-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
608 lines (548 loc) · 18 KB
/
main.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
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
import argparse
import horovod.torch as hvd
import json
import logging
import math
import os
import signal
import sys
import torch.multiprocessing as mp
import torch.utils.data.distributed
from argparse import Namespace
from ast import literal_eval
from datetime import datetime
from os import path, makedirs
import backbone
import models
from data_regime import DataRegime
from optimizer_regime import OptimizerRegime
from train.train import train
from utils.log import setup_logging, ResultsLog, PerformanceResultsLog
from utils.model import save_checkpoint
from utils.yaml_params import YParams
from utils.utils import move_cuda
backbone_model_names = sorted(
name
for name in backbone.__dict__
if not name.startswith("__") and callable(backbone.__dict__[name])
)
model_names = sorted(
name
for name in models.__dict__
if not name.startswith("__") and callable(models.__dict__[name])
)
parser = argparse.ArgumentParser(
description="Distributed deep/continual learning with Horovod + PyTorch"
)
parser.add_argument(
"--yaml-config",
default="config.yaml",
type=str,
help="path to yaml file containing training configs",
)
parser.add_argument(
"--config",
default="",
type=str,
help="name of desired config in yaml file",
)
parser.add_argument(
"--wandb-project",
default="distributed-continual-learning",
type=str,
help="name of your Weigths & Biases project",
)
parser.add_argument(
"--dataset", metavar="DATASET", default="mnist", help="dataset name"
)
parser.add_argument(
"--dataset-dir",
default="./datasets",
help="location of the training dataset in the local filesystem (will be downloaded if needed)",
)
parser.add_argument(
"--tasksets-config",
default="{}",
help="additional taskset configuration (useful for continual learning)",
)
parser.add_argument(
"--backbone",
metavar="BACKBONE",
default="resnet50",
choices=backbone_model_names,
help="available backbone models: " + " | ".join(backbone_model_names),
)
parser.add_argument(
"--backbone-config", default="{}", help="additional backbone configuration"
)
parser.add_argument(
"--model",
metavar="MODEL",
default="Vanilla",
choices=model_names,
help="available models: " + " | ".join(model_names),
)
parser.add_argument(
"--model-config",
default="{}",
help="model configuration",
)
parser.add_argument(
"--buffer-config",
default="{}",
help="rehearsal buffer configuration",
)
parser.add_argument(
"--load-checkpoint",
default="",
type=str,
metavar="PATH",
help="path to latest checkpoint (default: none)",
)
parser.add_argument(
"--batch-size",
type=int,
default=64,
metavar="N",
help="input batch size for training (default: 64)",
)
parser.add_argument(
"--eval-batch-size",
type=int,
default=-1,
metavar="N",
help="input batch size for testing (default: same as training)",
)
parser.add_argument(
"--dataloader-workers",
type=int,
default=0,
help="number of dataloaders workers to spawn",
)
parser.add_argument(
"--epochs",
type=int,
default=25,
metavar="N",
help="number of epochs to train (default: 25)",
)
parser.add_argument(
"--warmup-epochs",
type=int,
default=5,
metavar="N",
help="number of epochs to warmup LR, if scheduler supports (default: 5)",
)
parser.add_argument(
"--training-only",
action="store_true",
help="don't validate after every epoch, only after training on a new task",
)
parser.add_argument(
"--lr",
type=float,
default=0.01,
metavar="LR",
help="learning rate for a single GPU (default: 0.01)",
)
parser.add_argument(
"--momentum",
type=float,
default=0.5,
metavar="M",
help="SGD momentum (default: 0.5)",
)
parser.add_argument(
"--optimizer-regime",
default="",
help="optimizer regime, as an array of dicts containing the epoch key",
)
parser.add_argument(
"--no-cuda",
action="store_true",
default=False,
help="disables CUDA training",
)
parser.add_argument(
"--seed",
type=int,
default=42,
metavar="S",
help="random seed (default: 42)",
)
