool),
...         ("next", "terminated"): torch.zeros(*batch, 1, dtype=torch.bool),
...         ("next", "reward"): torch.randn(*batch, 1),
...         ("next", "observation"): torch.randn(*batch, n_obs),
...     }, batch)
>>> loss(data)
TensorDict(
    fields={
        entropy: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        loss_actor: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        loss_qvalue: 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)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)

此类也兼容非基于 tensordict 的模块，并且无需使用任何 tensordict 相关原语即可使用。在这种情况下，预期的关键字参数为：["action", "next_reward", "next_done", "next_terminated"] + actor、value 和 qvalue 网络的 in_keys。返回值是一个按以下顺序排列的张量元组：["loss_actor", "loss_qvalue", "loss_value", "entropy"]。

示例

>>> import torch
>>> import torch
>>> from torch import nn
>>> from torchrl.data.tensor_specs import OneHot
>>> from torchrl.modules.distributions.discrete import OneHotCategorical
>>> from torchrl.modules.tensordict_module.actors import ProbabilisticActor
>>> from torchrl.modules.tensordict_module.common import SafeModule
>>> from torchrl.objectives.iql import DiscreteIQLLoss
>>> _ = torch.manual_seed(42)
>>> n_act, n_obs = 4, 3
>>> spec = OneHot(n_act)
>>> module = SafeModule(nn.Linear(n_obs, n_act), in_keys=["observation"], out_keys=["logits"])
>>> actor = ProbabilisticActor(
...     module=module,
...     in_keys=["logits"],
...     out_keys=["action"],
...     spec=spec,
...     distribution_class=OneHotCategorical)
>>> qvalue = SafeModule(
...     nn.Linear(n_obs, n_act),
...     in_keys=["observation"],
...     out_keys=["state_action_value"],
... )
>>> value = SafeModule(
...     nn.Linear(n_obs, 1),
...     in_keys=["observation"],
...     out_keys=["state_value"],
... )
>>> loss = DiscreteIQLLoss(actor, qvalue, value)
>>> batch = [2, ]
>>> action = spec.rand(batch).long()
>>> loss_actor, loss_qvalue, loss_value, entropy = loss(
...     observation=torch.randn(*batch, n_obs),
...     action=action,
...     next_done=torch.zeros(*batch, 1, dtype=torch.bool),
...     next_terminated=torch.zeros(*batch, 1, dtype=torch.bool),
...     next_observation=torch.zeros(*batch, n_obs),
...     next_reward=torch.randn(*batch, 1))
>>> loss_actor.backward()

输出键也可以使用 DiscreteIQLLoss.select_out_keys() 方法进行过滤。

示例

>>> _ = loss.select_out_keys('loss_actor', 'loss_qvalue', 'loss_value')
>>> loss_actor, loss_qvalue, loss_value = loss(
...     observation=torch.randn(*batch, n_obs),
...     action=action,
...     next_done=torch.zeros(*batch, 1, dtype=torch.bool),
...     next_terminated=torch.zeros(*batch, 1, dtype=torch.bool),
...     next_observation=torch.zeros(*batch, n_obs),
...     next_reward=torch.randn(*batch, 1))
>>> loss_actor.backward()

default_keys

_AcceptedKeys 的别名

forward(tensordict: TensorDictBase = None) → TensorDictBase

它旨在读取输入的 TensorDict 并返回另一个包含名为“loss*”的损失键的 tensordict。

将损失分解到其组成部分，然后训练器可以使用这些组件来记录训练过程中的各种损失值。输出 tensordict 中存在的其他标量值也将被记录。

参数:

tensordict – 包含计算损失所需值的输入 tensordict。

返回值:

一个不包含批处理维度的新 tensordict，其中包含各种损失标量，这些标量将命名为“loss*”。将损失值以这个名称返回至关重要，因为训练器会在反向传播之前读取它们。