-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainval.py
440 lines (393 loc) · 16.4 KB
/
trainval.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
import os
import torch
import sys
from torch import nn
import numpy as np
import time
from src import datasets
from haven import haven_wizard as hw
import json
from src.metrics import models_clf
from src.fid import fid_score
from src.models import InfoGAN
from src.models.utils import (Argument,
FidWrapper,
validate_and_insert_defaults,
precompute_fid_stats,
load_json_from_file)
from src.models.haven_utils import get_checkpoint, save_checkpoint
# Parallel stuff
import torch.multiprocessing as mp
import torch.distributed as dist
from src import setup_logger
logger = setup_logger.get_logger(__name__)
def setup_mpi(rank, world_size):
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355"
# initialize the process group
dist.init_process_group("nccl", rank=rank, world_size=world_size)
NONETYPE = type(None)
DEFAULTS = {
"load_strict": Argument('load_strict', True, [bool]),
"epochs": Argument('epochs', 200, [int]),
"finetune": Argument('finetune', False, [bool]),
"save_every": Argument('save_every', None, [int, NONETYPE]),
"fid_every": Argument('fid_every', 5, [int]), # compute fid every this many epochs
"fid_N": Argument('fid_N', 5000, [int]), # how many real/fake samples for FID computation
"fid_pretrained": Argument('fid_pretrained', None, [str, NONETYPE]), # fid feature extractor
"semi_sup": Argument('semi_sup', False, [bool]),
"pretrained": Argument('pretrained', None, [str, NONETYPE]),
"dataset": {
"name": Argument("name", "emnist", [str]),
"k_shot": Argument("k_shot", 5, [int]),
"seed": Argument("seed", 0, [int]),
"input_size": Argument("input_size", 32, [int]),
"pad_length": Argument("pad_length", 0, [int]),
"transform_kwargs": Argument("transform_kwargs", {}, [dict, NONETYPE])
},
"batch_size": Argument('batch_size', 32, [int]),
"model": Argument("model", {}, [dict]), # model kwargs, see infogan.py
"optim": {
"beta1": Argument("beta1", 0.9, [float]),
"beta2": Argument("beta2", 0.9, [float]),
"lr": Argument("lr", 2e-4, [float]),
"n_gen": Argument("n_gen", 1, [int]),
"weight_decay": Argument("weight_decay", 1e-6, [float]),
"eps": Argument("eps", 1e-8, [float])
},
# args that are no longer used, added as dummy args to stop
# script from crashing:
"pretrained_enc": Argument("pretrained_enc", None, [NONETYPE])
}
def validate_args(dd):
if not dd["finetune"] and dd["dataset"]["pad_length"] > 0:
raise Exception("dataset.pad_length only supported for finetuning")
if dd["semi_sup"] and not dd["finetune"]:
raise Exception("semi-sup mode only supported if finetune==True")
def trainval(exp_dict, savedir, args):
logger.info("Validating and inserting defaults...")
validate_and_insert_defaults(exp_dict, DEFAULTS)
logger.info("Extra validating args...")
validate_args(exp_dict)
if not os.path.exists("{}/exp_dict.json".format(savedir)):
# This would trigger if we're launching an experiment
# using launch.py.
with open("{}/exp_dict.json".format(savedir), "w") as f:
f.write(json.dumps(exp_dict))
else:
# If we are using `launch_haven.py`, it does write its own `exp_dict.json`
# into `savedir` but that is before the arg validation, and here we want
# `exp_dict.json` to inherit the defaults from `DEFAULTS`. Because of
# this, launch_haven.py has a specific arg called "--disable_rewrite"
# which should be set to true if an experiment is being resumed.
pass
world_size = int(os.environ["WORLD_SIZE"])
if world_size == 0:
logger.info("WORLD_SIZE==0, running on a single process")
_trainval(rank=0, world_size=1, exp_dict=exp_dict, savedir=savedir, args=args)
else:
logger.info("WORLD_SIZE>0, running on multiprocess...")
