/
imitation.py
482 lines (386 loc) · 20.3 KB
/
imitation.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
480
481
482
import copy
import gym
import time
import datetime
import numpy as np
import sys
import itertools
import torch
from babyai.evaluate import batch_evaluate
import babyai.utils as utils
from babyai.rl import DictList
from babyai.model import ACModel
import multiprocessing
import os
import json
import logging
logger = logging.getLogger(__name__)
import numpy
class EpochIndexSampler:
"""
Generate smart indices for epochs that are smaller than the dataset size.
The usecase: you have a code that has a strongly baken in notion of an epoch,
e.g. you can only validate in the end of the epoch. That ties a lot of
aspects of training to the size of the dataset. You may want to validate
more often than once per a complete pass over the dataset.
This class helps you by generating a sequence of smaller epochs that
use different subsets of the dataset, as long as this is possible.
This allows you to keep the small advantage that sampling without replacement
provides, but also enjoy smaller epochs.
"""
def __init__(self, n_examples, epoch_n_examples):
self.n_examples = n_examples
self.epoch_n_examples = epoch_n_examples
self._last_seed = None
def _reseed_indices_if_needed(self, seed):
if seed == self._last_seed:
return
rng = numpy.random.RandomState(seed)
self._indices = list(range(self.n_examples))
rng.shuffle(self._indices)
logger.info('reshuffle the dataset')
self._last_seed = seed
def get_epoch_indices(self, epoch):
"""Return indices corresponding to a particular epoch.
Tip: if you call this function with consecutive epoch numbers,
you will avoid expensive reshuffling of the index list.
"""
seed = epoch * self.epoch_n_examples // self.n_examples
offset = epoch * self.epoch_n_examples % self.n_examples
indices = []
while len(indices) < self.epoch_n_examples:
self._reseed_indices_if_needed(seed)
n_lacking = self.epoch_n_examples - len(indices)
indices += self._indices[offset:offset + min(n_lacking, self.n_examples - offset)]
offset = 0
seed += 1
return indices
class ImitationLearning(object):
def __init__(self, args, ):
self.args = args
utils.seed(self.args.seed)
self.val_seed = self.args.val_seed
# args.env is a list when training on multiple environments
if getattr(args, 'multi_env', None):
self.env = [gym.make(item) for item in args.multi_env]
self.train_demos = []
for demos, episodes in zip(args.multi_demos, args.multi_episodes):
demos_path = utils.get_demos_path(demos, None, None, valid=False)
logger.info('loading {} of {} demos'.format(episodes, demos))
train_demos = utils.load_demos(demos_path)
logger.info('loaded demos')
if episodes > len(train_demos):
raise ValueError("there are only {} train demos in {}".format(len(train_demos), demos))
self.train_demos.extend(train_demos[:episodes])
logger.info('So far, {} demos loaded'.format(len(self.train_demos)))
self.val_demos = []
for demos, episodes in zip(args.multi_demos, [args.val_episodes] * len(args.multi_demos)):
demos_path_valid = utils.get_demos_path(demos, None, None, valid=True)
logger.info('loading {} of {} valid demos'.format(episodes, demos))
valid_demos = utils.load_demos(demos_path_valid)
logger.info('loaded demos')
if episodes > len(valid_demos):
logger.info('Using all the available {} demos to evaluate valid. accuracy'.format(len(valid_demos)))
self.val_demos.extend(valid_demos[:episodes])
logger.info('So far, {} valid demos loaded'.format(len(self.val_demos)))
logger.info('Loaded all demos')
observation_space = self.env[0].observation_space
action_space = self.env[0].action_space
else:
self.env = gym.make(self.args.env)
demos_path = utils.get_demos_path(args.demos, args.env, args.demos_origin, valid=False)
demos_path_valid = utils.get_demos_path(args.demos, args.env, args.demos_origin, valid=True)
logger.info('loading demos')
self.train_demos = utils.load_demos(demos_path)
logger.info('loaded demos')
if args.episodes:
if args.episodes > len(self.train_demos):
raise ValueError("there are only {} train demos".format(len(self.train_demos)))
