-
Notifications
You must be signed in to change notification settings - Fork 28
/
sac.py
479 lines (408 loc) · 21.9 KB
/
sac.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
"""
Soft Actor-Critic
Based on SAC implementation from https://github.com/denisyarats/pytorch_sac
License:
MIT License
Copyright (c) 2019 Denis Yarats
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
and https://github.com/rail-berkeley/softlearning
Licence:
MIT License
Copyright (c) 2018 Softlearning authors and contributors
Softlearning uses a shared copyright model: each contributor holds copyright over
their contributions to Softlearning. The project versioning records all such
contribution and copyright details.
By contributing to the Softlearning repository through pull-request, comment,
or otherwise, the contributor releases their content to the license and
copyright terms herein.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
from copy import deepcopy
import csv
from pathlib import Path
from time import time
from typing import Union
import numpy as np
import torch
from torch import nn
import torch.nn.functional as F
from torch import distributions as pyd
from torch.utils.tensorboard import SummaryWriter
from yarll.agents.agent import Agent
from yarll.agents.env_runner import EnvRunner
from yarll.memory.prealloc_memory import PreAllocMemory
from yarll.misc import summary_writer
# TODO: put this in separate file
class DeterministicPolicy:
def __init__(self, env, policy_fn):
self.env = env
self.policy_fn = policy_fn
self.initial_features = None
self.action_low = env.action_space.low
self.action_high = env.action_space.high
def choose_action(self, state, features):
res = self.policy_fn(state[None, :])[0]
return {"action": res}
def get_env_action(self, action):
return self.action_low + (action + 1.0) * 0.5 * (self.action_high - self.action_low)
def new_trajectory(self):
pass
def hard_update(source_network, target_network):
target_network.load_state_dict(source_network.state_dict())
def soft_update(source_network, target_network, tau):
for source_param, target_param in zip(source_network.parameters(), target_network.parameters()):
target_param.data.copy_(tau * source_param.data +
(1 - tau) * target_param.data)
def weight_init(m):
"""Custom weight init for Conv2D and Linear layers."""
if isinstance(m, nn.Linear):
nn.init.orthogonal_(m.weight.data)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.0)
def to_numpy(x):
return x.cpu().detach().numpy()
class SAC(Agent):
def __init__(self, env, monitor_path: Union[Path, str], **usercfg) -> None:
super().__init__(**usercfg)
self.env = env
self.monitor_path = Path(monitor_path)
self.monitor_path.mkdir(parents=True, exist_ok=True)
self.config.update(
max_steps=100000,
actor_learning_rate=3e-4,
softq_learning_rate=3e-4,
alpha_learning_rate=1e-4,
reward_scale=1.0,
n_hidden_layers=2,
n_hidden_units=256,
gamma=0.99,
batch_size=256,
tau=0.005,
init_log_alpha=0.1,
actor_update_frequency=1,
critic_target_update_frequency=2,
target_entropy=None,
logprob_epsilon=1e-6, # For numerical stability when computing log
n_train_steps=1, # Number of parameter update steps per iteration
replay_buffer_size=1e6,
replay_start_size=256, # Required number of replay buffer entries to start training
gradient_clip_value=1.0,
hidden_layer_activation="relu",
device="cpu",
normalize_inputs=False, # TODO: handle this
summaries=True,
summaries_every_steps=None, # If None, add them every time
checkpoints=True,
checkpoint_every_episodes=10,
checkpoints_max_to_keep=None,
save_model=True,
test_frequency=0,
n_test_episodes=5,
write_train_rewards=False
)
self.config.update(usercfg)
self.state_shape: list = list(env.observation_space.shape)
