maxhuettenrauch 522f7fbf98
Feature/dataclasses (#996)
This PR adds strict typing to the output of `update` and `learn` in all
policies. This will likely be the last large refactoring PR before the
next release (0.6.0, not 1.0.0), so it requires some attention. Several
difficulties were encountered on the path to that goal:

1. The policy hierarchy is actually "broken" in the sense that the keys
of dicts that were output by `learn` did not follow the same enhancement
(inheritance) pattern as the policies. This is a real problem and should
be addressed in the near future. Generally, several aspects of the
policy design and hierarchy might deserve a dedicated discussion.
2. Each policy needs to be generic in the stats return type, because one
might want to extend it at some point and then also extend the stats.
Even within the source code base this pattern is necessary in many
places.
3. The interaction between learn and update is a bit quirky, we
currently handle it by having update modify special field inside
TrainingStats, whereas all other fields are handled by learn.
4. The IQM module is a policy wrapper and required a
TrainingStatsWrapper. The latter relies on a bunch of black magic.

They were addressed by:
1. Live with the broken hierarchy, which is now made visible by bounds
in generics. We use type: ignore where appropriate.
2. Make all policies generic with bounds following the policy
inheritance hierarchy (which is incorrect, see above). We experimented a
bit with nested TrainingStats classes, but that seemed to add more
complexity and be harder to understand. Unfortunately, mypy thinks that
the code below is wrong, wherefore we have to add `type: ignore` to the
return of each `learn`

```python

T = TypeVar("T", bound=int)


def f() -> T:
  return 3
```

3. See above
4. Write representative tests for the `TrainingStatsWrapper`. Still, the
black magic might cause nasty surprises down the line (I am not proud of
it)...

Closes #933

---------

Co-authored-by: Maximilian Huettenrauch <m.huettenrauch@appliedai.de>
Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
2023-12-30 11:09:03 +01:00

