Source code for torchrl.objectives.cql
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations
import math
import warnings
from copy import deepcopy
from dataclasses import dataclass
from typing import List, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn as nn
from tensordict import TensorDict, TensorDictBase, TensorDictParams
from tensordict.nn import dispatch, TensorDictModule
from tensordict.utils import NestedKey, unravel_key
from torch import Tensor
from torchrl.data.tensor_specs import Composite
from torchrl.data.utils import _find_action_space
from torchrl.envs.utils import ExplorationType, set_exploration_type
from torchrl.modules import ProbabilisticActor, QValueActor
from torchrl.modules.tensordict_module.common import ensure_tensordict_compatible
from torchrl.objectives.common import LossModule
from torchrl.objectives.utils import (
_cache_values,
_GAMMA_LMBDA_DEPREC_ERROR,
_reduce,
_vmap_func,
default_value_kwargs,
distance_loss,
ValueEstimators,
)
from torchrl.objectives.value import TD0Estimator, TD1Estimator, TDLambdaEstimator
[docs]class CQLLoss(LossModule):
"""TorchRL implementation of the continuous CQL loss.
Presented in "Conservative Q-Learning for Offline Reinforcement Learning" https://arxiv.org/abs/2006.04779
Args:
actor_network (ProbabilisticActor): stochastic actor
qvalue_network (TensorDictModule or list of TensorDictModule): Q(s, a) parametric model.
This module typically outputs a ``"state_action_value"`` entry.
If a single instance of `qvalue_network` is provided, it will be duplicated ``N``
times (where ``N=2`` for this loss). If a list of modules is passed, their
parameters will be stacked unless they share the same identity (in which case
the original parameter will be expanded).
.. warning:: When a list of parameters if passed, it will __not__ be compared against the policy parameters
and all the parameters will be considered as untied.
Keyword args:
loss_function (str, optional): loss function to be used with
the value function loss. Default is `"smooth_l1"`.
alpha_init (:obj:`float`, optional): initial entropy multiplier.
Default is 1.0.
min_alpha (:obj:`float`, optional): min value of alpha.
Default is None (no minimum value).
max_alpha (:obj:`float`, optional): max value of alpha.
Default is None (no maximum value).
action_spec (TensorSpec, optional): the action tensor spec. If not provided
and the target entropy is ``"auto"``, it will be retrieved from
the actor.
fixed_alpha (bool, optional): if ``True``, alpha will be fixed to its
initial value. Otherwise, alpha will be optimized to
match the 'target_entropy' value.
Default is ``False``.
target_entropy (float or str, optional): Target entropy for the
stochastic policy. Default is "auto", where target entropy is
computed as :obj:`-prod(n_actions)`.
delay_actor (bool, optional): Whether to separate the target actor
networks from the actor networks used for data collection.
Default is ``False``.
delay_qvalue (bool, optional): Whether to separate the target Q value
networks from the Q value networks used for data collection.
Default is ``True``.
gamma (:obj:`float`, optional): Discount factor. Default is ``None``.
temperature (:obj:`float`, optional): CQL temperature. Default is 1.0.
min_q_weight (:obj:`float`, optional): Minimum Q weight. Default is 1.0.
max_q_backup (bool, optional): Whether to use the max-min Q backup.
Default is ``False``.
deterministic_backup (bool, optional): Whether to use the deterministic. Default is ``True``.
num_random (int, optional): Number of random actions to sample for the CQL loss.
Default is 10.
with_lagrange (bool, optional): Whether to use the Lagrange multiplier.
Default is ``False``.
lagrange_thresh (:obj:`float`, optional): Lagrange threshold. Default is 0.0.
reduction (str, optional): Specifies the reduction to apply to the output:
``"none"`` | ``"mean"`` | ``"sum"``. ``"none"``: no reduction will be applied,
``"mean"``: the sum of the output will be divided by the number of
elements in the output, ``"sum"``: the output will be summed. Default: ``"mean"``.
Examples:
>>> import torch
>>> from torch import nn
>>> from torchrl.data import Bounded
>>> from torchrl.modules.distributions import NormalParamExtractor, TanhNormal
>>> from torchrl.modules.tensordict_module.actors import ProbabilisticActor, ValueOperator
>>> from torchrl.modules.tensordict_module.common import SafeModule
>>> from torchrl.objectives.cql import CQLLoss
>>> from tensordict import TensorDict
>>> n_act, n_obs = 4, 3
>>> spec = Bounded(-torch.ones(n_act), torch.ones(n_act), (n_act,))
>>> net = nn.Sequential(nn.Linear(n_obs, 2 * n_act), NormalParamExtractor())
>>> module = SafeModule(net, in_keys=["observation"], out_keys=["loc", "scale"])
>>> actor = ProbabilisticActor(
... module=module,
... in_keys=["loc", "scale"],
... spec=spec,
... distribution_class=TanhNormal)
>>> class ValueClass(nn.Module):
... def __init__(self):
... super().__init__()
... self.linear = nn.Linear(n_obs + n_act, 1)
... def forward(self, obs, act):
... return self.linear(torch.cat([obs, act], -1))
>>> module = ValueClass()
>>> qvalue = ValueOperator(
... module=module,
... in_keys=['observation', 'action'])
>>> loss = CQLLoss(actor, qvalue)
>>> batch = [2, ]
>>> action = spec.rand(batch)
>>> 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={
alpha: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
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_actor_bc: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
loss_alpha: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
loss_cql: 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)},
batch_size=torch.Size([]),
device=None,
is_shared=False)
This class is compatible with non-tensordict based modules too and can be
used without recurring to any tensordict-related primitive. In this case,
the expected keyword arguments are:
``["action", "next_reward", "next_done", "next_terminated"]`` + in_keys of the actor, value, and qvalue network.