self.n_actions: int = env.action_space.shape[0]
self.action_low = self.env.action_space.low
self.action_high = self.env.action_space.high
self.device = torch.device(self.config["device"])
self.target_entropy = self.config["target_entropy"]
if self.target_entropy is None:
self.target_entropy = -np.prod(env.action_space.shape)
# Make networks
# action_output are the squashed actions and action_original those straight from the normal distribution
input_dim = self.state_shape[0]
self.actor_network = ActorNetwork(input_dim,
self.n_actions,
self.config["n_hidden_units"],
self.config["n_hidden_layers"]).to(self.device)
self.softq_networks = DoubleQCriticNetwork(input_dim,
self.n_actions,
self.config["n_hidden_units"],
self.config["n_hidden_layers"]).to(self.device)
self.target_softq_networks = DoubleQCriticNetwork(input_dim,
self.n_actions,
self.config["n_hidden_units"],
self.config["n_hidden_layers"]).to(self.device)
hard_update(self.softq_networks, self.target_softq_networks)
self.log_alpha = torch.tensor(self.config["init_log_alpha"]).to(self.device)
self.log_alpha.requires_grad = True
# Make train ops
# TODO: gradient_clip_value
self.actor_optimizer = torch.optim.Adam(self.actor_network.parameters(),
lr=self.config["actor_learning_rate"])
# ! TF2 code has 1 optimizer per softq network
self.softqs_optimizer = torch.optim.Adam(self.softq_networks.parameters(),
lr=self.config["softq_learning_rate"])
self.alpha_optimizer = torch.optim.Adam([self.log_alpha],
lr=self.config["alpha_learning_rate"])
self.replay_buffer = PreAllocMemory(
int(self.config["replay_buffer_size"]), self.state_shape, env.action_space.shape)
self.n_updates = 0
self.total_steps = 0
self.total_episodes = 0
if self.config["summaries"]:
self.summaries_every_steps = self.config["summaries_every_steps"] or 1
self.summary_writer = SummaryWriter(str(self.monitor_path))
summary_writer.set(self.summary_writer)
self.env_runner = EnvRunner(self.env,
self,
usercfg,
transition_preprocessor=self.config.get("transition_preprocessor", None),
summaries=self.config["summaries"],
summaries_every_episodes=self.config.get("env_summaries_every_episodes", None),
episode_rewards_file=(
self.monitor_path / "train_rewards.txt" if self.config["write_train_rewards"] else None)
)
if self.config["checkpoints"]:
self.checkpoint_directory = self.monitor_path / "checkpoints"
self.checkpoint_directory.mkdir(exist_ok=True)
if self.config["test_frequency"] > 0 and self.config["n_test_episodes"] > 0:
test_env = deepcopy(env)
unw = test_env.unwrapped
if hasattr(unw, "summaries"):
unw.summaries = False
if hasattr(unw, "log_data"):
unw.log_data = False
deterministic_policy = DeterministicPolicy(test_env, self.deterministic_actions)
self.test_env_runner = EnvRunner(test_env,
deterministic_policy,
usercfg,
summaries=False,
transition_preprocessor=self.config.get("transition_preprocessor", None),
episode_rewards_file=(
self.monitor_path / "test_rewards.txt")
)
header = [""] # (epoch) id has no name in header
header += [f"rew_{i}" for i in range(self.config["n_test_episodes"])]
header += ["rew_mean", "rew_std"]
self.test_results_file = self.monitor_path / "test_results.csv"
with open(self.test_results_file, "w") as f:
writer = csv.writer(f)
writer.writerow(header)
self.total_rewards = np.empty((self.config["n_test_episodes"],), dtype=np.float32)
def deterministic_actions(self, states: np.ndarray) -> np.ndarray:
"""Get the actions for a batch of states."""
dist = self.actor_network(torch.FloatTensor(states).to(self.device))
return to_numpy(dist.mean)
def train(self, mode: bool = True) -> None:
self.actor_network.train(mode)
self.softq_networks.train(mode)
def action(self, state: np.ndarray) -> np.ndarray:
"""Get the action for a single state."""
