-
Notifications
You must be signed in to change notification settings - Fork 1
/
philipp_runner.py
366 lines (317 loc) · 13.7 KB
/
philipp_runner.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
# Continuous Environment Meta Reinforcement Learning (CEMRL)
import os
from collections import OrderedDict
from pathlib import Path
import numpy as np
import click
import json
import torch
import gym
from cemrl_edited.exploration_agent import construct_exploration_agent
gym.logger.set_level(40)
from cemrl_edited.policy_networks import SingleSAC, MultipleSAC
from rlkit.envs.wrappers import NormalizedBoxEnv, CameraWrapper
from rlkit.launchers.launcher_util import setup_logger
import rlkit.torch.pytorch_util as ptu
from configs.default import default_config
from cemrl_edited.encoder_decoder_networks import PriorPz, Encoder, DecoderMDP
from cemrl_edited.sac import PolicyTrainer
from cemrl_edited.stacked_replay_buffer import StackedReplayBuffer
from cemrl_edited.reconstruction_trainer import ReconstructionTrainer, NoOpReconstructionTrainer
from cemrl_edited.rollout_worker import RolloutCoordinator
from cemrl_edited.agent import CEMRLAgent, ScriptedPolicyAgent
from cemrl_edited.cemrl_algorithm import CEMRLAlgorithm
from meta_rand_envs.wrappers import ENVS
def setup_environment(variant):
# optional GPU mode
ptu.set_gpu_mode(variant['util_params']['use_gpu'], variant['util_params']['gpu_id'])
torch.set_num_threads(1)
if variant['algo_params']['use_fixed_seeding']:
torch.manual_seed(variant['algo_params']['seed'])
np.random.seed(variant['algo_params']['seed'])
# create logging directory
encoding_save_epochs = [0, 25, 50, 75, 100, 125, 150, 175, 200, 225, 250, 275, 300, 325, 350, 375, 400, 425, 450,
475, 500, 600, 750, 1000, 1250, 1500, 1750, 2000, 3000, 5000, 10000, 15000, 20000]
experiment_log_dir = setup_logger(variant['config_name'], variant=variant, exp_id=None,
base_log_dir=variant['util_params']['base_log_dir'], snapshot_mode='specific',
snapshot_gap=variant['algo_params']['snapshot_gap'],
snapshot_points=encoding_save_epochs)
# create temp folder
if not os.path.exists(variant['util_params']['temp_dir']):
os.makedirs(variant['util_params']['temp_dir'])
# debugging triggers a lot of printing and logs to a debug directory
DEBUG = variant['util_params']['debug']
PLOT = variant['util_params']['plot']
os.environ['DEBUG'] = str(int(DEBUG))
os.environ['PLOT'] = str(int(PLOT))
# create multi-task environment and sample tasks
env = ENVS[variant['env_name']](**variant['env_params'])
if variant['env_params']['use_normalized_env']:
env = NormalizedBoxEnv(env)
if variant['train_or_showcase'] == 'showcase':
env = CameraWrapper(env)
return env, experiment_log_dir
def initialize_networks(variant, env, experiment_log_dir):
obs_dim = int(np.prod(env.observation_space.shape))
action_dim = int(np.prod(env.action_space.shape))
reward_dim = 1
tasks = list(range(len(env.tasks)))
train_tasks = list(range(len(env.train_tasks)))
test_tasks = tasks[-variant['env_params']['n_eval_tasks']:]
# instantiate networks
net_complex_enc_dec = variant['reconstruction_params']['net_complex_enc_dec']
latent_dim = variant['algo_params']['latent_size']
time_steps = variant['algo_params']['time_steps']
num_classes = variant['reconstruction_params']['num_classes']
# set parameters if not given
if variant['algo_params']['decoder_time_window'] is None:
variant['algo_params']['decoder_time_window'] = [-time_steps, 0]
# encoder used: single transitions or trajectories
encoder = Encoder(
obs_dim,
action_dim,
reward_dim,
net_complex_enc_dec,
variant['algo_params']['encoder_type'],
variant['algo_params']['encoder_exclude_padding'],
latent_dim,
variant['algo_params']['batch_size_reconstruction'],
num_classes,
variant['reconstruction_params']['state_preprocessing_dim'],
variant['reconstruction_params']['simplified_state_preprocessor'],
time_steps,
variant['algo_params']['encoder_merge_mode'],
relevant_input_indices=variant['algo_params']['encoder_omit_input']
)
decoder = DecoderMDP(
action_dim,
obs_dim,