self.train_demos = self.train_demos[:args.episodes]
self.val_demos = utils.load_demos(demos_path_valid)
if args.val_episodes > len(self.val_demos):
logger.info('Using all the available {} demos to evaluate valid. accuracy'.format(len(self.val_demos)))
self.val_demos = self.val_demos[:self.args.val_episodes]
observation_space = self.env.observation_space
action_space = self.env.action_space
self.obss_preprocessor = utils.ObssPreprocessor(args.model, observation_space,
getattr(self.args, 'pretrained_model', None))
# Define actor-critic model
self.acmodel = utils.load_model(args.model, raise_not_found=False)
if self.acmodel is None:
if getattr(self.args, 'pretrained_model', None):
self.acmodel = utils.load_model(args.pretrained_model, raise_not_found=True)
else:
logger.info('Creating new model')
self.acmodel = ACModel(self.obss_preprocessor.obs_space, action_space,
args.image_dim, args.memory_dim, args.instr_dim,
not self.args.no_instr, self.args.instr_arch,
not self.args.no_mem, self.args.arch)
self.obss_preprocessor.vocab.save()
utils.save_model(self.acmodel, args.model)
self.acmodel.train()
if torch.cuda.is_available():
self.acmodel.cuda()
self.optimizer = torch.optim.Adam(self.acmodel.parameters(), self.args.lr, eps=self.args.optim_eps)
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=100, gamma=0.9)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@staticmethod
def default_model_name(args):
if getattr(args, 'multi_env', None):
# It's better to specify one's own model name for this scenario
named_envs = '-'.join(args.multi_env)
else:
named_envs = args.env
# Define model name
suffix = datetime.datetime.now().strftime("%y-%m-%d-%H-%M-%S")
instr = args.instr_arch if args.instr_arch else "noinstr"
model_name_parts = {
'envs': named_envs,
'arch': args.arch,
'instr': instr,
'seed': args.seed,
'suffix': suffix}
default_model_name = "{envs}_IL_{arch}_{instr}_seed{seed}_{suffix}".format(**model_name_parts)
if getattr(args, 'pretrained_model', None):
default_model_name = args.pretrained_model + '_pretrained_' + default_model_name
return default_model_name
def starting_indexes(self, num_frames):
if num_frames % self.args.recurrence == 0:
return np.arange(0, num_frames, self.args.recurrence)
else:
return np.arange(0, num_frames, self.args.recurrence)[:-1]
def run_epoch_recurrence(self, demos, is_training=False, indices=None):
if not indices:
indices = list(range(len(demos)))
if is_training:
np.random.shuffle(indices)
batch_size = min(self.args.batch_size, len(demos))
offset = 0
if not is_training:
self.acmodel.eval()
# Log dictionary
log = {"entropy": [], "policy_loss": [], "accuracy": []}
start_time = time.time()
frames = 0
for batch_index in range(len(indices) // batch_size):
logger.info("batch {}, FPS so far {}".format(
batch_index, frames / (time.time() - start_time) if frames else 0))
batch = [demos[i] for i in indices[offset: offset + batch_size]]
frames += sum([len(demo[3]) for demo in batch])
_log = self.run_epoch_recurrence_one_batch(batch, is_training=is_training)
log["entropy"].append(_log["entropy"])
log["policy_loss"].append(_log["policy_loss"])
log["accuracy"].append(_log["accuracy"])
offset += batch_size
log['total_frames'] = frames
if not is_training:
self.acmodel.train()
return log
def run_epoch_recurrence_one_batch(self, batch, is_training=False):
batch = utils.demos.transform_demos(batch)
batch.sort(key=len, reverse=True)
# Constructing flat batch and indices pointing to start of each demonstration
flat_batch = []
inds = [0]
for demo in batch:
flat_batch += demo
inds.append(inds[-1] + len(demo))
flat_batch = np.array(flat_batch)
inds = inds[:-1]
num_frames = len(flat_batch)
mask = np.ones([len(flat_batch)], dtype=np.float64)
mask[inds] = 0
mask = torch.tensor(mask, device=self.device, dtype=torch.float).unsqueeze(1)
# Observations, true action, values and done for each of the stored demostration
obss, action_true, done = flat_batch[:, 0], flat_batch[:, 1], flat_batch[:, 2]
action_true = torch.tensor([action for action in action_true], device=self.device, dtype=torch.long)
