b900fdf6f2
Closes #947 This removes all kwargs from all policy constructors. While doing that, I also improved several names and added a whole lot of TODOs. ## Functional changes: 1. Added possibility to pass None as `critic2` and `critic2_optim`. In fact, the default behavior then should cover the absolute majority of cases 2. Added a function called `clone_optimizer` as a temporary measure to support passing `critic2_optim=None` ## Breaking changes: 1. `action_space` is no longer optional. In fact, it already was non-optional, as there was a ValueError in BasePolicy.init. So now several examples were fixed to reflect that 2. `reward_normalization` removed from DDPG and children. It was never allowed to pass it as `True` there, an error would have been raised in `compute_n_step_reward`. Now I removed it from the interface 3. renamed `critic1` and similar to `critic`, in order to have uniform interfaces. Note that the `critic` in DDPG was optional for the sole reason that child classes used `critic1`. I removed this optionality (DDPG can't do anything with `critic=None`) 4. Several renamings of fields (mostly private to public, so backwards compatible) ## Additional changes: 1. Removed type and default declaration from docstring. This kind of duplication is really not necessary 2. Policy constructors are now only called using named arguments, not a fragile mixture of positional and named as before 5. Minor beautifications in typing and code 6. Generally shortened docstrings and made them uniform across all policies (hopefully) ## Comment: With these changes, several problems in tianshou's inheritance hierarchy become more apparent. I tried highlighting them for future work. --------- Co-authored-by: Dominik Jain <d.jain@appliedai.de>
109 lines
4.3 KiB
Python
109 lines
4.3 KiB
Python
from typing import Any
|
|
|
|
import gymnasium as gym
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
from tianshou.data import to_torch
|
|
from tianshou.data.types import RolloutBatchProtocol
|
|
from tianshou.policy import QRDQNPolicy
|
|
from tianshou.policy.base import TLearningRateScheduler
|
|
|
|
|
|
class DiscreteCQLPolicy(QRDQNPolicy):
|
|
"""Implementation of discrete Conservative Q-Learning algorithm. arXiv:2006.04779.
|
|
|
|
:param model: a model following the rules in
|
|
:class:`~tianshou.policy.BasePolicy`. (s -> logits)
|
|
:param optim: a torch.optim for optimizing the model.
|
|
:param action_space: Env's action space.
|
|
:param min_q_weight: the weight for the cql loss.
|
|
:param discount_factor: in [0, 1].
|
|
:param num_quantiles: the number of quantile midpoints in the inverse
|
|
cumulative distribution function of the value.
|
|
:param estimation_step: the number of steps to look ahead.
|
|
:param target_update_freq: the target network update frequency (0 if
|
|
you do not use the target network).
|
|
:param reward_normalization: normalize the **returns** to Normal(0, 1).
|
|
TODO: rename to return_normalization?
|
|
:param is_double: use double dqn.
|
|
:param clip_loss_grad: clip the gradient of the loss in accordance
|
|
with nature14236; this amounts to using the Huber loss instead of
|
|
the MSE loss.
|
|
:param observation_space: Env's observation space.
|
|
:param lr_scheduler: if not None, will be called in `policy.update()`.
|
|
|
|
.. seealso::
|
|
Please refer to :class:`~tianshou.policy.QRDQNPolicy` for more detailed
|
|
explanation.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
*,
|
|
model: torch.nn.Module,
|
|
optim: torch.optim.Optimizer,
|
|
action_space: gym.spaces.Discrete,
|
|
min_q_weight: float = 10.0,
|
|
discount_factor: float = 0.99,
|
|
num_quantiles: int = 200,
|
|
estimation_step: int = 1,
|
|
target_update_freq: int = 0,
|
|
reward_normalization: bool = False,
|
|
is_double: bool = True,
|
|
clip_loss_grad: bool = False,
|
|
observation_space: gym.Space | None = None,
|
|
lr_scheduler: TLearningRateScheduler | None = None,
|
|
) -> None:
|
|
super().__init__(
|
|
model=model,
|
|
optim=optim,
|
|
action_space=action_space,
|
|
discount_factor=discount_factor,
|
|
num_quantiles=num_quantiles,
|
|
estimation_step=estimation_step,
|
|
target_update_freq=target_update_freq,
|
|
reward_normalization=reward_normalization,
|
|
is_double=is_double,
|
|
clip_loss_grad=clip_loss_grad,
|
|
observation_space=observation_space,
|
|
lr_scheduler=lr_scheduler,
|
|
)
|
|
self.min_q_weight = min_q_weight
|
|
|
|
def learn(self, batch: RolloutBatchProtocol, *args: Any, **kwargs: Any) -> dict[str, float]:
|
|
if self._target and self._iter % self.freq == 0:
|
|
self.sync_weight()
|
|
self.optim.zero_grad()
|
|
weight = batch.pop("weight", 1.0)
|
|
all_dist = self(batch).logits
|
|
act = to_torch(batch.act, dtype=torch.long, device=all_dist.device)
|
|
curr_dist = all_dist[np.arange(len(act)), act, :].unsqueeze(2)
|
|
target_dist = batch.returns.unsqueeze(1)
|
|
# calculate each element's difference between curr_dist and target_dist
|
|
dist_diff = F.smooth_l1_loss(target_dist, curr_dist, reduction="none")
|
|
huber_loss = (
|
|
(dist_diff * (self.tau_hat - (target_dist - curr_dist).detach().le(0.0).float()).abs())
|
|
.sum(-1)
|
|
.mean(1)
|
|
)
|
|
qr_loss = (huber_loss * weight).mean()
|
|
# ref: https://github.com/ku2482/fqf-iqn-qrdqn.pytorch/
|
|
# blob/master/fqf_iqn_qrdqn/agent/qrdqn_agent.py L130
|
|
batch.weight = dist_diff.detach().abs().sum(-1).mean(1) # prio-buffer
|
|
# add CQL loss
|
|
q = self.compute_q_value(all_dist, None)
|
|
dataset_expec = q.gather(1, act.unsqueeze(1)).mean()
|
|
negative_sampling = q.logsumexp(1).mean()
|
|
min_q_loss = negative_sampling - dataset_expec
|
|
loss = qr_loss + min_q_loss * self.min_q_weight
|
|
loss.backward()
|
|
self.optim.step()
|
|
self._iter += 1
|
|
return {
|
|
"loss": loss.item(),
|
|
"loss/qr": qr_loss.item(),
|
|
"loss/cql": min_q_loss.item(),
|
|
}
|