mp.spawn(
_trainval,
args=(world_size, exp_dict, savedir, args),
nprocs=world_size,
join=True,
)
def _trainval(rank, world_size, exp_dict, savedir, args):
logger.info("RANK: {}, WORLD SIZE: {}".format(rank, world_size))
setup_mpi(rank, world_size)
dataset_name = exp_dict["dataset"]["name"]
dataset_input_size = exp_dict["dataset"]["input_size"]
dataset_seed = exp_dict["dataset"]["seed"]
dataset_transform = exp_dict["dataset"]["transform_kwargs"]
batch_size = exp_dict["batch_size"]
load_strict = True
if "load_strict" in exp_dict:
load_strict = exp_dict["load_strict"]
if not load_strict:
logger.warning(
"`load_strict` is not set, which means loading pre-trained weights "
" may work even when the model definition has changed"
)
# Load datasets
# -------------
train_dataset = datasets.get_dataset(
class_split="train",
datadir=args.datadir,
dataset=dataset_name,
seed=dataset_seed,
input_size=dataset_input_size,
which_set=None,
k_shot=None,
return_pairs=True,
transform_kwargs=dataset_transform,
)
distributed = True if world_size > 0 else False
train_loader, train_sampler = datasets.get_loader(
train_dataset,
batch_size=batch_size,
num_workers=args.num_workers,
distributed=distributed,
)
logger.info("train dataset: {}\n".format(train_dataset))
dev_sampler = None
# Load model and get checkpoint
model = InfoGAN(
input_size=dataset_input_size,
exp_dict=exp_dict,
n_classes=train_loader.dataset.n_classes,
train=True,
rank=rank,
finetune=exp_dict["finetune"],
verbose=True,
)
if exp_dict["fid_pretrained"] is not None:
fid_model = models_clf.get_model(
n_classes=train_dataset.n_classes,
freeze_all_except=None,
)
clf_state_dict = torch.load(exp_dict["fid_pretrained"])
clf_cfg = load_json_from_file(
"{}/exp_dict.json".format(os.path.dirname(exp_dict["fid_pretrained"]))
)
# Verify that the k_shot here is the same as what is
# specified in the pretrained exp_dict.json.
if exp_dict["dataset"]["seed"] != clf_cfg["dataset"]["seed"]:
raise Exception(
"dataset.seed does not match that of FID pretrained dataset.seed"
)
logger.info(
"fid_pretrained is set, so using classifier features: {}".format(
exp_dict["fid_pretrained"]
)
)
fid_model.set_state_dict(clf_state_dict)
fid_model = FidWrapper(fid_model.model.f)
else:
fid_model = None
fid_every = exp_dict["fid_every"]
unsup_loader = None
if fid_every >= 0:
fid_N = exp_dict["fid_N"]
train_fid_mean, train_fid_sd = precompute_fid_stats(
train_loader, batch_size, fid_N, model=fid_model
)
if exp_dict["finetune"]:
dev_dataset = datasets.get_dataset(
class_split="valid",
datadir=args.datadir,
dataset=dataset_name,
seed=dataset_seed,
input_size=dataset_input_size,
k_shot=exp_dict["dataset"]["k_shot"],
which_set="supports",
return_pairs=True,
transform_kwargs=dataset_transform,
)
logger.info("dev dataset: {}\n".format(dev_dataset))
dataset_pad_M = exp_dict["dataset"]["pad_length"]
if dataset_pad_M > 0:
dev_dataset = datasets.DuplicateDatasetMTimes(
dev_dataset, M=dataset_pad_M
)
dev_loader, dev_sampler = datasets.get_loader(
dev_dataset,
batch_size=batch_size,
num_workers=args.num_workers,
distributed=distributed,
)
valid_dataset = datasets.get_dataset(
class_split="valid",
datadir=args.datadir,
dataset=dataset_name,
seed=dataset_seed,
input_size=dataset_input_size,
k_shot=exp_dict["dataset"]["k_shot"],
which_set="valid",
return_pairs=True,
transform_kwargs=dataset_transform,
)
logger.info("valid dataset: {}".format(valid_dataset))
valid_loader, _ = datasets.get_loader(
valid_dataset,
batch_size=batch_size,
num_workers=args.num_workers,
distributed=distributed,
)
logger.info("Precomputing global FID stats for valid set")
valid_fid_mean, valid_fid_sd = precompute_fid_stats(
valid_loader, batch_size, fid_N, model=fid_model
)
#logger.info("Precomputing FID stats per class for valid set")
#valid_class_to_stats = _precompute_fid_stats_per_class(
# valid_dataset, batch_size, fid_N, model=fid_model
#)
if exp_dict["semi_sup"]:
unsup_loader, _ = datasets.get_loader(
valid_dataset,
batch_size=batch_size,
num_workers=args.num_workers,
distributed=distributed,
)
# Explicitly set what gpu to put the weights on.
# If map_location is not set, each rank (gpu) will
# load these onto presumably gpu0, causing an OOM
# if we run this code under a resuming script.
chk_dict = get_checkpoint(
savedir,
return_model_state_dict=True,
map_location=lambda storage, loc: storage.cuda(rank),
)
if exp_dict["pretrained"] is not None:
pretrained_chkpt = torch.load(exp_dict["pretrained"])
pretrained_cfg = load_json_from_file(
"{}/exp_dict.json".format(os.path.dirname(exp_dict["pretrained"]))
)
logger.info("Loading pretrained model: {}".format(exp_dict["pretrained"]))
model.set_state_dict(
pretrained_chkpt,
load_opt=False if exp_dict["finetune"] else True,
strict=True,
)