# Memory to be stored
memories = torch.zeros([len(flat_batch), self.acmodel.memory_size], device=self.device)
episode_ids = np.zeros(len(flat_batch))
memory = torch.zeros([len(batch), self.acmodel.memory_size], device=self.device)
preprocessed_first_obs = self.obss_preprocessor(obss[inds], device=self.device)
instr_embedding = self.acmodel._get_instr_embedding(preprocessed_first_obs.instr)
# Loop terminates when every observation in the flat_batch has been handled
while True:
# taking observations and done located at inds
obs = obss[inds]
done_step = done[inds]
preprocessed_obs = self.obss_preprocessor(obs, device=self.device)
with torch.no_grad():
# taking the memory till len(inds), as demos beyond that have already finished
new_memory = self.acmodel(
preprocessed_obs,
memory[:len(inds), :], instr_embedding[:len(inds)])['memory']
memories[inds, :] = memory[:len(inds), :]
memory[:len(inds), :] = new_memory
episode_ids[inds] = range(len(inds))
# Updating inds, by removing those indices corresponding to which the demonstrations have finished
inds = inds[:len(inds) - sum(done_step)]
if len(inds) == 0:
break
# Incrementing the remaining indices
inds = [index + 1 for index in inds]
# Here, actual backprop upto args.recurrence happens
final_loss = 0
final_entropy, final_policy_loss, final_value_loss = 0, 0, 0
indexes = self.starting_indexes(num_frames)
memory = memories[indexes]
accuracy = 0
total_frames = len(indexes) * self.args.recurrence
for _ in range(self.args.recurrence):
obs = obss[indexes]
preprocessed_obs = self.obss_preprocessor(obs, device=self.device)
action_step = action_true[indexes]
mask_step = mask[indexes]
model_results = self.acmodel(
preprocessed_obs, memory * mask_step,
instr_embedding[episode_ids[indexes]])
dist = model_results['dist']
memory = model_results['memory']
entropy = dist.entropy().mean()
policy_loss = -dist.log_prob(action_step).mean()
loss = policy_loss - self.args.entropy_coef * entropy
action_pred = dist.probs.max(1, keepdim=True)[1]
accuracy += float((action_pred == action_step.unsqueeze(1)).sum()) / total_frames
final_loss += loss
final_entropy += entropy
final_policy_loss += policy_loss
indexes += 1
final_loss /= self.args.recurrence
if is_training:
self.optimizer.zero_grad()
final_loss.backward()
self.optimizer.step()
log = {}
log["entropy"] = float(final_entropy / self.args.recurrence)
log["policy_loss"] = float(final_policy_loss / self.args.recurrence)
log["accuracy"] = float(accuracy)
return log
def validate(self, episodes, verbose=True):
if verbose:
logger.info("Validating the model")
if getattr(self.args, 'multi_env', None):
agent = utils.load_agent(self.env[0], model_name=self.args.model, argmax=True)
else:
agent = utils.load_agent(self.env, model_name=self.args.model, argmax=True)
# Setting the agent model to the current model
agent.model = self.acmodel
agent.model.eval()
logs = []
for env_name in ([self.args.env] if not getattr(self.args, 'multi_env', None)
else self.args.multi_env):
logs += [batch_evaluate(agent, env_name, self.val_seed, episodes)]
self.val_seed += episodes
agent.model.train()
return logs
def collect_returns(self):
logs = self.validate(episodes=self.args.eval_episodes, verbose=False)
mean_return = {tid: np.mean(log["return_per_episode"]) for tid, log in enumerate(logs)}
return mean_return
def train(self, train_demos, writer, csv_writer, status_path, header, reset_status=False):
# Load the status
def initial_status():
return {'i': 0,
'num_frames': 0,
'patience': 0}
status = initial_status()
if os.path.exists(status_path) and not reset_status:
with open(status_path, 'r') as src:
status = json.load(src)
elif not os.path.exists(os.path.dirname(status_path)):
# Ensure that the status directory exists
os.makedirs(os.path.dirname(status_path))
# If the batch size is larger than the number of demos, we need to lower the batch size
if self.args.batch_size > len(train_demos):
self.args.batch_size = len(train_demos)
logger.info("Batch size too high. Setting it to the number of train demos ({})".format(len(train_demos)))
