mirror of
https://github.com/microsoft/qlib.git
synced 2026-07-02 02:21:18 +08:00
256 lines
10 KiB
Python
256 lines
10 KiB
Python
import torch
|
|
|
|
import numpy as np
|
|
from torch import nn
|
|
import torch.nn.functional as F
|
|
from typing import Dict, List, Tuple, Union, Optional
|
|
|
|
from tianshou.policy import PGPolicy
|
|
from tianshou.data import Batch, ReplayBuffer
|
|
from tianshou.data import to_torch
|
|
from numba import njit
|
|
import sys
|
|
|
|
sys.path.append("..")
|
|
from util import to_numpy, to_torch_as
|
|
|
|
|
|
def _episodic_return(
|
|
v_s_: np.ndarray, rew: np.ndarray, done: np.ndarray, gamma: float, gae_lambda: float,
|
|
) -> np.ndarray:
|
|
"""Numba speedup: 4.1s -> 0.057s."""
|
|
returns = np.roll(v_s_, 1)
|
|
m = (1.0 - done) * gamma
|
|
delta = rew + v_s_ * m - returns
|
|
m *= gae_lambda
|
|
gae = 0.0
|
|
for i in range(len(rew) - 1, -1, -1):
|
|
gae_new = delta[i] + m[i] * gae
|
|
gae = gae_new
|
|
returns[i] += gae
|
|
return returns
|
|
|
|
|
|
class PPO(PGPolicy):
|
|
""" The PPO policy with Teacher supervision"""
|
|
|
|
def __init__(
|
|
self,
|
|
actor: torch.nn.Module,
|
|
critic: torch.nn.Module,
|
|
optim: torch.optim.Optimizer,
|
|
dist_fn: torch.distributions.Distribution,
|
|
teacher=None,
|
|
discount_factor: float = 0.99,
|
|
max_grad_norm: Optional[float] = None,
|
|
eps_clip: float = 0.2,
|
|
vf_clip_para=10.0,
|
|
vf_coef: float = 0.5,
|
|
kl_coef=0.5,
|
|
kl_target=0.01,
|
|
ent_coef: float = 0.01,
|
|
sup_coef=0.1,
|
|
action_range: Optional[Tuple[float, float]] = None,
|
|
gae_lambda: float = 0.95,
|
|
dual_clip: Optional[float] = None,
|
|
value_clip: bool = True,
|
|
reward_normalization: bool = True,
|
|
**kwargs
|
|
) -> None:
|
|
super().__init__(None, None, dist_fn, discount_factor, **kwargs)
|
|
self._max_grad_norm = max_grad_norm
|
|
self._eps_clip = eps_clip
|
|
self._vf_clip_para = vf_clip_para
|
|
self._w_vf = vf_coef
|
|
self._w_ent = ent_coef
|
|
self._range = action_range
|
|
self.actor = actor
|
|
self.critic = critic
|
|
self.optim = optim
|
|
self.sup_coef = sup_coef
|
|
self.kl_target = kl_target
|
|
self.kl_coef = kl_coef
|
|
self._batch = 64
|
|
assert 0 <= gae_lambda <= 1, "GAE lambda should be in [0, 1]."
|
|
self._lambda = gae_lambda
|
|
assert dual_clip is None or dual_clip > 1, "Dual-clip PPO parameter should greater than 1."
|
|
self._dual_clip = dual_clip
|
|
self._value_clip = value_clip
|
|
self._rew_norm = reward_normalization
|
|
if not teacher is None:
|
|
self.teacher = torch.load(teacher, map_location=torch.device("cpu"))
|
|
self.teacher.to(self.actor.device)
|
|
self.teacher.actor.extractor.device = self.actor.device
|
|
else:
|
|
self.teacher = None
|
|
|
|
@staticmethod
|
|
def compute_episodic_return(
|
|
batch: Batch,
|
|
v_s_: Optional[Union[np.ndarray, torch.Tensor]] = None,
|
|
gamma: float = 0.99,
|
|
gae_lambda: float = 0.95,
|
|
rew_norm: bool = False,
|
|
) -> Batch:
|
|
"""Compute returns over given full-length episodes.