dist = self.actor_network(torch.FloatTensor(state[None, :]).to(self.device))
sample = dist.sample()
return to_numpy(sample)[0]
@property
def alpha(self):
return self.log_alpha.exp()
def train_critics(self, state0_batch, action_batch, reward_batch, state1_batch, terminal1_batch):
# Calculate critic targets
next_action_batch_dist = self.actor_network(state1_batch)
next_action_batch = next_action_batch_dist.rsample()
next_logprob_batch = next_action_batch_dist.log_prob(next_action_batch).sum(-1, keepdim=True)
next_qs_values = self.target_softq_networks(state1_batch, next_action_batch)
next_q_values = torch.min(*next_qs_values)
next_values = next_q_values - self.alpha.detach() * next_logprob_batch
next_values = (1.0 - terminal1_batch) * next_values
softq_targets = self.config["reward_scale"] * reward_batch + self.config["gamma"] * next_values
softq_targets = softq_targets.detach()
# Update critics
softq1_values, softq2_values = self.softq_networks(state0_batch, action_batch)
softqs_loss = F.mse_loss(softq1_values, softq_targets) + F.mse_loss(softq2_values, softq_targets)
self.softqs_optimizer.zero_grad()
softqs_loss.backward()
self.softqs_optimizer.step()
softq = torch.cat((softq1_values, softq2_values))
softq_std, softq_mean = torch.std_mean(softq)
return to_numpy(softq_mean), to_numpy(softq_std), to_numpy(softq_targets), to_numpy(softqs_loss)
def train_actor(self, state0_batch):
dist = self.actor_network(state0_batch)
action = dist.rsample()
log_prob = dist.log_prob(action).sum(-1, keepdim=True)
softqs_pred = self.softq_networks(state0_batch, action)
min_softq_pred = torch.min(*softqs_pred)
actor_losses = self.alpha.detach() * log_prob - min_softq_pred
actor_loss = actor_losses.mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
return to_numpy(actor_loss), to_numpy(log_prob.mean())
def train_alpha(self, state0_batch):
dist = self.actor_network(state0_batch)
action = dist.rsample()
log_prob = dist.log_prob(action).sum(-1, keepdim=True)
alpha_losses = -1.0 * self.alpha * (log_prob + self.target_entropy).detach()
alpha_loss = alpha_losses.mean()
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
return to_numpy(alpha_loss)
def do_test_episodes(self, step) -> None:
for i in range(self.config["n_test_episodes"]):
test_trajectory = self.test_env_runner.get_trajectory(stop_at_trajectory_end=True)
self.total_rewards[i] = np.sum(test_trajectory.rewards)
test_rewards_mean = np.mean(self.total_rewards)
test_rewards_std = np.std(self.total_rewards)
to_write = [step] + self.total_rewards.tolist() + [test_rewards_mean, test_rewards_std]
with open(self.test_results_file, "a") as f:
writer = csv.writer(f)
writer.writerow(to_write)
def learn(self):
self.train()
# Arrays to keep results from train function over different train steps in
softq_means = np.empty((self.config["n_train_steps"],), np.float32)
softq_stds = np.empty((self.config["n_train_steps"],), np.float32)
softq_losses = np.empty((self.config["n_train_steps"],), np.float32)
actor_losses = np.empty((self.config["n_train_steps"],), np.float32)
alpha_losses = np.empty((self.config["n_train_steps"],), np.float32)
action_logprob_means = np.empty((self.config["n_train_steps"],), np.float32)
total_episodes = 0
episode_start_time = time()
summary_writer.start()
for step in range(self.config["max_steps"]):
if self.config["test_frequency"] > 0 and (step % self.config["test_frequency"]) == 0 and self.config["n_test_episodes"] > 0:
self.do_test_episodes(step)
experience = self.env_runner.get_steps(1)[0]
self.total_steps += 1
self.replay_buffer.add(experience.state, experience.action, experience.reward,
experience.next_state, experience.terminal)
if self.replay_buffer.n_entries > self.config["replay_start_size"]:
for i in range(self.config["n_train_steps"]):
sample = self.replay_buffer.get_batch(self.config["batch_size"])
states0 = torch.as_tensor(sample["states0"], device=self.device)
softq_mean, softq_std, softq_targets, softq_loss = self.train_critics(
states0,
torch.as_tensor(sample["actions"], device=self.device),
torch.as_tensor(sample["rewards"], device=self.device),
torch.as_tensor(sample["states1"], device=self.device),
torch.as_tensor(sample["terminals1"], device=self.device))
if (step % self.config["actor_update_frequency"]) == 0:
