-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconfigs.yaml
348 lines (311 loc) · 6.1 KB
/
configs.yaml
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
defaults:
logdir: null
traindir: null
evaldir: null
video_dir: null
pretrain_model_dir: null
offline_traindir: ''
offline_evaldir: ''
seed: 0
steps: 1e7
eval_every: 1e4
log_every: 1e4
reset_every: 0
#gpu_growth: True
device: 'cuda:0'
precision: 16
debug: False
expl_gifs: False
camera: corner
num_teachers: 6
eval_num: 10
is_adaptive: True
source_tasks: null
# Environment
task: 'dmc_walker_walk'
size: [64, 64]
envs: 1
action_repeat: 2
time_limit: 1000
grayscale: False
prefill: 2500
eval_noise: 0.0
clip_rewards: 'identity'
# Model
dyn_cell: 'gru'
dyn_hidden: 200
dyn_deter: 200
dyn_stoch: 50
dyn_discrete: 0
dyn_input_layers: 1
dyn_output_layers: 1
dyn_rec_depth: 1
dyn_shared: False
dyn_mean_act: 'none'
dyn_std_act: 'sigmoid2'
dyn_min_std: 0.1
dyn_temp_post: True
grad_heads: ['image', 'reward']
units: 400
reward_layers: 2
discount_layers: 3
value_layers: 3
actor_layers: 4
act: 'ELU'
cnn_depth: 32
encoder_kernels: [4, 4, 4, 4]
decoder_kernels: [5, 5, 6, 6]
decoder_thin: True
value_head: 'normal'
kl_scale: '1.0'
kl_balance: '0.8'
kl_free: '1.0'
kl_forward: False
pred_discount: False
discount_scale: 1.0
reward_scale: 1.0
weight_decay: 0.0
# Training
batch_size: 50
batch_length: 50
train_every: 5
train_steps: 1
pretrain: 100
model_lr: 3e-4
value_lr: 8e-5
actor_lr: 8e-5
opt_eps: 1e-5
grad_clip: 100
value_grad_clip: 100
actor_grad_clip: 100
dataset_size: 0
oversample_ends: False
slow_value_target: True
slow_actor_target: True
slow_target_update: 100
slow_target_fraction: 1
opt: 'adam'
# Behavior.
discount: 0.99
discount_lambda: 0.95
imag_horizon: 15
imag_gradient: 'dynamics'
imag_gradient_mix: '0.1'
imag_sample: True
actor_dist: 'trunc_normal'
actor_entropy: '1e-4'
actor_state_entropy: 0.0
actor_init_std: 1.0
actor_min_std: 0.1
actor_disc: 5
actor_temp: 0.1
actor_outscale: 0.0
expl_amount: 0.0
eval_state_mean: False
collect_dyn_sample: True
behavior_stop_grad: True
value_decay: 0.0
future_entropy: False
# Exploration
expl_behavior: 'greedy'
expl_until: 0
expl_extr_scale: 0.0
expl_intr_scale: 1.0
disag_target: 'stoch'
disag_log: True
disag_models: 10
disag_offset: 1
disag_layers: 4
disag_units: 400
disag_action_cond: False
dmlab:
# General
task: 'dmlab_rooms_watermaze'
steps: 2e8
eval_every: 1e5
log_every: 1e4
prefill: 50000
dataset_size: 2e6
pretrain: 0
# Environment
time_limit: 108000 # 30 minutes of game play.
#grayscale: True
action_repeat: 4
eval_noise: 0.0
train_every: 16
train_steps: 1
clip_rewards: 'tanh'
# Model
grad_heads: ['image', 'reward', 'discount']
dyn_cell: 'gru_layer_norm'
pred_discount: True
cnn_depth: 48
dyn_deter: 600
dyn_hidden: 600
dyn_stoch: 32
dyn_discrete: 32
reward_layers: 4
discount_layers: 4
value_layers: 4
actor_layers: 4
# Behavior
actor_dist: 'onehot'
actor_entropy: 'linear(3e-3,3e-4,2.5e6)'
expl_amount: 0.0
discount: 0.999
imag_gradient: 'both'
imag_gradient_mix: 'linear(0.1,0,2.5e6)'
# Training
discount_scale: 5.0
reward_scale: 1
weight_decay: 1e-6
model_lr: 2e-4
kl_scale: 0.1
kl_free: 0.0
actor_lr: 4e-5
value_lr: 1e-4
oversample_ends: True
atari:
# General
task: 'atari_pong'
steps: 2e8
eval_every: 1e5
log_every: 1e4
prefill: 50000
dataset_size: 2e6
pretrain: 0
# Environment
time_limit: 108000 # 30 minutes of game play.