The return value is a tuple of tensors in the following order:
``["loss_actor", "loss_qvalue", "loss_alpha", "loss_alpha_prime", "alpha", "entropy"]``.
Examples:
>>> import torch
>>> from torch import nn
>>> from torchrl.data import Bounded
>>> from torchrl.modules.distributions import NormalParamExtractor, TanhNormal
>>> from torchrl.modules.tensordict_module.actors import ProbabilisticActor, ValueOperator
>>> from torchrl.modules.tensordict_module.common import SafeModule
>>> from torchrl.objectives.cql import CQLLoss
>>> _ = torch.manual_seed(42)
>>> n_act, n_obs = 4, 3
>>> spec = Bounded(-torch.ones(n_act), torch.ones(n_act), (n_act,))
>>> net = nn.Sequential(nn.Linear(n_obs, 2 * n_act), NormalParamExtractor())
>>> module = SafeModule(net, in_keys=["observation"], out_keys=["loc", "scale"])
>>> actor = ProbabilisticActor(
... module=module,
... in_keys=["loc", "scale"],
... spec=spec,
... distribution_class=TanhNormal)
>>> class ValueClass(nn.Module):
... def __init__(self):
... super().__init__()
... self.linear = nn.Linear(n_obs + n_act, 1)
... def forward(self, obs, act):
... return self.linear(torch.cat([obs, act], -1))
>>> module = ValueClass()
>>> qvalue = ValueOperator(
... module=module,
... in_keys=['observation', 'action'])
>>> loss = CQLLoss(actor, qvalue)
>>> batch = [2, ]
>>> action = spec.rand(batch)
>>> loss_actor, loss_actor_bc, loss_qvalue, loss_cql, *_ = 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()
The output keys can also be filtered using the :meth:`CQLLoss.select_out_keys`
method.
Examples:
>>> _ = loss.select_out_keys('loss_actor', 'loss_qvalue')
>>> loss_actor, loss_qvalue = 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()
"""
@dataclass
class _AcceptedKeys:
"""Maintains default values for all configurable tensordict keys.
This class defines which tensordict keys can be set using '.set_keys(key_name=key_value)' and their
default values.
Attributes:
action (NestedKey): The input tensordict key where the action is expected.
Defaults to ``"advantage"``.
value (NestedKey): The input tensordict key where the state value is expected.
Will be used for the underlying value estimator. Defaults to ``"state_value"``.
state_action_value (NestedKey): The input tensordict key where the
state action value is expected. Defaults to ``"state_action_value"``.
log_prob (NestedKey): The input tensordict key where the log probability is expected.
Defaults to ``"_log_prob"``.
pred_q1 (NestedKey): The input tensordict key where the predicted Q1 values are expected.
Defaults to ``"pred_q1"``.
pred_q2 (NestedKey): The input tensordict key where the predicted Q2 values are expected.
Defaults to ``"pred_q2"``.
priority (NestedKey): The input tensordict key where the target priority is written to.
Defaults to ``"td_error"``.
cql_q1_loss (NestedKey): The input tensordict key where the CQL Q1 loss is expected.
Defaults to ``"cql_q1_loss"``.
cql_q2_loss (NestedKey): The input tensordict key where the CQL Q2 loss is expected.
Defaults to ``"cql_q2_loss"``.
reward (NestedKey): The input tensordict key where the reward is expected.
Defaults to ``"reward"``.
done (NestedKey): The input tensordict key where the done flag is expected.
Defaults to ``"done"``.
terminated (NestedKey): The input tensordict key where the terminated flag is expected.
Defaults to ``"terminated"``.
"""
action: NestedKey = "action"
value: NestedKey = "state_value"
state_action_value: NestedKey = "state_action_value"
log_prob: NestedKey = "_log_prob"
pred_q1: NestedKey = "pred_q1"
pred_q2: NestedKey = "pred_q2"
priority: NestedKey = "td_error"
cql_q1_loss: NestedKey = "cql_q1_loss"
cql_q2_loss: NestedKey = "cql_q2_loss"
priority: NestedKey = "td_error"
reward: NestedKey = "reward"
done: NestedKey = "done"
terminated: NestedKey = "terminated"
default_keys = _AcceptedKeys()
default_value_estimator = ValueEstimators.TD0
actor_network: TensorDictModule
qvalue_network: TensorDictModule
actor_network_params: TensorDictParams
qvalue_network_params: TensorDictParams
target_actor_network_params: TensorDictParams
target_qvalue_network_params: TensorDictParams
def __init__(
self,
actor_network: ProbabilisticActor,
qvalue_network: TensorDictModule | List[TensorDictModule],
*,
loss_function: str = "smooth_l1",
alpha_init: float = 1.0,
min_alpha: float = None,
max_alpha: float = None,
action_spec=None,
fixed_alpha: bool = False,
target_entropy: Union[str, float] = "auto",
delay_actor: bool = False,
delay_qvalue: bool = True,
gamma: float = None,
temperature: float = 1.0,
min_q_weight: float = 1.0,
max_q_backup: bool = False,
deterministic_backup: bool = True,
num_random: int = 10,
with_lagrange: bool = False,
lagrange_thresh: float = 0.0,
reduction: str = None,
) -> None:
self._out_keys = None
if reduction is None:
reduction = "mean"
super().__init__()
# Actor
self.delay_actor = delay_actor
self.convert_to_functional(
actor_network,
"actor_network",
create_target_params=self.delay_actor,
)
# Q value
self.delay_qvalue = delay_qvalue
self.num_qvalue_nets = 2
self.convert_to_functional(
qvalue_network,
"qvalue_network",
self.num_qvalue_nets,
create_target_params=self.delay_qvalue,
compare_against=list(actor_network.parameters()),
)
self.loss_function = loss_function
try:
device = next(self.parameters()).device
except AttributeError:
device = torch.device("cpu")
self.register_buffer("alpha_init", torch.tensor(alpha_init, device=device))
if bool(min_alpha) ^ bool(max_alpha):
min_alpha = min_alpha if min_alpha else 0.0
if max_alpha == 0:
raise ValueError("max_alpha must be either None or greater than 0.")
