94bfb32cc1
1. add policy.eval() in all test scripts' "watch performance" 2. remove dict return support for collector preprocess_fn 3. add `__contains__` and `pop` in batch: `key in batch`, `batch.pop(key, deft)` 4. exact n_episode for a list of n_episode limitation and save fake data in cache_buffer when self.buffer is None (#184) 5. fix tensorboard logging: h-axis stands for env step instead of gradient step; add test results into tensorboard 6. add test_returns (both GAE and nstep) 7. change the type-checking order in batch.py and converter.py in order to meet the most often case first 8. fix shape inconsistency for torch.Tensor in replay buffer 9. remove `**kwargs` in ReplayBuffer 10. remove default value in batch.split() and add merge_last argument (#185) 11. improve nstep efficiency 12. add max_batchsize in onpolicy algorithms 13. potential bugfix for subproc.wait 14. fix RecurrentActorProb 15. improve the code-coverage (from 90% to 95%) and remove the dead code 16. fix some incorrect type annotation The above improvement also increases the training FPS: on my computer, the previous version is only ~1800 FPS and after that, it can reach ~2050 (faster than v0.2.4.post1).
154 lines
5.7 KiB
Python
154 lines
5.7 KiB
Python
import torch
|
|
import numpy as np
|
|
from torch import nn
|
|
|
|
from tianshou.data import to_torch, to_torch_as
|
|
|
|
|
|
class Actor(nn.Module):
|
|
"""For advanced usage (how to customize the network), please refer to
|
|
:ref:`build_the_network`.
|
|
"""
|
|
|
|
def __init__(self, preprocess_net, action_shape, max_action=1.,
|
|
device='cpu', hidden_layer_size=128):
|
|
super().__init__()
|
|
self.preprocess = preprocess_net
|
|
self.last = nn.Linear(hidden_layer_size, np.prod(action_shape))
|
|
self._max = max_action
|
|
|
|
def forward(self, s, state=None, info={}):
|
|
"""s -> logits -> action"""
|
|
logits, h = self.preprocess(s, state)
|
|
logits = self._max * torch.tanh(self.last(logits))
|
|
return logits, h
|
|
|
|
|
|
class Critic(nn.Module):
|
|
"""For advanced usage (how to customize the network), please refer to
|
|
:ref:`build_the_network`.
|
|
"""
|
|
|
|
def __init__(self, preprocess_net, device='cpu', hidden_layer_size=128):
|
|
super().__init__()
|
|
self.device = device
|
|
self.preprocess = preprocess_net
|
|
self.last = nn.Linear(hidden_layer_size, 1)
|
|
|
|
def forward(self, s, a=None, info={}):
|
|
"""(s, a) -> logits -> Q(s, a)"""
|
|
s = to_torch(s, device=self.device, dtype=torch.float32)
|
|
s = s.flatten(1)
|
|
if a is not None:
|
|
a = to_torch(a, device=self.device, dtype=torch.float32)
|
|
a = a.flatten(1)
|
|
s = torch.cat([s, a], dim=1)
|
|
logits, h = self.preprocess(s)
|
|
logits = self.last(logits)
|
|
return logits
|
|
|
|
|
|
class ActorProb(nn.Module):
|
|
"""For advanced usage (how to customize the network), please refer to
|
|
:ref:`build_the_network`.
|
|
"""
|
|
|
|
def __init__(self, preprocess_net, action_shape, max_action=1.,
|
|
device='cpu', unbounded=False, hidden_layer_size=128):
|
|
super().__init__()
|
|
self.preprocess = preprocess_net
|
|
self.device = device
|
|
self.mu = nn.Linear(hidden_layer_size, np.prod(action_shape))
|
|
self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
|
|
self._max = max_action
|
|
self._unbounded = unbounded
|
|
|
|
def forward(self, s, state=None, info={}):
|
|
"""s -> logits -> (mu, sigma)"""
|
|
logits, h = self.preprocess(s, state)
|
|
mu = self.mu(logits)
|
|
if not self._unbounded:
|
|
mu = self._max * torch.tanh(mu)
|
|
shape = [1] * len(mu.shape)
|
|
shape[1] = -1
|
|
sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
|
|
return (mu, sigma), None
|
|
|
|
|
|
class RecurrentActorProb(nn.Module):
|
|
"""For advanced usage (how to customize the network), please refer to
|
|
:ref:`build_the_network`.
