forked from rraileanu/auto-drac
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
319 lines (271 loc) · 11.4 KB
/
train.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
import copy
import os
import time
from collections import deque
from tqdm import trange
import sys
import numpy as np
import torch
from ucb_rl2_meta import algo, utils
from ucb_rl2_meta.model import Policy, AugCNN, ModelBasedPolicy, PlanningPolicy
from ucb_rl2_meta.storage import RolloutStorage, BiggerRolloutStorage
from test import evaluate
from baselines import logger
from procgen import ProcgenEnv
from baselines.common.vec_env import (
VecExtractDictObs,
VecMonitor,
VecNormalize
)
from ucb_rl2_meta.envs import VecPyTorchProcgen, TransposeImageProcgen
from ucb_rl2_meta.arguments import parser
import data_augs
aug_to_func = {
'crop': data_augs.Crop,
'random-conv': data_augs.RandomConv,
'grayscale': data_augs.Grayscale,
'flip': data_augs.Flip,
'rotate': data_augs.Rotate,
'cutout': data_augs.Cutout,
'cutout-color': data_augs.CutoutColor,
'color-jitter': data_augs.ColorJitter,
}
modelbased = True
def train(args):
args.cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
log_dir = os.path.expanduser(args.log_dir)
utils.cleanup_log_dir(log_dir)
torch.set_num_threads(1)
device = torch.device("cuda:0" if args.cuda else "cpu")
log_file = '-{}-{}-reproduce-s{}'.format(args.run_name, args.env_name, args.seed)
logger.configure(dir=args.log_dir, format_strs=['csv', 'stdout'], log_suffix=log_file)
venv = ProcgenEnv(num_envs=args.num_processes, env_name=args.env_name, \
num_levels=args.num_levels, start_level=args.start_level, \
distribution_mode=args.distribution_mode)
venv = VecExtractDictObs(venv, "rgb")
venv = VecMonitor(venv=venv, filename=None, keep_buf=100)
venv = VecNormalize(venv=venv, ob=False)
envs = VecPyTorchProcgen(venv, device)
obs_shape = envs.observation_space.shape
actor_critic = Policy(
obs_shape,
envs.action_space.n,
base_kwargs={'recurrent': False, 'hidden_size': args.hidden_size})
actor_critic.to(device)
if modelbased:
rollouts = BiggerRolloutStorage(args.num_steps, args.num_processes,
envs.observation_space.shape, envs.action_space,
actor_critic.recurrent_hidden_state_size,
aug_type=args.aug_type, split_ratio=args.split_ratio)
else:
rollouts = RolloutStorage(args.num_steps, args.num_processes,
envs.observation_space.shape, envs.action_space,
actor_critic.recurrent_hidden_state_size,
aug_type=args.aug_type, split_ratio=args.split_ratio)
batch_size = int(args.num_processes * args.num_steps / args.num_mini_batch)
if args.use_ucb:
aug_id = data_augs.Identity
aug_list = [aug_to_func[t](batch_size=batch_size)
for t in list(aug_to_func.keys())]
agent = algo.UCBDrAC(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
aug_list=aug_list,
aug_id=aug_id,
aug_coef=args.aug_coef,
num_aug_types=len(list(aug_to_func.keys())),
ucb_exploration_coef=args.ucb_exploration_coef,
ucb_window_length=args.ucb_window_length)
elif args.use_meta_learning:
aug_id = data_augs.Identity
aug_list = [aug_to_func[t](batch_size=batch_size) \
for t in list(aug_to_func.keys())]
aug_model = AugCNN()
aug_model.to(device)
agent = algo.MetaDrAC(
actor_critic,
aug_model,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
meta_grad_clip=args.meta_grad_clip,
meta_num_train_steps=args.meta_num_train_steps,
meta_num_test_steps=args.meta_num_test_steps,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
aug_id=aug_id,
aug_coef=args.aug_coef)
elif args.use_rl2:
aug_id = data_augs.Identity
aug_list = [aug_to_func[t](batch_size=batch_size)
for t in list(aug_to_func.keys())]
rl2_obs_shape = [envs.action_space.n + 1]
rl2_learner = Policy(
rl2_obs_shape,
len(list(aug_to_func.keys())),
base_kwargs={'recurrent': True, 'hidden_size': args.rl2_hidden_size})
rl2_learner.to(device)
agent = algo.RL2DrAC(
actor_critic,
rl2_learner,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
args.rl2_entropy_coef,
lr=args.lr,
eps=args.eps,
rl2_lr=args.rl2_lr,
rl2_eps=args.rl2_eps,
max_grad_norm=args.max_grad_norm,
aug_list=aug_list,
aug_id=aug_id,
aug_coef=args.aug_coef,
num_aug_types=len(list(aug_to_func.keys())),
recurrent_hidden_size=args.rl2_hidden_size,