max_alpha = max_alpha if max_alpha else 1e9
if min_alpha:
self.register_buffer(
"min_log_alpha", torch.tensor(min_alpha, device=device).log()
)
else:
self.min_log_alpha = None
if max_alpha:
self.register_buffer(
"max_log_alpha", torch.tensor(max_alpha, device=device).log()
)
else:
self.max_log_alpha = None
self.fixed_alpha = fixed_alpha
if fixed_alpha:
self.register_buffer(
"log_alpha", torch.tensor(math.log(alpha_init), device=device)
)
else:
self.register_parameter(
"log_alpha",
torch.nn.Parameter(torch.tensor(math.log(alpha_init), device=device)),
)
self._target_entropy = target_entropy
self._action_spec = action_spec
self.target_entropy_buffer = None
if gamma is not None:
raise TypeError(_GAMMA_LMBDA_DEPREC_ERROR)
self.temperature = temperature
self.min_q_weight = min_q_weight
self.max_q_backup = max_q_backup
self.deterministic_backup = deterministic_backup
self.num_random = num_random
self.with_lagrange = with_lagrange
if self.with_lagrange:
self.target_action_gap = lagrange_thresh
self.register_parameter(
"log_alpha_prime",
torch.nn.Parameter(torch.tensor(math.log(1.0), device=device)),
)
self._make_vmap()
self.reduction = reduction
_ = self.target_entropy
def _make_vmap(self):
self._vmap_qvalue_networkN0 = _vmap_func(
self.qvalue_network, (None, 0), randomness=self.vmap_randomness
)
self._vmap_qvalue_network00 = _vmap_func(
self.qvalue_network, randomness=self.vmap_randomness
)
@property
def target_entropy(self):
target_entropy = self.target_entropy_buffer
if target_entropy is None:
delattr(self, "target_entropy_buffer")
target_entropy = self._target_entropy
action_spec = self._action_spec
actor_network = self.actor_network
device = next(self.parameters()).device
if target_entropy == "auto":
action_spec = (
action_spec
if action_spec is not None
else getattr(actor_network, "spec", None)
)
if action_spec is None:
raise RuntimeError(
"Cannot infer the dimensionality of the action. Consider providing "
"the target entropy explicitely or provide the spec of the "
"action tensor in the actor network."
)
if not isinstance(action_spec, Composite):
action_spec = Composite({self.tensor_keys.action: action_spec})
if (
isinstance(self.tensor_keys.action, tuple)
and len(self.tensor_keys.action) > 1
):
action_container_shape = action_spec[
self.tensor_keys.action[:-1]
].shape
else:
action_container_shape = action_spec.shape
target_entropy = -float(
action_spec[self.tensor_keys.action]
.shape[len(action_container_shape) :]
.numel()
)
self.register_buffer(
"target_entropy_buffer", torch.tensor(target_entropy, device=device)
)
return self.target_entropy_buffer
return target_entropy
def _forward_value_estimator_keys(self, **kwargs) -> None:
if self._value_estimator is not None:
self._value_estimator.set_keys(
value=self._tensor_keys.value,
reward=self.tensor_keys.reward,
done=self.tensor_keys.done,
terminated=self.tensor_keys.terminated,
)
[docs] def make_value_estimator(self, value_type: ValueEstimators = None, **hyperparams):
if value_type is None:
value_type = self.default_value_estimator
self.value_type = value_type
# we will take care of computing the next value inside this module
value_net = None
hp = dict(default_value_kwargs(value_type))
hp.update(hyperparams)
if value_type is ValueEstimators.TD1:
self._value_estimator = TD1Estimator(
**hp,
value_network=value_net,
)
elif value_type is ValueEstimators.TD0:
self._value_estimator = TD0Estimator(
**hp,
value_network=value_net,
)
elif value_type is ValueEstimators.GAE:
raise NotImplementedError(
f"Value type {value_type} it not implemented for loss {type(self)}."
