forked from MichalBortkiewicz/JaxGCRL
-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
316 lines (291 loc) · 13.4 KB
/
utils.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
import argparse
from collections import namedtuple
from datetime import datetime
from matplotlib import pyplot as plt
import wandb
from envs.ant import Ant
from envs.debug_env import Debug
from envs.half_cheetah import Halfcheetah
from envs.reacher import Reacher
from envs.pusher import Pusher, PusherReacher
from envs.ant_ball import AntBall
from envs.ant_maze import AntMaze
from envs.humanoid import Humanoid
from envs.ant_push import AntPush
Config = namedtuple(
"Config",
"debug discount obs_dim goal_start_idx goal_end_idx unroll_length episode_length repr_dim random_goals disable_entropy_actor use_traj_idx_wrapper",
)
def create_parser():
parser = argparse.ArgumentParser(description="Training script arguments")
parser.add_argument("--exp_name", type=str, default="test", help="Name of the wandb experiment")
parser.add_argument("--group_name", type=str, default="test", help="Name of the wandb group of experiment")
parser.add_argument("--project_name", type=str, default="crl", help="Name of the wandb project of experiment")
parser.add_argument("--num_timesteps", type=int, default=1000000, help="Number of training timesteps")
parser.add_argument("--max_replay_size", type=int, default=10000, help="Maximum size of replay buffer")
parser.add_argument("--min_replay_size", type=int, default=8192, help="Minimum size of replay buffer")
parser.add_argument("--num_evals", type=int, default=50, help="Number of evaluations")
parser.add_argument("--episode_length", type=int, default=50, help="Maximum length of each episode")
parser.add_argument("--action_repeat", type=int, default=2, help="Number of times to repeat each action")
parser.add_argument("--discounting", type=float, default=0.997, help="Discounting factor for rewards")
parser.add_argument("--num_envs", type=int, default=256, help="Number of environments")
parser.add_argument("--batch_size", type=int, default=512, help="Batch size for training")
parser.add_argument("--seed", type=int, default=0, help="Seed for reproducibility")
parser.add_argument("--unroll_length", type=int, default=50, help="Length of the env unroll")
parser.add_argument("--multiplier_num_sgd_steps", type=int, default=1, help="Multiplier of total number of gradient steps resulting from other args.",)
parser.add_argument("--env_name", type=str, default="reacher", help="Name of the environment to train on")
parser.add_argument("--normalize_observations", default=False, action="store_true", help="Whether to normalize observations")
parser.add_argument("--log_wandb", default=False, action="store_true", help="Whether to log to wandb")
parser.add_argument('--policy_lr', type=float, default=6e-4)
parser.add_argument('--alpha_lr', type=float, default=3e-4)
parser.add_argument('--critic_lr', type=float, default=3e-4)
parser.add_argument('--contrastive_loss_fn', type=str, default='symmetric_infonce')
parser.add_argument('--energy_fn', type=str, default='l2')
parser.add_argument('--backend', type=str, default=None)
parser.add_argument('--no_resubs', default=False, action='store_true', help="Not use resubstitution (diagonal) for logsumexp in contrastive cross entropy")
parser.add_argument('--use_ln', default=False, action='store_true', help="Whether to use layer normalization for preactivations in hidden layers")
parser.add_argument('--use_c_target', default=False, action='store_true', help="Use learnable c_target param in contrastive loss")
parser.add_argument('--logsumexp_penalty', type=float, default=0.0)
parser.add_argument('--l2_penalty', type=float, default=0.0)
parser.add_argument('--exploration_coef', type=float, default=0.0)
parser.add_argument('--random_goals', type=float, default=0.0, help="Propotion of random goals to use in the actor loss")
parser.add_argument('--disable_entropy_actor', default=False, action="store_true", help="Whether to disable entropy in actor")
parser.add_argument('--use_traj_idx_wrapper', default=False, action="store_true", help="Whether to use debug wrapper with info about envs, seeds and trajectories")
parser.add_argument('--eval_env', type=str, default=None, help="Whether to use separate environment for evaluation")
parser.add_argument("--h_dim", type=int, default=256, help="Width of hidden layers")
parser.add_argument("--n_hidden", type=int, default=2, help="Number of hidden layers")
parser.add_argument('--repr_dim', type=int, default=64, help="Dimension of the representation")
return parser
def create_env(args: argparse.Namespace) -> object:
env_name = args.env_name
if env_name == "reacher":
env = Reacher(backend=args.backend or "generalized")
elif env_name == "ant":
env = Ant(
backend=args.backend or "spring",
exclude_current_positions_from_observation=False,
terminate_when_unhealthy=True,
)
elif env_name == "ant_ball":