num_actions=envs.action_space.n,
device=device)
elif False: # Regular Drac
aug_id = data_augs.Identity
aug_func = aug_to_func[args.aug_type](batch_size=batch_size)
agent = algo.DrAC(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
aug_id=aug_id,
aug_func=aug_func,
aug_coef=args.aug_coef,
env_name=args.env_name)
elif False: # Model Free Planning Drac
aug_id = data_augs.Identity
aug_func = aug_to_func[args.aug_type](batch_size=batch_size)
actor_critic = PlanningPolicy(
obs_shape,
envs.action_space.n,
base_kwargs={'recurrent': False, 'hidden_size': args.hidden_size})
actor_critic.to(device)
agent = algo.DrAC(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
aug_id=aug_id,
aug_func=aug_func,
aug_coef=args.aug_coef,
env_name=args.env_name)
else: # Model based Drac
aug_id = data_augs.Identity
aug_func = aug_to_func[args.aug_type](batch_size=batch_size)
actor_critic = ModelBasedPolicy(
obs_shape,
envs.action_space.n,
base_kwargs={'recurrent': False, 'hidden_size': args.hidden_size})
actor_critic.to(device)
agent = algo.ConvDrAC(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
aug_id=aug_id,
aug_func=aug_func,
aug_coef=args.aug_coef,
env_name=args.env_name)
obs = envs.reset()
rollouts.obs[0].copy_(obs)
if modelbased:
rollouts.next_obs[0].copy_(obs) # TODO: is this right?
rollouts.to(device)
episode_rewards = deque(maxlen=10)
num_updates = int(
args.num_env_steps) // args.num_steps // args.num_processes
for j in trange(num_updates):
actor_critic.train()
for step in range(args.num_steps):
# Sample actions
with torch.no_grad():
obs_id = aug_id(rollouts.obs[step])
value, action, action_log_prob, recurrent_hidden_states = actor_critic.act(
obs_id, rollouts.recurrent_hidden_states[step],
rollouts.masks[step])
# Obser reward and next obs
obs, reward, done, infos = envs.step(action)
for info in infos:
if 'episode' in info.keys():
episode_rewards.append(info['episode']['r'])
# If done then clean the history of observations.
masks = torch.FloatTensor(
[[0.0] if done_ else [1.0] for done_ in done])
bad_masks = torch.FloatTensor(
[[0.0] if 'bad_transition' in info.keys() else [1.0]
for info in infos])
rollouts.insert(obs, recurrent_hidden_states, action,
action_log_prob, value, reward, masks, bad_masks)
with torch.no_grad():
obs_id = aug_id(rollouts.obs[-1])
next_value = actor_critic.get_value(
obs_id, rollouts.recurrent_hidden_states[-1],
rollouts.masks[-1]).detach()
rollouts.compute_returns(next_value, args.gamma, args.gae_lambda)
if args.use_ucb and j > 0:
agent.update_ucb_values(rollouts)
if isinstance(agent, algo.ConvDrAC):
value_loss, action_loss, dist_entropy, transition_model_loss, reward_model_loss = agent.update(rollouts)
else:
value_loss, action_loss, dist_entropy = agent.update(rollouts)
rollouts.after_update()
# save for every interval-th episode or for the last epoch
total_num_steps = (j + 1) * args.num_processes * args.num_steps
if j % args.log_interval == 0 and len(episode_rewards) > 1:
total_num_steps = (j + 1) * args.num_processes * args.num_steps
print("\nUpdate {}, step {} \n Last {} training episodes: mean/median reward {:.1f}/{:.1f}"
.format(j, total_num_steps,
len(episode_rewards), np.mean(episode_rewards),
np.median(episode_rewards), dist_entropy, value_loss,
action_loss))
logger.logkv("train/nupdates", j)
logger.logkv("train/total_num_steps", total_num_steps)
logger.logkv("losses/dist_entropy", dist_entropy)
logger.logkv("losses/value_loss", value_loss)
logger.logkv("losses/action_loss", action_loss)
if isinstance(agent, algo.ConvDrAC):
logger.logkv("losses/transition_model_loss", transition_model_loss)
logger.logkv("losses/reward_model_loss", reward_model_loss)
logger.logkv("train/mean_episode_reward", np.mean(episode_rewards))
logger.logkv("train/median_episode_reward", np.median(episode_rewards))
### Eval on the Full Distribution of Levels ###
eval_episode_rewards = evaluate(args, actor_critic, device, aug_id=aug_id)
logger.logkv("test/mean_episode_reward", np.mean(eval_episode_rewards))
logger.logkv("test/median_episode_reward", np.median(eval_episode_rewards))
logger.dumpkvs()
if __name__ == "__main__":
args = parser.parse_args()
train(args)