)
elif value_type is ValueEstimators.TDLambda:
self._value_estimator = TDLambdaEstimator(
**hp,
value_network=value_net,
)
else:
raise NotImplementedError(f"Unknown value type {value_type}")
tensor_keys = {
"value_target": "value_target",
"value": self.tensor_keys.value,
"reward": self.tensor_keys.reward,
"done": self.tensor_keys.done,
"terminated": self.tensor_keys.terminated,
}
self._value_estimator.set_keys(**tensor_keys)
@property
def in_keys(self):
keys = [
self.tensor_keys.action,
("next", self.tensor_keys.reward),
("next", self.tensor_keys.done),
("next", self.tensor_keys.terminated),
*self.actor_network.in_keys,
*[("next", key) for key in self.actor_network.in_keys],
*self.qvalue_network.in_keys,
]
return list(set(keys))
@property
def out_keys(self):
if self._out_keys is None:
keys = [
"loss_actor",
"loss_actor_bc",
"loss_qvalue",
"loss_cql",
"loss_alpha",
"alpha",
"entropy",
]
if self.with_lagrange:
keys.append("loss_alpha_prime")
self._out_keys = keys
return self._out_keys
@out_keys.setter
def out_keys(self, values):
self._out_keys = values
[docs] @dispatch
def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
q_loss, metadata = self.q_loss(tensordict)
cql_loss, cql_metadata = self.cql_loss(tensordict)
if self.with_lagrange:
alpha_prime_loss, alpha_prime_metadata = self.alpha_prime_loss(tensordict)
metadata.update(alpha_prime_metadata)
loss_actor_bc, bc_metadata = self.actor_bc_loss(tensordict)
loss_actor, actor_metadata = self.actor_loss(tensordict)
loss_alpha, alpha_metadata = self.alpha_loss(actor_metadata)
metadata.update(bc_metadata)
metadata.update(cql_metadata)
metadata.update(actor_metadata)
metadata.update(alpha_metadata)
tensordict.set(
self.tensor_keys.priority, metadata.pop("td_error").detach().max(0).values
)
out = {
"loss_actor": loss_actor,
"loss_actor_bc": loss_actor_bc,
"loss_qvalue": q_loss,
"loss_cql": cql_loss,
"loss_alpha": loss_alpha,
"alpha": self._alpha,
"entropy": -actor_metadata.get(self.tensor_keys.log_prob).mean().detach(),
}
if self.with_lagrange:
out["loss_alpha_prime"] = alpha_prime_loss.mean()
return TensorDict(out, [])
@property
@_cache_values
def _cached_detach_qvalue_params(self):
return self.qvalue_network_params.detach()
def actor_bc_loss(self, tensordict: TensorDictBase) -> Tensor:
with set_exploration_type(
ExplorationType.RANDOM
), self.actor_network_params.to_module(self.actor_network):
dist = self.actor_network.get_dist(
tensordict,
)
a_reparm = dist.rsample()
log_prob = dist.log_prob(a_reparm)
bc_log_prob = dist.log_prob(tensordict.get(self.tensor_keys.action))
bc_actor_loss = self._alpha * log_prob - bc_log_prob
bc_actor_loss = _reduce(bc_actor_loss, reduction=self.reduction)
metadata = {"bc_log_prob": bc_log_prob.mean().detach()}
return bc_actor_loss, metadata
def actor_loss(self, tensordict: TensorDictBase) -> Tuple[Tensor, dict]:
with set_exploration_type(
ExplorationType.RANDOM
), self.actor_network_params.to_module(self.actor_network):
dist = self.actor_network.get_dist(
tensordict,
)
a_reparm = dist.rsample()
log_prob = dist.log_prob(a_reparm)
td_q = tensordict.select(*self.qvalue_network.in_keys, strict=False)
if td_q is tensordict:
raise RuntimeError
td_q.set(self.tensor_keys.action, a_reparm)
td_q = self._vmap_qvalue_networkN0(
td_q,
self._cached_detach_qvalue_params,
)
min_q_logprob = (
td_q.get(self.tensor_keys.state_action_value).min(0)[0].squeeze(-1)
)
if log_prob.shape != min_q_logprob.shape:
raise RuntimeError(
f"Losses shape mismatch: {log_prob.shape} and {min_q_logprob.shape}"
)
metadata = {}
metadata[self.tensor_keys.log_prob] = log_prob.detach()
actor_loss = self._alpha * log_prob - min_q_logprob
actor_loss = _reduce(actor_loss, reduction=self.reduction)
return actor_loss, metadata
def _get_policy_actions(self, data, actor_params, num_actions=10):
batch_size = data.batch_size
batch_size = list(batch_size[:-1]) + [batch_size[-1] * num_actions]
in_keys = [unravel_key(key) for key in self.actor_network.in_keys]
def filter_and_repeat(name, x):
if name in in_keys:
return x.repeat_interleave(num_actions, dim=data.ndim - 1)
tensordict = data.named_apply(
filter_and_repeat, batch_size=batch_size, filter_empty=True
)
with torch.no_grad():
with set_exploration_type(ExplorationType.RANDOM), actor_params.to_module(
self.actor_network
):
dist = self.actor_network.get_dist(tensordict)
action = dist.rsample()
tensordict.set(self.tensor_keys.action, action)
sample_log_prob = dist.log_prob(action)
# tensordict.del_("loc")
# tensordict.del_("scale")
return (
tensordict.select(
*self.actor_network.in_keys, self.tensor_keys.action, strict=False
),