env = AntBall(
backend=args.backend or "spring",
exclude_current_positions_from_observation=False,
terminate_when_unhealthy=True,
)
elif env_name == "ant_push":
# This is stable only in mjx backend
assert args.backend == "mjx"
env = AntPush(
backend=args.backend,
exclude_current_positions_from_observation=False,
terminate_when_unhealthy=True,
)
elif "maze" in env_name:
# Possible env_name = {'ant_u_maze', 'ant_big_maze', 'ant_hardest_maze'}
env = AntMaze(
backend=args.backend or "spring",
exclude_current_positions_from_observation=False,
terminate_when_unhealthy=True,
maze_layout_name=env_name[4:]
)
elif env_name == "cheetah":
env = Halfcheetah(
backend="mjx",
exclude_current_positions_from_observation=False,
)
elif env_name == "debug":
env = Debug(backend=args.backend or "spring")
elif env_name == "pusher_easy":
env=Pusher(backend=args.backend or "generalized", kind="easy")
elif env_name == "pusher_hard":
env=Pusher(backend=args.backend or "generalized", kind="hard")
elif env_name == "pusher_reacher":
env=PusherReacher(backend=args.backend or "generalized")
elif env_name == "humanoid":
env=Humanoid(backend=args.backend)
else:
raise ValueError(f"Unknown environment: {env_name}")
return env
def create_eval_env(args: argparse.Namespace) -> object:
if not args.eval_env:
return None
eval_arg = argparse.Namespace(**vars(args))
eval_arg.env_name = args.eval_env
return create_env(eval_arg)
def get_env_config(args: argparse.Namespace):
if args.env_name == "debug":
config = Config(
debug=True,
discount=args.discounting,
obs_dim=3,
goal_start_idx=1,
goal_end_idx=3,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "reacher":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=10,
goal_start_idx=4,
goal_end_idx=7,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "cheetah":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=18,
goal_start_idx=0,
goal_end_idx=1,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "pusher_easy" or args.env_name == "pusher_hard":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=20,
goal_start_idx=10,
goal_end_idx=13,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "pusher_reacher":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=17,
goal_start_idx=14,
goal_end_idx=17,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "ant" or 'maze' in args.env_name:
config = Config(
debug=False,
discount=args.discounting,
obs_dim=29,
goal_start_idx=0,
goal_end_idx=2,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "ant_push":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=31,
goal_start_idx=0,
goal_end_idx=2,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "ant_ball":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=31,
goal_start_idx=-4,
goal_end_idx=-2,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
elif args.env_name == "humanoid":
config = Config(
debug=False,
discount=args.discounting,
obs_dim=268,
goal_start_idx=0,
goal_end_idx=3,
unroll_length=args.unroll_length,
episode_length=args.episode_length,
repr_dim=args.repr_dim,
random_goals=args.random_goals,
disable_entropy_actor=args.disable_entropy_actor,
use_traj_idx_wrapper=args.use_traj_idx_wrapper
)
else:
raise ValueError(f"Unknown environment: {args.env_name}")
return config
class MetricsRecorder:
def __init__(self, num_timesteps):
self.x_data = []
self.y_data = {}
self.y_data_err = {}
self.times = [datetime.now()]
self.max_x, self.min_x = num_timesteps * 1.1, 0
def record(self, num_steps, metrics):
self.times.append(datetime.now())
self.x_data.append(num_steps)
for key, value in metrics.items():
if key not in self.y_data:
self.y_data[key] = []
self.y_data_err[key] = []
self.y_data[key].append(value)
self.y_data_err[key].append(metrics.get(f"{key}_std", 0))
def log_wandb(self):
data_to_log = {}
for key, value in self.y_data.items():
data_to_log[key] = value[-1]
data_to_log["step"] = self.x_data[-1]
wandb.log(data_to_log, step=self.x_data[-1])
def plot_progress(self):
num_plots = len(self.y_data)
num_rows = (num_plots + 1) // 2 # Calculate number of rows needed for 2 columns
fig, axs = plt.subplots(num_rows, 2, figsize=(15, 5 * num_rows))
for idx, (key, y_values) in enumerate(self.y_data.items()):
row = idx // 2
col = idx % 2
axs[row, col].set_xlim(self.min_x, self.max_x)
axs[row, col].set_xlabel("# environment steps")
axs[row, col].set_ylabel(key)
axs[row, col].errorbar(self.x_data, y_values, yerr=self.y_data_err[key])
axs[row, col].set_title(f"{key}: {y_values[-1]:.3f}")
# Hide any empty subplots
for idx in range(num_plots, num_rows * 2):
row = idx // 2
col = idx % 2
axs[row, col].axis("off")
plt.tight_layout()
plt.show()
def print_progress(self):
for idx, (key, y_values) in enumerate(self.y_data.items()):
print(f"step: {self.x_data[-1]}, {key}: {y_values[-1]:.3f} +/- {self.y_data_err[key][-1]:.3f}")
def print_times(self):
print(f"time to jit: {self.times[1] - self.times[0]}")
print(f"time to train: {self.times[-1] - self.times[1]}")