|
|
Implementation of Generalized Advantage Estimator (arXiv:1506.02438).
|
|
:param batch: a data batch which contains several full-episode data
|
|
chronologically.
|
|
:type batch: :class:`~tianshou.data.Batch`
|
|
:param v_s_: the value function of all next states :math:`V(s')`.
|
|
:type v_s_: numpy.ndarray
|
|
:param float gamma: the discount factor, should be in [0, 1], defaults
|
|
to 0.99.
|
|
:param float gae_lambda: the parameter for Generalized Advantage
|
|
Estimation, should be in [0, 1], defaults to 0.95.
|
|
:param bool rew_norm: normalize the reward to Normal(0, 1), defaults
|
|
to False.
|
|
:return: a Batch. The result will be stored in batch.returns as a numpy
|
|
array with shape (bsz, ).
|
|
"""
|
|
rew = batch.rew
|
|
v_s_ = np.zeros_like(rew) if v_s_ is None else to_numpy(v_s_.flatten())
|
|
assert not np.isnan(v_s_).any()
|
|
assert not np.isnan(rew).any()
|
|
assert not np.isnan(batch.done).any()
|
|
returns = _episodic_return(v_s_, rew, batch.done, gamma, gae_lambda)
|
|
assert not np.isnan(returns).any()
|
|
if rew_norm and not np.isclose(returns.std(), 0.0, 1e-2):
|
|
returns = (returns - returns.mean()) / returns.std()
|
|
assert not np.isnan(returns).any()
|
|
batch.returns = returns
|
|
return batch
|
|
|
|
def process_fn(self, batch: Batch, buffer: ReplayBuffer, indice: np.ndarray) -> Batch:
|
|
if self._rew_norm:
|
|
mean, std = batch.rew.mean(), batch.rew.std()
|
|
if not np.isclose(std, 0):
|
|
batch.rew = (batch.rew - mean) / std
|
|
assert not np.isnan(batch.rew).any()
|
|
if self._lambda in [0, 1]:
|
|
return self.compute_episodic_return(batch, None, gamma=self._gamma, gae_lambda=self._lambda)
|
|
else:
|
|
v_ = []
|
|
with torch.no_grad():
|
|
for b in batch.split(self._batch, shuffle=False):
|
|
v_.append(self.critic(b.obs_next))
|
|
v_ = to_numpy(torch.cat(v_, dim=0))
|
|
assert not np.isnan(v_).any()
|
|
return self.compute_episodic_return(batch, v_, gamma=self._gamma, gae_lambda=self._lambda)
|
|
|
|
def forward(self, batch: Batch, state: Optional[Union[dict, Batch, np.ndarray]] = None, **kwargs) -> Batch:
|
|
"""Compute action over the given batch data."""
|
|
logits, h = self.actor(batch.obs, state=state, info=batch.info)
|
|
if isinstance(logits, tuple):
|
|
dist = self.dist_fn(*logits)
|
|
else:
|
|
dist = self.dist_fn(logits)
|
|
if self.training:
|
|
try:
|
|
act = dist.sample()
|
|
except:
|
|
print(logits)
|
|
act = dist.sample()
|
|
else:
|
|
act = torch.argmax(logits, dim=1)
|
|
if self._range:
|
|
act = act.clamp(self._range[0], self._range[1])
|
|
return Batch(logits=logits, act=act, state=h, dist=dist)
|
|
|
|
def learn(self, batch: Batch, batch_size: int, repeat: int, **kwargs) -> Dict[str, List[float]]:
|
|
self._batch = batch_size
|
|
losses, clip_losses, vf_losses, ent_losses, kl_losses = [], [], [], [], []
|
|
if self.teacher is not None:
|
|
supervision_losses = []
|
|
v = []
|
|
old_log_prob = []
|
|
feature = []
|
|
old_logits = []
|
|
with torch.no_grad():
|
|
for b in batch.split(batch_size, shuffle=False):
|
|
v.append(self.critic(b.obs))
|
|
b_ = self(b)
|
|
dist = b_.dist