sample_log_prob,
)
def _get_value_v(self, tensordict, _alpha, actor_params, qval_params):
tensordict = tensordict.clone(False)
# get actions and log-probs
with torch.no_grad():
with set_exploration_type(ExplorationType.RANDOM), actor_params.to_module(
self.actor_network
):
next_tensordict = tensordict.get("next").clone(False)
next_dist = self.actor_network.get_dist(next_tensordict)
next_action = next_dist.rsample()
next_tensordict.set(self.tensor_keys.action, next_action)
next_sample_log_prob = next_dist.log_prob(next_action)
# get q-values
if not self.max_q_backup:
next_tensordict_expand = self._vmap_qvalue_networkN0(
next_tensordict, qval_params
)
next_state_value = next_tensordict_expand.get(
self.tensor_keys.state_action_value
).min(0)[0]
if (
next_state_value.shape[-len(next_sample_log_prob.shape) :]
!= next_sample_log_prob.shape
):
next_sample_log_prob = next_sample_log_prob.unsqueeze(-1)
if not self.deterministic_backup:
next_state_value = next_state_value - _alpha * next_sample_log_prob
if self.max_q_backup:
next_tensordict, _ = self._get_policy_actions(
tensordict.get("next").copy(),
actor_params,
num_actions=self.num_random,
)
next_tensordict_expand = self._vmap_qvalue_networkN0(
next_tensordict, qval_params
)
state_action_value = next_tensordict_expand.get(
self.tensor_keys.state_action_value
)
# take max over actions
state_action_value = state_action_value.reshape(
torch.Size(
[self.num_qvalue_nets, *tensordict.shape, self.num_random, -1]
)
).max(-2)[0]
# take min over qvalue nets
next_state_value = state_action_value.min(0)[0]
tensordict.set(
("next", self.value_estimator.tensor_keys.value), next_state_value
)
target_value = self.value_estimator.value_estimate(tensordict).squeeze(-1)
return target_value
def q_loss(self, tensordict: TensorDictBase) -> Tuple[Tensor, dict]:
# we pass the alpha value to the tensordict. Since it's a scalar, we must erase the batch-size first.
target_value = self._get_value_v(
tensordict.copy(),
self._alpha,
self.actor_network_params,
self.target_qvalue_network_params,
)
tensordict_pred_q = tensordict.select(
*self.qvalue_network.in_keys, strict=False
)
q_pred = self._vmap_qvalue_networkN0(
tensordict_pred_q, self.qvalue_network_params
).get(self.tensor_keys.state_action_value)
# write pred values in tensordict for cql loss
tensordict.set(self.tensor_keys.pred_q1, q_pred[0])
tensordict.set(self.tensor_keys.pred_q2, q_pred[1])
q_pred = q_pred.squeeze(-1)
loss_qval = distance_loss(
q_pred,
target_value.expand_as(q_pred),
loss_function=self.loss_function,
).sum(0)
loss_qval = _reduce(loss_qval, reduction=self.reduction)
td_error = (q_pred - target_value).pow(2)
metadata = {"td_error": td_error.detach()}
return loss_qval, metadata
def cql_loss(self, tensordict: TensorDictBase) -> Tuple[Tensor, dict]:
pred_q1 = tensordict.get(self.tensor_keys.pred_q1)
pred_q2 = tensordict.get(self.tensor_keys.pred_q2)
if pred_q1 is None:
raise KeyError(
f"Couldn't find the pred_q1 with key {self.tensor_keys.pred_q1} in the input tensordict. "
"This could be caused by calling cql_loss method before q_loss method."
)
if pred_q2 is None:
raise KeyError(
f"Couldn't find the pred_q2 with key {self.tensor_keys.pred_q2} in the input tensordict. "
"This could be caused by calling cql_loss method before q_loss method."
)
random_actions_tensor = pred_q1.new_empty(
(
*tensordict.shape[:-1],
tensordict.shape[-1] * self.num_random,
tensordict[self.tensor_keys.action].shape[-1],
)
).uniform_(-1, 1)
curr_actions_td, curr_log_pis = self._get_policy_actions(
tensordict.copy(),
self.actor_network_params,
num_actions=self.num_random,
)
new_curr_actions_td, new_log_pis = self._get_policy_actions(
tensordict.get("next").copy(),
self.actor_network_params,
num_actions=self.num_random,
)
# process all in one forward pass
# stack qvalue params
qvalue_params = torch.cat(
[
self.qvalue_network_params,
self.qvalue_network_params,
self.qvalue_network_params,
],
0,
)
# select and stack input params
# q value random action
tensordict_q_random = tensordict.select(
*self.actor_network.in_keys, strict=False
)
batch_size = tensordict_q_random.batch_size
batch_size = list(batch_size[:-1]) + [batch_size[-1] * self.num_random]
in_keys = [unravel_key(key) for key in self.actor_network.in_keys]
def filter_and_repeat(name, x):
if name in in_keys:
return x.repeat_interleave(
self.num_random, dim=tensordict_q_random.ndim - 1
)
tensordict_q_random = tensordict_q_random.named_apply(
filter_and_repeat,
batch_size=batch_size,
filter_empty=True,
)
tensordict_q_random.set(self.tensor_keys.action, random_actions_tensor)
cql_tensordict = torch.cat(