parser.add_argument("--log-level", default="info", help="logging level")
parser.add_argument(
"--log-interval",
type=int,
default=10,
metavar="N",
help="how many batches to wait before logging training status",
)
parser.add_argument(
"--use-amp",
action="store_true",
default=False,
help="enable Automatic Mixed Precision training",
)
parser.add_argument(
"--use-dali",
action="store_true",
default=False,
help="enable DALI loading pipeline",
)
parser.add_argument(
"--fp16-dali",
action="store_true",
default=False,
help="load images in half precision",
)
parser.add_argument(
"--fp16-allreduce",
action="store_true",
default=False,
help="use fp16 compression during allreduce",
)
parser.add_argument(
"--use-adasum",
action="store_true",
default=False,
help="use adasum helping with convergence at scale",
)
parser.add_argument(
"--batches-per-allreduce",
type=int,
default=1,
help="number of batches processed locally before "
"executing allreduce across workers; it multiplies "
"effective batch size.",
)
parser.add_argument(
"--gradient-predivide-factor",
type=float,
default=1.0,
help="apply gradient predivide factor in optimizer (default: 1.0)",
)
parser.add_argument(
"--weight-decay",
"--wd",
type=float,
default=0,
metavar="W",
help="weight decay (default: 0)",
)
parser.add_argument(
"--results-dir",
metavar="RESULTS_DIR",
default="./results",
help="results dir",
)
parser.add_argument(
"--nsys-run",
action="store_true",
default=False,
help="override params to retrieve nsys metrics on one epoch only",
)
parser.add_argument("--save-dir", metavar="SAVE_DIR", default="", help="saved folder")
def main():
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
args.evaluate = not args.training_only
mp.set_start_method("spawn")
params = vars(args)
if args.config:
yparams = YParams(os.path.abspath(args.yaml_config), args.config)
for k, v in params.items():
yparam = yparams[k]
if yparam:
params[k] = yparam
if (
k == "buffer_config"
or k == "backbone_config"
or k == "tasksets_config"
or k == "optimizer_regime"
):
if v:
params[k] = str(literal_eval(v) | literal_eval(yparam))
args = Namespace(**params)
# Nsys mode: train for one epoch only
if args.nsys_run:
args.epochs = 1
# Horovod: initialize library.
hvd.init()
args.gpus = hvd.size()
torch.manual_seed(args.seed)
if args.cuda:
# Horovod: pin GPU to local rank.
torch.cuda.set_device(hvd.local_rank())
torch.cuda.manual_seed(args.seed)
# Horovod: limit # of CPU threads to be used per worker.
torch.set_num_threads(1)
save_path = ""
if hvd.rank() == 0:
try:
import wandb
wandb.init(project=args.wandb_project)
run_name = f"{datetime.now().strftime('%Y%m%d-%H%M%S')}-{wandb.run.name}"
if args.model is not None:
run_name = f"{args.model}-{run_name}"
wandb.run.name = run_name
if args.save_dir == "":
args.save_dir = run_name
save_path = path.join(args.results_dir, args.save_dir)
if not path.exists(save_path):
makedirs(save_path)
with open(path.join(save_path, "args.json"), "w", encoding="utf-8") as f:
json.dump(args.__dict__, f, indent=2)
wandb.save(path.join(save_path, "args.json"))
wandb.config.update(args)
logging.info("Saving to %s", save_path)
except:
pass
setup_logging(
path.join(save_path, f"log_{hvd.rank()}.txt"),
level=args.log_level,
)
device = "GPU" if args.cuda else "CPU"
logging.info("Number of %ss: %d", device, hvd.size())
logging.info("Run arguments: %s", args)
xp = Experiment(args, save_path)
xp.run()
try:
wandb.finish()
except:
pass
logging.info("Done 🎉🎉🎉")
sys.exit(0)
class Experiment:
resume_from_task = 0
resume_from_epoch = 0
def __init__(self, args, save_path=""):
self.args = args
self.save_path = save_path
total_num_classes = self.prepare_dataset()
batch_metrics_path = path.join(self.save_path, "batch_metrics")
batch_metrics = PerformanceResultsLog(batch_metrics_path)
self.create_model(total_num_classes, batch_metrics)
def create_model(self, total_num_classes, batch_metrics=None):
# -------------------------------------------------------------------------------------------------#
# --------------------------#
# ----- BACKBONE MODEL -----#
# --------------------------#
# Creating the model
backbone_config = {