reward_dim,
latent_dim,
net_complex_enc_dec,
variant['env_params']['state_reconstruction_clip'],
variant['env_params']['use_state_decoder'],
encoder.state_preprocessor,
variant['reconstruction_params']['use_next_state_for_reward_decoder'],
)
prior_pz = PriorPz(num_classes, latent_dim)
if variant['algo_params']['policy_mode'] == 'sac_single':
policy_networks = SingleSAC(
obs_dim,
latent_dim,
action_dim,
variant['algo_params']['sac_layer_size']
)
elif variant['algo_params']['policy_mode'] == 'sac_multiple':
policy_networks = MultipleSAC(
obs_dim,
latent_dim,
action_dim,
variant['algo_params']['sac_layer_size'],
variant['algo_params']['num_policy_nets']
)
else:
raise ValueError(f"{variant['algo_params']['policy_mode']} is not a valid policy_mode")
networks = {
'encoder': encoder,
'prior_pz': prior_pz,
'decoder': decoder,
**policy_networks.get_networks()
}
combined_trajectories = variant['algo_params']['num_trajectories_per_task'] \
+ variant['algo_params']['num_exploration_trajectories_per_task']
replay_buffer = StackedReplayBuffer(
variant['algo_params']['max_replay_buffer_size'],
time_steps,
variant['algo_params']['decoder_time_window'],
combined_trajectories * variant['algo_params']['max_path_length'],
obs_dim,
action_dim,
latent_dim,
variant['algo_params']['permute_samples'],
variant['algo_params']['sampling_mode']
)
# Agent
agent_class = ScriptedPolicyAgent if variant['env_params']['scripted_policy'] else CEMRLAgent
agent = agent_class(
encoder,
prior_pz,
policy_networks
)
if variant['algo_params']['exploration_agent'] is not None:
exploration_agent = construct_exploration_agent(
variant['algo_params']['exploration_agent'],
policy_networks,
replay_buffer,
variant['env_name'],
variant['env_params'],
experiment_log_dir if variant['path_to_weights'] is None else variant['path_to_weights'],
variant['algo_params']['max_path_length'],
variant['algo_params']['exploration_pretraining_steps'],
encoder.state_preprocessor,
variant['algo_params']['exploration_ensemble_agents'],
variant['showcase_itr']
)
else:
exploration_agent = agent
# Rollout Coordinator
rollout_coordinator = RolloutCoordinator(
env,
variant['env_name'],
variant['env_params'],
variant['train_or_showcase'],
agent,
exploration_agent,
replay_buffer,
time_steps,
variant['algo_params']['max_path_length'],
variant['algo_params']['permute_samples'],
variant['util_params']['use_multiprocessing'],
variant['algo_params']['use_data_normalization'],
variant['algo_params']['use_sac_data_normalization'],
variant['util_params']['num_workers'],
variant['util_params']['gpu_id'],
variant['env_params']['scripted_policy']
)
# ReconstructionTrainer
reconstruction_trainer = ReconstructionTrainer(
encoder,
decoder,
prior_pz,
replay_buffer,
variant['algo_params']['batch_size_reconstruction'],
variant['algo_params']['batch_size_validation'],
num_classes,
latent_dim,
time_steps,
variant['reconstruction_params']['lr_decoder'],
variant['reconstruction_params']['lr_encoder'],
variant['reconstruction_params']['alpha_kl_z'],
variant['reconstruction_params']['beta_kl_y'],
variant['reconstruction_params']['alpha_kl_z_query'],
variant['reconstruction_params']['beta_kl_y_query'],
variant['reconstruction_params']['use_state_diff'],
variant['reconstruction_params']['component_constraint_learning'],
variant['env_params']['state_reconstruction_clip'],
variant['env_params']['use_state_decoder'],
variant['algo_params']['use_data_normalization'],
variant['reconstruction_params']['train_val_percent'],
variant['reconstruction_params']['eval_interval'],
variant['reconstruction_params']['early_stopping_threshold'],
experiment_log_dir,
variant['util_params']['temp_dir'],
variant['reconstruction_params']['prior_mode'],
variant['reconstruction_params']['prior_sigma'],
variant['algo_params']['data_usage_reconstruction'],
variant['reconstruction_params']['reconstruct_all_timesteps']
)
if variant['algo_params']['encoder_type'] == 'NoEncoder':
# debug case: completely omit any encoding and only do SAC training
reconstruction_trainer = NoOpReconstructionTrainer()