[
tensordict_q_random.expand(
self.num_qvalue_nets, *curr_actions_td.batch_size
),
curr_actions_td.expand(
self.num_qvalue_nets, *curr_actions_td.batch_size
),
new_curr_actions_td.expand(
self.num_qvalue_nets, *curr_actions_td.batch_size
),
],
0,
)
cql_tensordict = cql_tensordict.contiguous()
cql_tensordict_expand = self._vmap_qvalue_network00(
cql_tensordict, qvalue_params
)
# get q values
state_action_value = cql_tensordict_expand.get(
self.tensor_keys.state_action_value
)
# split q values
(q_random, q_curr, q_new,) = state_action_value.split(
[
self.num_qvalue_nets,
self.num_qvalue_nets,
self.num_qvalue_nets,
],
dim=0,
)
# importance sammpled version
random_density = np.log(
0.5 ** curr_actions_td[self.tensor_keys.action].shape[-1]
)
cat_q1 = torch.cat(
[
q_random[0] - random_density,
q_new[0] - new_log_pis.detach().unsqueeze(-1),
q_curr[0] - curr_log_pis.detach().unsqueeze(-1),
],
-1,
)
cat_q2 = torch.cat(
[
q_random[1] - random_density,
q_new[1] - new_log_pis.detach().unsqueeze(-1),
q_curr[1] - curr_log_pis.detach().unsqueeze(-1),
],
-1,
)
min_qf1_loss = (
torch.logsumexp(cat_q1 / self.temperature, dim=-1)
* self.min_q_weight
* self.temperature
)
min_qf2_loss = (
torch.logsumexp(cat_q2 / self.temperature, dim=-1)
* self.min_q_weight
* self.temperature
)
# Subtract the log likelihood of data
cql_q1_loss = min_qf1_loss.flatten() - pred_q1 * self.min_q_weight
cql_q2_loss = min_qf2_loss.flatten() - pred_q2 * self.min_q_weight
# write cql losses in tensordict for alpha prime loss
tensordict.set(self.tensor_keys.cql_q1_loss, cql_q1_loss)
tensordict.set(self.tensor_keys.cql_q2_loss, cql_q2_loss)
cql_q_loss = (cql_q1_loss + cql_q2_loss).mean(-1)
cql_q_loss = _reduce(cql_q_loss, reduction=self.reduction)
return cql_q_loss, {}
def alpha_prime_loss(self, tensordict: TensorDictBase) -> Tensor:
cql_q1_loss = tensordict.get(self.tensor_keys.cql_q1_loss)
cql_q2_loss = tensordict.get(self.tensor_keys.cql_q2_loss)
if cql_q1_loss is None:
raise KeyError(
f"Couldn't find the cql_q1_loss with key {self.tensor_keys.cql_q1_loss} in the input tensordict. "
"This could be caused by calling alpha_prime_loss method before cql_loss method."
)
if cql_q2_loss is None:
raise KeyError(
f"Couldn't find the cql_q2_loss with key {self.tensor_keys.cql_q2_loss} in the input tensordict. "
"This could be caused by calling alpha_prime_loss method before cql_loss method."
)
alpha_prime = torch.clamp_max(self.log_alpha_prime.exp(), max=1000000.0)
min_qf1_loss = alpha_prime * (cql_q1_loss.mean() - self.target_action_gap)
min_qf2_loss = alpha_prime * (cql_q2_loss.mean() - self.target_action_gap)
alpha_prime_loss = (-min_qf1_loss - min_qf2_loss) * 0.5
alpha_prime_loss = _reduce(alpha_prime_loss, reduction=self.reduction)
return alpha_prime_loss, {}
def alpha_loss(self, tensordict: TensorDictBase) -> Tensor:
log_pi = tensordict.get(self.tensor_keys.log_prob)
if self.target_entropy is not None:
# we can compute this loss even if log_alpha is not a parameter
alpha_loss = -self.log_alpha * (log_pi.detach() + self.target_entropy)
else:
# placeholder
alpha_loss = torch.zeros_like(log_pi)
alpha_loss = _reduce(alpha_loss, reduction=self.reduction)
return alpha_loss, {}
@property
def _alpha(self):
if self.min_log_alpha is not None:
self.log_alpha.data.clamp_(self.min_log_alpha, self.max_log_alpha)
with torch.no_grad():
alpha = self.log_alpha.exp()
return alpha
[docs]class DiscreteCQLLoss(LossModule):
"""TorchRL implementation of the discrete CQL loss.
This class implements the discrete conservative Q-learning (CQL) loss function, as presented in the paper
"Conservative Q-Learning for Offline Reinforcement Learning" (https://arxiv.org/abs/2006.04779).
Args:
value_network (Union[QValueActor, nn.Module]): The Q-value network used to estimate state-action values.
Keyword Args:
loss_function (Optional[str]): The distance function used to calculate the distance between the predicted
Q-values and the target Q-values. Defaults to ``l2``.
delay_value (bool): Whether to separate the target Q value
networks from the Q value networks used for data collection.
Default is ``True``.
gamma (:obj:`float`, optional): Discount factor. Default is ``None``.
action_space: The action space of the environment. If None, it is inferred from the value network.
Defaults to None.
reduction (str, optional): Specifies the reduction to apply to the output:
``"none"`` | ``"mean"`` | ``"sum"``. ``"none"``: no reduction will be applied,
``"mean"``: the sum of the output will be divided by the number of
elements in the output, ``"sum"``: the output will be summed. Default: ``"mean"``.