actor_loss, action_logprob_mean = self.train_actor(states0)
alpha_loss = self.train_alpha(states0)
actor_losses[i] = actor_loss
alpha_losses[i] = alpha_loss
action_logprob_means[i] = action_logprob_mean
else:
print("WARNING: ACTOR NOT UPDATED")
softq_means[i] = softq_mean
softq_stds[i] = softq_std
softq_losses[i] = softq_loss
# Update the target networks
if (step % self.config["critic_target_update_frequency"]) == 0:
soft_update(self.softq_networks,
self.target_softq_networks,
self.config["tau"])
if self.config["summaries"] and (self.total_steps % self.summaries_every_steps) == 0:
summary_writer.add_scalar("model/predicted_softq_mean", np.mean(softq_means), self.total_steps)
summary_writer.add_scalar("model/predicted_softq_std", np.mean(softq_stds), self.total_steps)
summary_writer.add_scalar("model/softq_targets", softq_targets.mean(), self.total_steps)
summary_writer.add_scalar("model/softq_loss", np.mean(softq_losses), self.total_steps)
if (step % self.config["actor_update_frequency"]) == 0:
summary_writer.add_scalar("model/actor_loss", np.mean(actor_losses), self.total_steps)
summary_writer.add_scalar("model/alpha_loss", np.mean(alpha_losses), self.total_steps)
summary_writer.add_scalar("model/alpha", to_numpy(self.alpha), self.total_steps)
summary_writer.add_scalar("model/action_logprob_mean",
np.mean(action_logprob_means), self.total_steps)
self.n_updates += 1
if experience.terminal:
episode_end_time = time()
summary_writer.add_scalar("diagnostics/episode_duration_seconds",
episode_end_time - episode_start_time,
self.total_steps)
if self.config["checkpoints"] and (total_episodes % self.config["checkpoint_every_episodes"]) == 0:
torch.save(self.actor_network.state_dict(),
self.checkpoint_directory / f"actor_ep{total_episodes}.pt")
total_episodes += 1
episode_start_time = time()
summary_writer.stop()
if self.config["save_model"]:
torch.save(self.actor_network.state_dict(), self.checkpoint_directory / "actor_final.pt")
def choose_action(self, state, features):
if self.total_steps < self.config["replay_start_size"]:
action = np.random.uniform(-1.0, 1.0, self.env.action_space.shape)
else:
action = self.action(state)
return {"action": action}
def get_env_action(self, action):
"""
Converts an action from self.choose_action to an action to be given to the environment.
"""
return self.action_low + (action + 1.0) * 0.5 * (self.action_high - self.action_low)
class SquashedNormal(pyd.transformed_distribution.TransformedDistribution):
def __init__(self, loc, scale):
self.loc = loc
self.scale = scale
self.base_dist = pyd.Normal(loc, scale)
transforms = [pyd.TanhTransform(cache_size=1)]
super().__init__(self.base_dist, transforms)
@property
def mean(self):
mu = self.loc
for tr in self.transforms:
mu = tr(mu)
return mu
def mlp(input_dim, output_dim, n_hidden_units, n_hidden_layers, output_mod=None):
if n_hidden_layers == 0:
mods = [nn.Linear(input_dim, output_dim)]
else:
mods = [nn.Linear(input_dim, n_hidden_units), nn.ReLU(inplace=True)]
for i in range(n_hidden_layers - 1):
mods += [nn.Linear(n_hidden_units, n_hidden_units), nn.ReLU(inplace=True)]
mods.append(nn.Linear(n_hidden_units, output_dim))
if output_mod is not None:
mods.append(output_mod)
trunk = nn.Sequential(*mods)
return trunk
class ActorNetwork(nn.Module):
"""torch.distributions implementation of an diagonal Gaussian policy."""
def __init__(self, obs_dim, n_actions, n_hidden_units, n_hidden_layers):
super().__init__()
self.trunk = mlp(obs_dim, 2 * n_actions, n_hidden_units, n_hidden_layers)
self.apply(weight_init)
def forward(self, obs):
mu, scale = self.trunk(obs).chunk(2, dim=-1)
scale = F.softplus(scale) + 1e-5
dist = SquashedNormal(mu, scale)
return dist
class DoubleQCriticNetwork(nn.Module):
"""Critic network, employes double Q-learning."""
def __init__(self, obs_dim, action_dim, n_hidden_units, n_hidden_layers):
super().__init__()
self.Q1 = mlp(obs_dim + action_dim, 1, n_hidden_units, n_hidden_layers)
self.Q2 = mlp(obs_dim + action_dim, 1, n_hidden_units, n_hidden_layers)
self.apply(weight_init)
def forward(self, obs, action):
assert obs.size(0) == action.size(0)
obs_action = torch.cat([obs, action], dim=-1)
q1 = self.Q1(obs_action)
q2 = self.Q2(obs_action)
return q1, q2