# PolicyTrainer
policy_trainer = PolicyTrainer(
policy_networks,
replay_buffer,
encoder,
variant['algo_params']['batch_size_policy'],
action_dim,
variant['algo_params']['data_usage_sac'],
variant['algo_params']['use_data_normalization'],
variant['algo_params']['use_sac_data_normalization'],
variant['algo_params']['sac_uses_exploration_data'],
policy_lr=variant['algo_params']['lr_policy'],
qf_lr=variant['algo_params']['lr_qf'],
use_parametrized_alpha=variant['algo_params']['use_parametrized_alpha'],
target_entropy_factor=variant['algo_params']['target_entropy_factor'],
alpha=variant['algo_params']['sac_alpha']
)
algorithm = CEMRLAlgorithm(
replay_buffer,
rollout_coordinator,
reconstruction_trainer,
policy_trainer,
agent,
exploration_agent,
networks,
train_tasks,
test_tasks,
variant['algo_params']['num_train_epochs'],
variant['algo_params']['exploration_pretraining_epochs'],
variant['algo_params']['exploration_pretraining_epoch_steps'],
variant['algo_params']['exploration_epoch_steps'],
variant['algo_params']['num_reconstruction_steps'],
variant['algo_params']['num_policy_steps'],
variant['algo_params']['num_train_tasks_per_episode'],
variant['algo_params']['num_initial_collection_cycles_per_task'],
variant['algo_params']['num_trajectories_per_task'],
variant['algo_params']['num_exploration_trajectories_per_task'],
variant['algo_params']['num_eval_trajectories'],
variant['algo_params']['showcase_every'],
variant['algo_params']['snapshot_gap'],
variant['algo_params']['num_showcase_deterministic'],
variant['algo_params']['num_showcase_non_deterministic'],
variant['algo_params']['exploration_agent'] is not None,
variant['algo_params']['exploration_by_probability'],
variant['algo_params']['exploration_fixed_probability'],
experiment_log_dir,
latent_dim,
)
return algorithm, networks, rollout_coordinator, replay_buffer, train_tasks, test_tasks
def load_networks(variant, networks, cemrl_compatibility):
itr = variant['showcase_itr']
path = variant['path_to_weights']
for name, net in networks.items():
state_dict = torch.load(os.path.join(path, name + '_itr_' + str(itr) + '.pth'), map_location='cpu')
if cemrl_compatibility:
if name == 'encoder':
state_dict = OrderedDict((key.replace('shared_encoder', 'shared_encoder.layers'), val)
for key, val in state_dict.items())
net.load_state_dict(state_dict)
def experiment(variant):
env, experiment_log_dir = setup_environment(variant)
algorithm, networks, *_ = initialize_networks(variant, env, experiment_log_dir)
if variant['path_to_weights'] is not None:
load_networks(variant, networks)
if ptu.gpu_enabled():
algorithm.to()
# run the algorithm
algorithm.train()
def deep_update_dict(fr, to):
''' update dict of dicts with new values '''
# assume dicts have same keys
for k, v in fr.items():
if type(v) is dict:
deep_update_dict(v, to[k])
else:
to[k] = v
return to
@click.command()
@click.argument('config', default="configs/thesis/mw-reach-line.json")
@click.option('--weights', default=None)
@click.option('--weights_itr', default=None)
@click.option('--gpu', default=None, type=int)
@click.option('--num_workers', default=None, type=int)
@click.option('--use_mp', default=None, type=bool)
@click.option('--docker', is_flag=True, default=False)
@click.option('--debug', is_flag=True, default=False)
def main(config, weights, weights_itr, gpu, use_mp, num_workers, docker, debug):
variant = default_config
if config:
with open(os.path.join(config)) as f:
exp_params = json.load(f)
variant = deep_update_dict(exp_params, variant)
# Overwrite machine defaults only when explicitly given
if gpu is not None:
variant['util_params']['gpu_id'] = gpu
if num_workers is not None:
variant['util_params']['num_workers'] = num_workers
if use_mp is not None:
variant['util_params']['use_multiprocessing'] = use_mp
if weights is not None:
variant['path_to_weights'] = weights
if weights_itr is not None:
variant['showcase_itr'] = weights_itr
variant['config_name'] = Path(config).stem
experiment(variant)
if __name__ == "__main__":
main()