Examples:
>>> from torchrl.modules import MLP, QValueActor
>>> from torchrl.data import OneHot
>>> from torchrl.objectives import DiscreteCQLLoss
>>> n_obs, n_act = 4, 3
>>> value_net = MLP(in_features=n_obs, out_features=n_act)
>>> spec = OneHot(n_act)
>>> actor = QValueActor(value_net, in_keys=["observation"], action_space=spec)
>>> loss = DiscreteCQLLoss(actor, action_space=spec)
>>> batch = [10,]
>>> data = TensorDict({
... "observation": torch.randn(*batch, n_obs),
... "action": spec.rand(batch),
... ("next", "observation"): torch.randn(*batch, n_obs),
... ("next", "done"): torch.zeros(*batch, 1, dtype=torch.bool),
... ("next", "terminated"): torch.zeros(*batch, 1, dtype=torch.bool),
... ("next", "reward"): torch.randn(*batch, 1)
... }, batch)
>>> loss(data)
TensorDict(
fields={
loss_cql: 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),
pred_value: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
target_value: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
td_error: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False)},
batch_size=torch.Size([]),
device=None,
is_shared=False)
This class is compatible with non-tensordict based modules too and can be
used without recurring to any tensordict-related primitive. In this case,
the expected keyword arguments are:
``["observation", "next_observation", "action", "next_reward", "next_done", "next_terminated"]``,
and a single loss value is returned.
Examples:
>>> from torchrl.objectives import DiscreteCQLLoss
>>> from torchrl.data import OneHot
>>> from torch import nn
>>> import torch
>>> n_obs = 3
>>> n_action = 4
>>> action_spec = OneHot(n_action)
>>> value_network = nn.Linear(n_obs, n_action) # a simple value model
>>> dcql_loss = DiscreteCQLLoss(value_network, action_space=action_spec)
>>> # define data
>>> observation = torch.randn(n_obs)
>>> next_observation = torch.randn(n_obs)
>>> action = action_spec.rand()
>>> next_reward = torch.randn(1)
>>> next_done = torch.zeros(1, dtype=torch.bool)
>>> next_terminated = torch.zeros(1, dtype=torch.bool)
>>> loss_val = dcql_loss(
... observation=observation,
... next_observation=next_observation,
... next_reward=next_reward,
... next_done=next_done,
... next_terminated=next_terminated,
... action=action)
"""
@dataclass
class _AcceptedKeys:
"""Maintains default values for all configurable tensordict keys.
This class defines which tensordict keys can be set using '.set_keys(key_name=key_value)' and their
default values.
Attributes:
value_target (NestedKey): The input tensordict key where the target state value is expected.
Will be used for the underlying value estimator Defaults to ``"value_target"``.
value (NestedKey): The input tensordict key where the chosen action value is expected.
Will be used for the underlying value estimator. Defaults to ``"chosen_action_value"``.
action_value (NestedKey): The input tensordict key where the action value is expected.
Defaults to ``"action_value"``.
action (NestedKey): The input tensordict key where the action is expected.
Defaults to ``"action"``.
priority (NestedKey): The input tensordict key where the target priority is written to.
Defaults to ``"td_error"``.
reward (NestedKey): The input tensordict key where the reward is expected.
Will be used for the underlying value estimator. Defaults to ``"reward"``.
done (NestedKey): The key in the input TensorDict that indicates
whether a trajectory is done. Will be used for the underlying value estimator.
Defaults to ``"done"``.
terminated (NestedKey): The key in the input TensorDict that indicates
whether a trajectory is terminated. Will be used for the underlying value estimator.
Defaults to ``"terminated"``.
pred_val (NestedKey): The key where the predicted value will be written
in the input tensordict. This value is subsequently used by cql_loss.
Defaults to ``"pred_val"``.
"""
value_target: NestedKey = "value_target"
value: NestedKey = "chosen_action_value"
action_value: NestedKey = "action_value"
action: NestedKey = "action"
priority: NestedKey = "td_error"
reward: NestedKey = "reward"
done: NestedKey = "done"
terminated: NestedKey = "terminated"
pred_val: NestedKey = "pred_val"
default_keys = _AcceptedKeys()
default_value_estimator = ValueEstimators.TD0
out_keys = [
"loss_qvalue",
"loss_cql",
]
value_network: TensorDictModule
value_network_params: TensorDictParams
target_value_network_params: TensorDictParams
def __init__(
self,
value_network: Union[QValueActor, nn.Module],
*,
loss_function: Optional[str] = "l2",
delay_value: bool = True,
gamma: float = None,
action_space=None,
reduction: str = None,
) -> None:
self._in_keys = None
if reduction is None:
reduction = "mean"
super().__init__()
self.delay_value = delay_value
value_network = ensure_tensordict_compatible(
module=value_network,
wrapper_type=QValueActor,
action_space=action_space,
)
self.convert_to_functional(
value_network,
"value_network",
create_target_params=self.delay_value,
)
self.value_network_in_keys = value_network.in_keys
self.loss_function = loss_function
if action_space is None:
# infer from value net
if hasattr(value_network, "action_space"):
action_space = value_network.spec
else:
# let's try with action_space then
try:
action_space = value_network.action_space
except AttributeError:
raise ValueError(self.ACTION_SPEC_ERROR)
if action_space is None:
warnings.warn(
"action_space was not specified. DiscreteCQLLoss will default to 'one-hot'. "
"This behavior will be deprecated soon and a space will have to be passed. "
"Check the DiscreteCQLLoss documentation to see how to pass the action space."