|
|
logits = b_.logits
|
|
old_log_prob.append(dist.log_prob(to_torch_as(b.act, v[0])))
|
|
old_logits.append(logits)
|
|
if not self.teacher is None:
|
|
with torch.no_grad():
|
|
for b in batch.split(batch_size, shuffle=False):
|
|
self.teacher(b)
|
|
feature.append(self.teacher.actor.feature)
|
|
batch.old_feature = torch.cat(feature, dim=0)
|
|
batch.old_logits = torch.cat(old_logits, dim=0)
|
|
batch.v = torch.cat(v, dim=0) # old value
|
|
batch.act = to_torch_as(batch.act, v[0])
|
|
batch.logp_old = torch.cat(old_log_prob, dim=0)
|
|
batch.returns = to_torch_as(batch.returns, v[0]).reshape(batch.v.shape)
|
|
if self._rew_norm:
|
|
mean, std = batch.returns.mean(), batch.returns.std()
|
|
if not np.isclose(std.item(), 0):
|
|
batch.returns = (batch.returns - mean) / std
|
|
batch.adv = batch.returns - batch.v
|
|
if self._rew_norm:
|
|
mean, std = batch.adv.mean(), batch.adv.std()
|
|
if not np.isclose(std.item(), 0):
|
|
batch.adv = (batch.adv - mean) / std
|
|
for _ in range(repeat):
|
|
for b in batch.split(batch_size):
|
|
dist = self(b).dist
|
|
value = self.critic(b.obs)
|
|
if not self.teacher is None:
|
|
feature = self.actor.feature
|
|
# print(feature.pow(2).mean())
|
|
ratio = (dist.log_prob(b.act) - b.logp_old).exp().float()
|
|
surr1 = ratio * b.adv
|
|
surr2 = ratio.clamp(1.0 - self._eps_clip, 1.0 + self._eps_clip) * b.adv
|
|
if self._dual_clip:
|
|
clip_loss = -torch.max(torch.min(surr1, surr2), self._dual_clip * b.adv).mean()
|
|
else:
|
|
clip_loss = -torch.min(surr1, surr2).mean()
|
|
clip_losses.append(clip_loss.item())
|
|
if self._value_clip:
|
|
v_clip = b.v + (value - b.v).clamp(-self._vf_clip_para, self._vf_clip_para)
|
|
vf1 = (b.returns - value).pow(2)
|
|
vf2 = (b.returns - v_clip).pow(2)
|
|
vf_loss = torch.max(vf1, vf2).mean()
|
|
else:
|
|
vf_loss = (b.returns - value).pow(2).mean()
|
|
if not self.teacher is None:
|
|
supervision_loss = (b.old_feature - feature).pow(2).mean()
|
|
supervision_losses.append(supervision_loss.item())
|
|
kl = torch.distributions.kl.kl_divergence(self.dist_fn(b.old_logits), dist)
|
|
kl_loss = kl.mean()
|
|
kl_losses.append(kl_loss.item())
|
|
vf_losses.append(vf_loss.item())
|
|
e_loss = dist.entropy().mean()
|
|
ent_losses.append(e_loss.item())
|
|
loss = clip_loss + self._w_vf * vf_loss - self._w_ent * e_loss + self.kl_coef * kl_loss
|
|
if self.teacher is not None:
|
|
loss += self.sup_coef * supervision_loss
|
|
losses.append(loss.item())
|
|
self.optim.zero_grad()
|
|
loss.backward()
|
|
nn.utils.clip_grad_norm_(
|
|
list(self.actor.parameters()) + list(self.critic.parameters()), self._max_grad_norm,
|
|
)
|
|
self.optim.step()
|
|
cur_kl = np.mean(kl_losses)
|
|
if cur_kl > 2.0 * self.kl_target:
|
|
self.kl_coef *= 1.5
|
|
elif cur_kl < 0.5 * self.kl_target:
|
|
self.kl_coef *= 0.5
|
|
res = {
|
|
"loss/total_loss": losses,
|
|
"loss/policy": clip_losses,
|
|
"loss/vf": vf_losses,
|
|
"loss/entropy": ent_losses,
|
|
"loss/kl": kl_losses,
|
|
}
|
|
if not self.teacher is None:
|
|
res["loss/supervision"] = supervision_losses
|
|
return res
|
|
|
|
|
|
Student_new = PPO
|