|
|
"""
|
|
|
|
def __init__(self, layer_num, state_shape, action_shape, max_action=1.,
|
|
device='cpu', unbounded=False, hidden_layer_size=128):
|
|
super().__init__()
|
|
self.device = device
|
|
self.nn = nn.LSTM(input_size=np.prod(state_shape),
|
|
hidden_size=hidden_layer_size,
|
|
num_layers=layer_num, batch_first=True)
|
|
self.mu = nn.Linear(hidden_layer_size, np.prod(action_shape))
|
|
self.sigma = nn.Parameter(torch.zeros(np.prod(action_shape), 1))
|
|
self._max = max_action
|
|
self._unbounded = unbounded
|
|
|
|
def forward(self, s, state=None, info={}):
|
|
"""Almost the same as :class:`~tianshou.utils.net.common.Recurrent`."""
|
|
s = to_torch(s, device=self.device, dtype=torch.float32)
|
|
# s [bsz, len, dim] (training) or [bsz, dim] (evaluation)
|
|
# In short, the tensor's shape in training phase is longer than which
|
|
# in evaluation phase.
|
|
if len(s.shape) == 2:
|
|
s = s.unsqueeze(-2)
|
|
self.nn.flatten_parameters()
|
|
if state is None:
|
|
s, (h, c) = self.nn(s)
|
|
else:
|
|
# we store the stack data in [bsz, len, ...] format
|
|
# but pytorch rnn needs [len, bsz, ...]
|
|
s, (h, c) = self.nn(s, (state['h'].transpose(0, 1).contiguous(),
|
|
state['c'].transpose(0, 1).contiguous()))
|
|
logits = s[:, -1]
|
|
mu = self.mu(logits)
|
|
if not self._unbounded:
|
|
mu = self._max * torch.tanh(mu)
|
|
shape = [1] * len(mu.shape)
|
|
shape[1] = -1
|
|
sigma = (self.sigma.view(shape) + torch.zeros_like(mu)).exp()
|
|
# please ensure the first dim is batch size: [bsz, len, ...]
|
|
return (mu, sigma), {'h': h.transpose(0, 1).detach(),
|
|
'c': c.transpose(0, 1).detach()}
|
|
|
|
|
|
class RecurrentCritic(nn.Module):
|
|
"""For advanced usage (how to customize the network), please refer to
|
|
:ref:`build_the_network`.
|
|
"""
|
|
|
|
def __init__(self, layer_num, state_shape,
|
|
action_shape=0, device='cpu', hidden_layer_size=128):
|
|
super().__init__()
|
|
self.state_shape = state_shape
|
|
self.action_shape = action_shape
|
|
self.device = device
|
|
self.nn = nn.LSTM(input_size=np.prod(state_shape),
|
|
hidden_size=hidden_layer_size,
|
|
num_layers=layer_num, batch_first=True)
|
|
self.fc2 = nn.Linear(hidden_layer_size + np.prod(action_shape), 1)
|
|
|
|
def forward(self, s, a=None):
|
|
"""Almost the same as :class:`~tianshou.utils.net.common.Recurrent`."""
|
|
s = to_torch(s, device=self.device, dtype=torch.float32)
|
|
# s [bsz, len, dim] (training) or [bsz, dim] (evaluation)
|
|
# In short, the tensor's shape in training phase is longer than which
|
|
# in evaluation phase.
|
|
assert len(s.shape) == 3
|
|
self.nn.flatten_parameters()
|
|
s, (h, c) = self.nn(s)
|
|
s = s[:, -1]
|
|
if a is not None:
|
|
a = to_torch_as(a, s)
|
|
s = torch.cat([s, a], dim=1)
|
|
s = self.fc2(s)
|
|
return s
|