# Model saved initially to avoid "Model not found Exception" during first validation step
utils.save_model(self.acmodel, self.args.model)
# best mean return to keep track of performance on validation set
best_success_rate, patience, i = 0, 0, 0
total_start_time = time.time()
epoch_length = self.args.epoch_length
if not epoch_length:
epoch_length = len(train_demos)
index_sampler = EpochIndexSampler(len(train_demos), epoch_length)
while status['i'] < getattr(self.args, 'epochs', int(1e9)):
if 'patience' not in status: # if for some reason you're finetuining with IL an RL pretrained agent
status['patience'] = 0
# Do not learn if using a pre-trained model that already lost patience
if status['patience'] > self.args.patience:
break
if status['num_frames'] > self.args.frames:
break
update_start_time = time.time()
indices = index_sampler.get_epoch_indices(status['i'])
log = self.run_epoch_recurrence(train_demos, is_training=True, indices=indices)
# Learning rate scheduler
self.scheduler.step()
status['num_frames'] += log['total_frames']
status['i'] += 1
update_end_time = time.time()
# Print logs
if status['i'] % self.args.log_interval == 0:
total_ellapsed_time = int(time.time() - total_start_time)
fps = log['total_frames'] / (update_end_time - update_start_time)
duration = datetime.timedelta(seconds=total_ellapsed_time)
for key in log:
log[key] = np.mean(log[key])
train_data = [status['i'], status['num_frames'], fps, total_ellapsed_time,
log["entropy"], log["policy_loss"], log["accuracy"]]
logger.info(
"U {} | F {:06} | FPS {:04.0f} | D {} | H {:.3f} | pL {: .3f} | A {: .3f}".format(*train_data))
# Log the gathered data only when we don't evaluate the validation metrics. It will be logged anyways
# afterwards when status['i'] % self.args.val_interval == 0
if status['i'] % self.args.val_interval != 0:
# instantiate a validation_log with empty strings when no validation is done
validation_data = [''] * len([key for key in header if 'valid' in key])
assert len(header) == len(train_data + validation_data)
if self.args.tb:
for key, value in zip(header, train_data):
writer.add_scalar(key, float(value), status['num_frames'])
csv_writer.writerow(train_data + validation_data)
if status['i'] % self.args.val_interval == 0:
valid_log = self.validate(self.args.val_episodes)
mean_return = [np.mean(log['return_per_episode']) for log in valid_log]
success_rate = [np.mean([1 if r > 0 else 0 for r in log['return_per_episode']]) for log in
valid_log]
val_log = self.run_epoch_recurrence(self.val_demos)
validation_accuracy = np.mean(val_log["accuracy"])
if status['i'] % self.args.log_interval == 0:
validation_data = [validation_accuracy] + mean_return + success_rate
logger.info(("Validation: A {: .3f} " + ("| R {: .3f} " * len(mean_return) +
"| S {: .3f} " * len(success_rate))
).format(*validation_data))
assert len(header) == len(train_data + validation_data)
if self.args.tb:
for key, value in zip(header, train_data + validation_data):
writer.add_scalar(key, float(value), status['num_frames'])
csv_writer.writerow(train_data + validation_data)
# In case of a multi-env, the update condition would be "better mean success rate" !
if np.mean(success_rate) > best_success_rate:
best_success_rate = np.mean(success_rate)
status['patience'] = 0
with open(status_path, 'w') as dst:
json.dump(status, dst)
# Saving the model
logger.info("Saving best model")
if torch.cuda.is_available():
self.acmodel.cpu()
utils.save_model(self.acmodel, self.args.model + "_best")
self.obss_preprocessor.vocab.save(utils.get_vocab_path(self.args.model + "_best"))
if torch.cuda.is_available():
self.acmodel.cuda()
else:
status['patience'] += 1
logger.info(
"Losing patience, new value={}, limit={}".format(status['patience'], self.args.patience))
if torch.cuda.is_available():
self.acmodel.cpu()
utils.save_model(self.acmodel, self.args.model)
self.obss_preprocessor.vocab.save()
if torch.cuda.is_available():
self.acmodel.cuda()
with open(status_path, 'w') as dst:
json.dump(status, dst)
return best_success_rate