grayscale: True
action_repeat: 4
eval_noise: 0.0
train_every: 16
train_steps: 1
clip_rewards: 'tanh'
# Model
grad_heads: ['image', 'reward', 'discount']
dyn_cell: 'gru_layer_norm'
pred_discount: True
cnn_depth: 48
dyn_deter: 600
dyn_hidden: 600
dyn_stoch: 32
dyn_discrete: 32
reward_layers: 4
discount_layers: 4
value_layers: 4
actor_layers: 4
# Behavior
actor_dist: 'onehot'
actor_entropy: 'linear(3e-3,3e-4,2.5e6)'
expl_amount: 0.0
discount: 0.999
imag_gradient: 'both'
imag_gradient_mix: 'linear(0.1,0,2.5e6)'
# Training
discount_scale: 5.0
reward_scale: 1
weight_decay: 1e-6
model_lr: 2e-4
kl_scale: 0.1
kl_free: 0.0
actor_lr: 4e-5
value_lr: 1e-4
oversample_ends: True
dmc:
# General
task: 'dmc_walker_walk'
steps: 5e5
eval_every: 3e3
log_every: 3e3
prefill: 2500
dataset_size: 0
pretrain: 100
eval_num: 1
num_teachers: 4
source_tasks: ['cheetah_run', 'hopper_stand', 'walker_walk', 'walker_run']
# Environment
time_limit: 1000
action_repeat: 2
train_every: 5
train_steps: 1
# Model
grad_heads: ['image', 'reward']
dyn_cell: 'gru_layer_norm'
pred_discount: False
cnn_depth: 32
dyn_deter: 200
dyn_stoch: 50
dyn_discrete: 0
reward_layers: 2
discount_layers: 3
value_layers: 3
actor_layers: 4
# Behavior
actor_dist: 'trunc_normal'
expl_amount: 0.0
actor_entropy: '1e-4'
discount: 0.99
imag_gradient: 'dynamics'
imag_gradient_mix: 1.0
# Training
reward_scale: 2
weight_decay: 0.0
model_lr: 3e-4
value_lr: 8e-5
actor_lr: 8e-5
opt_eps: 1e-5
kl_free: '1.0'
kl_scale: '1.0'
metaworld:
task: metaworld_door_lock
# General
steps: 3e5
eval_every: 6e3
log_every: 6e3
prefill: 5000
dataset_size: 0
pretrain: 100
eval_num: 10
num_teachers: 6
source_tasks: ['door_close', 'faucet_open', 'handle_press', 'plate_slide', 'reach_wall', 'window_close']
# Environment
time_limit: 500
action_repeat: 1
train_every: 5
train_steps: 1
# Model
grad_heads: ['image', 'reward']
dyn_cell: 'gru_layer_norm'
pred_discount: False
cnn_depth: 32
dyn_deter: 200
dyn_stoch: 50
dyn_discrete: 0
reward_layers: 2
discount_layers: 3
value_layers: 3
actor_layers: 4
# Behavior
actor_dist: 'trunc_normal'
expl_amount: 0.0
actor_entropy: '1e-4'
discount: 0.99
imag_gradient: 'dynamics'
imag_gradient_mix: 1.0
# Training
reward_scale: 2
weight_decay: 0.0
model_lr: 3e-4
value_lr: 8e-5
actor_lr: 8e-5
opt_eps: 1e-5
kl_free: '1.0'
kl_scale: '1.0'
debug:
debug: True
pretrain: 1
prefill: 1
train_steps: 1
batch_size: 10
batch_length: 20