221 lines
8.7 KiB
Python

from dataclasses import dataclass
from typing import Any, Literal, TypeVar, cast
import gymnasium as gym
import numpy as np
import torch
import torch.nn.functional as F
from tianshou.data import Batch, ReplayBuffer, to_numpy
from tianshou.data.types import FQFBatchProtocol, ObsBatchProtocol, RolloutBatchProtocol
from tianshou.policy import DQNPolicy, QRDQNPolicy
from tianshou.policy.base import TLearningRateScheduler
from tianshou.policy.modelfree.qrdqn import QRDQNTrainingStats
from tianshou.utils.net.discrete import FractionProposalNetwork, FullQuantileFunction
@dataclass(kw_only=True)
class FQFTrainingStats(QRDQNTrainingStats):
quantile_loss: float
fraction_loss: float
entropy_loss: float
TFQFTrainingStats = TypeVar("TFQFTrainingStats", bound=FQFTrainingStats)
class FQFPolicy(QRDQNPolicy[TFQFTrainingStats]):
"""Implementation of Fully-parameterized Quantile Function. arXiv:1911.02140.
:param model: a model following the rules in
:class:`~tianshou.policy.BasePolicy`. (s -> logits)
:param optim: a torch.optim for optimizing the model.
:param fraction_model: a FractionProposalNetwork for
proposing fractions/quantiles given state.
:param fraction_optim: a torch.optim for optimizing
the fraction model above.
:param action_space: Env's action space.
:param discount_factor: in [0, 1].
:param num_fractions: the number of fractions to use.
:param ent_coef: the coefficient for entropy loss.
: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: FullQuantileFunction,
optim: torch.optim.Optimizer,
fraction_model: FractionProposalNetwork,
fraction_optim: torch.optim.Optimizer,
action_space: gym.spaces.Discrete,
discount_factor: float = 0.99,
# TODO: used as num_quantiles in QRDQNPolicy, but num_fractions in FQFPolicy.
# Rename? Or at least explain what happens here.
num_fractions: int = 32,
ent_coef: float = 0.0,
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_fractions,
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.fraction_model = fraction_model
self.ent_coef = ent_coef
self.fraction_optim = fraction_optim
def _target_q(self, buffer: ReplayBuffer, indices: np.ndarray) -> torch.Tensor:
obs_next_batch = Batch(
obs=buffer[indices].obs_next,
info=[None] * len(indices),
) # obs_next: s_{t+n}
if self._target:
result = self(obs_next_batch)
act, fractions = result.act, result.fractions
next_dist = self(obs_next_batch, model="model_old", fractions=fractions).logits
else:
next_batch = self(obs_next_batch)
act = next_batch.act
next_dist = next_batch.logits
return next_dist[np.arange(len(act)), act, :]
# TODO: fix Liskov substitution principle violation
def forward( # type: ignore
self,
batch: ObsBatchProtocol,
state: dict | Batch | np.ndarray | None = None,
model: Literal["model", "model_old"] = "model",
fractions: Batch | None = None,
**kwargs: Any,
) -> FQFBatchProtocol:
model = getattr(self, model)
obs = batch.obs
# TODO: this is convoluted! See also other places where this is done
obs_next = obs.obs if hasattr(obs, "obs") else obs
if fractions is None:
(logits, fractions, quantiles_tau), hidden = model(
obs_next,
propose_model=self.fraction_model,
state=state,
info=batch.info,
)
else:
(logits, _, quantiles_tau), hidden = model(
obs_next,
propose_model=self.fraction_model,
fractions=fractions,
state=state,
info=batch.info,
)
weighted_logits = (fractions.taus[:, 1:] - fractions.taus[:, :-1]).unsqueeze(1) * logits
q = DQNPolicy.compute_q_value(self, weighted_logits.sum(2), getattr(obs, "mask", None))
if self.max_action_num is None: # type: ignore
# TODO: see same thing in DQNPolicy! Also reduce code duplication.
self.max_action_num = q.shape[1]
act = to_numpy(q.max(dim=1)[1])
result = Batch(
logits=logits,
act=act,
state=hidden,
fractions=fractions,
quantiles_tau=quantiles_tau,
)
return cast(FQFBatchProtocol, result)
def learn(self, batch: RolloutBatchProtocol, *args: Any, **kwargs: Any) -> TFQFTrainingStats:
if self._target and self._iter % self.freq == 0:
self.sync_weight()
weight = batch.pop("weight", 1.0)
out = self(batch)
curr_dist_orig = out.logits
taus, tau_hats = out.fractions.taus, out.fractions.tau_hats
act = batch.act
curr_dist = curr_dist_orig[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
* (tau_hats.unsqueeze(2) - (target_dist - curr_dist).detach().le(0.0).float()).abs()
)
.sum(-1)
.mean(1)
)
quantile_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
# calculate fraction loss
with torch.no_grad():
sa_quantile_hats = curr_dist_orig[np.arange(len(act)), act, :]
sa_quantiles = out.quantiles_tau[np.arange(len(act)), act, :]
# ref: https://github.com/ku2482/fqf-iqn-qrdqn.pytorch/
# blob/master/fqf_iqn_qrdqn/agent/fqf_agent.py L169
values_1 = sa_quantiles - sa_quantile_hats[:, :-1]
signs_1 = sa_quantiles > torch.cat(
[sa_quantile_hats[:, :1], sa_quantiles[:, :-1]],
dim=1,
)
values_2 = sa_quantiles - sa_quantile_hats[:, 1:]
signs_2 = sa_quantiles < torch.cat(
[sa_quantiles[:, 1:], sa_quantile_hats[:, -1:]],
dim=1,
)
gradient_of_taus = torch.where(signs_1, values_1, -values_1) + torch.where(
signs_2,
values_2,
-values_2,
)
fraction_loss = (gradient_of_taus * taus[:, 1:-1]).sum(1).mean()
# calculate entropy loss
entropy_loss = out.fractions.entropies.mean()
fraction_entropy_loss = fraction_loss - self.ent_coef * entropy_loss
self.fraction_optim.zero_grad()
fraction_entropy_loss.backward(retain_graph=True)
self.fraction_optim.step()
self.optim.zero_grad()
quantile_loss.backward()
self.optim.step()
self._iter += 1
return FQFTrainingStats( # type: ignore[return-value]
loss=quantile_loss.item() + fraction_entropy_loss.item(),
quantile_loss=quantile_loss.item(),
fraction_loss=fraction_loss.item(),
entropy_loss=entropy_loss.item(),
)