# Verify that the k_shot here is the same as what is
# specified in the pretrained exp_dict.json.
if exp_dict["dataset"]["seed"] != pretrained_cfg["dataset"]["seed"]:
raise Exception(
"dataset.seed does not match that of pretrained dataset.seed"
)
if len(chk_dict["model_state_dict"]):
model.set_state_dict(chk_dict["model_state_dict"], strict=load_strict)
# Run Train-Val loop
# -----------------------------
max_epochs = exp_dict["epochs"]
save_every = exp_dict["save_every"]
if exp_dict["finetune"]:
# If we're just doing supervised finetuning, use global FID
# otherwise, use averaged per-class FID
if exp_dict["semi_sup"]:
# aoc = average over classes
chk_metric = "fid_valid" # used to be cfid_Valid but too expensive for og
else:
chk_metric = "fid_valid"
else:
chk_metric = "fid"
logger.info("chk_metric: {}".format(chk_metric))
if len(chk_dict["score_list"]) == 0:
best_metric = np.inf
else:
metric_scores = [
score[chk_metric] for score in chk_dict["score_list"] if chk_metric in score
]
if len(metric_scores) == 0:
best_metric = np.inf
else:
best_metric = min(metric_scores)
if args.dry_run:
sys.stderr.write("dry run set, terminating...\n")
return
logger.info("Starting epoch: {}".format(chk_dict["epoch"]))
for epoch in range(chk_dict["epoch"], max_epochs):
t0 = time.time()
# TODO: reduce cpu stats as well???
if rank == 0:
score_dict = {}
score_dict["epoch"] = epoch
if train_sampler is not None:
train_sampler.set_epoch(epoch)
if dev_sampler is not None:
dev_sampler.set_epoch(epoch)
# (1) Train GAN.
train_dict_ = model.train_on_loader(
train_loader if not exp_dict["finetune"] else dev_loader,
unsup_loader,
epoch=epoch,
savedir=savedir,
pbar=world_size <= 1,
)
train_dict = {("train_" + key): val for key, val in train_dict_.items()}
if rank == 0:
# TODO: currently we don't do barriers, we only
# save the metrics that are on gpu0
score_dict.update(train_dict)
# score_dict.update(valid_dict)
score_dict["time"] = time.time() - t0
if fid_every >= 0 and epoch % fid_every == 0 and epoch > 0:
logger.info("Computing FID between train and generated...")
generated_imgs_train = model.sample_from_loader(train_loader, N=fid_N)
this_fid_train = fid_score.calculate_fid_given_imgs_and_stats(
generated_imgs_train,
train_fid_mean,
train_fid_sd,
batch_size,
device=0,
model=fid_model,
)
score_dict["fid"] = this_fid_train
if exp_dict["finetune"]:
logger.info("Computing FID between generated and valid")
generated_imgs_dev = model.sample_from_loader(valid_loader, N=fid_N)
this_fid_valid = fid_score.calculate_fid_given_imgs_and_stats(
generated_imgs_dev,
valid_fid_mean,
valid_fid_sd,
batch_size,
device=0,
model=fid_model,
)
score_dict["fid_valid"] = this_fid_valid
"""
logger.info(
"Computing avg per-class FID between generated and valid"
)
avg_per_class_fids = []
for key in valid_class_to_stats.keys():
generated_imgs_dev_this_class = model.sample_from_loader(
dev_loader, N=fid_N, label=key
)
this_fidpc_valid = fid_score.calculate_fid_given_imgs_and_stats(
generated_imgs_dev_this_class,
valid_class_to_stats[key][0],
valid_class_to_stats[key][1],
batch_size,
device=0,
model=fid_model,
)
avg_per_class_fids.append(this_fidpc_valid)
logger.info("avg_per_class_fids:" + str(avg_per_class_fids))
score_dict["cfid_valid"] = np.mean(avg_per_class_fids)
score_dict["cfid_valid_sd"] = np.std(avg_per_class_fids)
"""
if score_dict[chk_metric] < best_metric:
logger.info(
"Best metric: from {}={} to {}={}".format(
chk_metric, best_metric, chk_metric, score_dict[chk_metric]
)
)
best_metric = score_dict[chk_metric]
save_checkpoint(
savedir,
fname_suffix="." + chk_metric,
score_list=chk_dict["score_list"],
model_state_dict=model.get_state_dict(),
verbose=False,
)
chk_dict["score_list"] += [score_dict]
# Save checkpoint
save_checkpoint(
savedir,
score_list=chk_dict["score_list"],
model_state_dict=model.get_state_dict(),
verbose=False,
)
# If `save_every` is defined, save every
# this many epochs.
if save_every is not None:
if epoch > 0 and epoch % save_every == 0:
save_checkpoint(
savedir,
fname_suffix="." + str(epoch),
score_list=chk_dict["score_list"],
model_state_dict=model.get_state_dict(),
verbose=False,
)
print("Experiment completed")