)
action_space = "one-hot"
self.action_space = _find_action_space(action_space)
self.reduction = reduction
if gamma is not None:
raise TypeError(_GAMMA_LMBDA_DEPREC_ERROR)
def _forward_value_estimator_keys(self, **kwargs) -> None:
if self._value_estimator is not None:
self._value_estimator.set_keys(
value_target=self.tensor_keys.value_target,
value=self._tensor_keys.value,
reward=self._tensor_keys.reward,
done=self._tensor_keys.done,
terminated=self._tensor_keys.terminated,
)
self._set_in_keys()
def _set_in_keys(self):
in_keys = {
self.tensor_keys.action,
unravel_key(("next", self.tensor_keys.reward)),
unravel_key(("next", self.tensor_keys.done)),
unravel_key(("next", self.tensor_keys.terminated)),
*self.value_network.in_keys,
*[unravel_key(("next", key)) for key in self.value_network.in_keys],
}
self._in_keys = sorted(in_keys, key=str)
[docs] def make_value_estimator(self, value_type: ValueEstimators = None, **hyperparams):
if value_type is None:
value_type = self.default_value_estimator
self.value_type = value_type
# we will take care of computing the next value inside this module
value_net = deepcopy(self.value_network)
self.value_network_params.to_module(value_net, return_swap=False)
hp = dict(default_value_kwargs(value_type))
hp.update(hyperparams)
if value_type is ValueEstimators.TD1:
self._value_estimator = TD1Estimator(
**hp,
value_network=value_net,
)
elif value_type is ValueEstimators.TD0:
self._value_estimator = TD0Estimator(
**hp,
value_network=value_net,
)
elif value_type is ValueEstimators.GAE:
raise NotImplementedError(
f"Value type {value_type} it not implemented for loss {type(self)}."
)
elif value_type is ValueEstimators.TDLambda:
self._value_estimator = TDLambdaEstimator(
**hp,
value_network=value_net,
)
else:
raise NotImplementedError(f"Unknown value type {value_type}")
tensor_keys = {
"value_target": "value_target",
"value": self.tensor_keys.value,
"reward": self.tensor_keys.reward,
"done": self.tensor_keys.done,
"terminated": self.tensor_keys.terminated,
}
self._value_estimator.set_keys(**tensor_keys)
@property
def in_keys(self):
if self._in_keys is None:
self._set_in_keys()
return self._in_keys
@in_keys.setter
def in_keys(self, values):
self._in_keys = values
@dispatch
def value_loss(
self,
tensordict: TensorDictBase,
) -> Tuple[torch.Tensor, dict]:
td_copy = tensordict.clone(False)
with self.value_network_params.to_module(self.value_network):
self.value_network(td_copy)
action = tensordict.get(self.tensor_keys.action)
pred_val = td_copy.get(self.tensor_keys.action_value)
if self.action_space == "categorical":
if action.shape != pred_val.shape:
# unsqueeze the action if it lacks on trailing singleton dim
action = action.unsqueeze(-1)
pred_val_index = torch.gather(pred_val, -1, index=action).squeeze(-1)
else:
action = action.to(torch.float)
pred_val_index = (pred_val * action).sum(-1)
# calculate target value
with torch.no_grad():
target_value = self.value_estimator.value_estimate(
td_copy, params=self._cached_detached_target_value_params
).squeeze(-1)
with torch.no_grad():
td_error = (pred_val_index - target_value).pow(2)
td_error = td_error.unsqueeze(-1)
tensordict.set(
self.tensor_keys.priority,
td_error,
inplace=True,
)
tensordict.set(
self.tensor_keys.pred_val,
pred_val,
inplace=True,
)
loss = 0.5 * distance_loss(pred_val_index, target_value, self.loss_function)
loss = _reduce(loss, reduction=self.reduction)
metadata = {
"td_error": td_error.mean(0).detach(),
"pred_value": pred_val.mean().detach(),
"target_value": target_value.mean().detach(),
}
return loss, metadata
[docs] @dispatch
def forward(self, tensordict: TensorDictBase) -> TensorDict:
"""Computes the (DQN) CQL loss given a tensordict sampled from the replay buffer.
This function will also write a "td_error" key that can be used by prioritized replay buffers to assign
a priority to items in the tensordict.
Args:
tensordict (TensorDictBase): a tensordict with keys ["action"] and the in_keys of
the value network (observations, "done", "terminated", "reward" in a "next" tensordict).
Returns:
a tensor containing the CQL loss.
"""
loss_qval, metadata = self.value_loss(tensordict)
loss_cql, _ = self.cql_loss(tensordict)
source = {
"loss_qvalue": loss_qval,
"loss_cql": loss_cql,
}
source.update(metadata)
td_out = TensorDict(
source=source,
batch_size=[],
)
return td_out
@property
@_cache_values
def _cached_detached_target_value_params(self):
return self.target_value_network_params.detach()
def cql_loss(self, tensordict):
qvalues = tensordict.get(self.tensor_keys.pred_val, default=None)
if qvalues is None:
raise KeyError(
"Couldn't find the predicted qvalue with key {self.tensor_keys.pred_val} in the input tensordict. "
"This could be caused by calling cql_loss method before value_loss."
)
current_action = tensordict.get(self.tensor_keys.action)
logsumexp = torch.logsumexp(qvalues, dim=-1, keepdim=True)
if self.action_space == "categorical":
if current_action.shape != qvalues.shape:
# unsqueeze the action if it lacks on trailing singleton dim
current_action = current_action.unsqueeze(-1)
q_a = qvalues.gather(-1, current_action)
else:
q_a = (qvalues * current_action).sum(dim=-1, keepdim=True)
loss_cql = (logsumexp - q_a).squeeze(-1)
loss_cql = _reduce(loss_cql, reduction=self.reduction)
return loss_cql, {}