"world_size": hvd.size(),
"num_classes": total_num_classes,
"lr": self.args.lr,
"batches_per_allreduce": self.args.batches_per_allreduce,
"warmup_epochs": self.args.warmup_epochs,
"num_epochs": self.args.epochs,
"total_num_samples": self.train_data_regime.total_num_samples,
"num_steps_per_epoch": len(self.train_data_regime.get_loader(0)),
"batch_size": self.args.batch_size,
}
if self.args.backbone_config != "":
backbone_config = dict(
backbone_config, **literal_eval(self.args.backbone_config)
)
backbone_model = getattr(backbone, self.args.backbone)(backbone_config)
logging.info(
"Created backbone model %s with configuration: %s",
self.args.backbone,
json.dumps(backbone_config, indent=2),
)
num_parameters = sum([l.nelement() for l in backbone_model.parameters()])
logging.info("Number of parameters: %d", num_parameters)
backbone_model = move_cuda(backbone_model, self.args.cuda)
# Building the optimizer regime
if self.args.optimizer_regime != "":
optimizer_regime_dict = literal_eval(self.args.optimizer_regime)
else:
optimizer_regime_dict = getattr(backbone_model, "regime")
logging.info("Optimizer regime: %s", optimizer_regime_dict)
optimizer_regime = OptimizerRegime(
backbone_model,
hvd.Compression.fp16 if self.args.fp16_allreduce else hvd.Compression.none,
self.args.batches_per_allreduce,
hvd.Adasum if self.args.use_adasum else hvd.Average,
self.args.gradient_predivide_factor,
optimizer_regime_dict,
self.args.use_amp,
)
# -------------------------------------------------------------------------------------------------#
# ------------------#
# ----- BUFFER -----#
# ------------------#
buffer_config = literal_eval(self.args.buffer_config)
rehearsal_ratio = buffer_config.pop("rehearsal_ratio", 30)
if bool(buffer_config):
total_num_samples = self.train_data_regime.total_num_samples
if (
literal_eval(self.args.tasksets_config).get("scenario", "class")
== "reconstruction"
):
# This is a workaround, as the ptycho dataset contains perspectives
# and not individual diffraction patterns i.e., samples, contained
# in a perspective. After filtering out empty diffractions, ~450 of
# them are contained in a perspective.
total_num_samples *= 450
budget_per_class = math.floor(
total_num_samples
* rehearsal_ratio
/ 100
/ total_num_classes
/ hvd.size()
)
assert (
budget_per_class > 0
), "Choose rehearsal_ratio so as to to store at least some samples per class on all processes"
buffer_config |= {"budget_per_class": budget_per_class}
# -------------------------------------------------------------------------------------------------#
# -----------------#
# ----- MODEL -----#
# -----------------#
model_config = literal_eval(self.args.model_config)
# Creating the continual learning model
model = getattr(models, self.args.model)
self.model = model(
backbone_model,
optimizer_regime,
self.args.use_amp,
self.args.nsys_run,
self.args.batch_size,
model_config,
buffer_config,
batch_metrics,
)
logging.info(
"Created model with buffer configuration: %s",
json.dumps(buffer_config, indent=2),
)
# -------------------------------------------------------------------------------------------------#
# ----------------------#
# ----- CHECKPOINT -----#
# ----------------------#
# Loading the checkpoint if given
resume_from_task = 0
resume_from_epoch = 0
if hvd.rank() == 0 and self.args.load_checkpoint:
if not path.isfile(self.args.load_checkpoint):
parser.error(f"Invalid checkpoint: {self.args.load_checkpoint}")
checkpoint = torch.load(self.args.load_checkpoint, map_location="cpu")
# Load checkpoint
logging.info("Loading model %s...", self.args.load_checkpoint)
self.model.backbone.load_state_dict(checkpoint["state_dict"])
# optimizer_regime.load_state_dict(checkpoint["optimizer_state_dict"])
# Broadcast resume information
resume_from_task = checkpoint["task"]
resume_from_epoch = checkpoint["epoch"]
logging.info(
"Resuming from task %d epoch %d", resume_from_task, resume_from_epoch
)
self.resume_from_task = hvd.broadcast(
torch.tensor(resume_from_task), root_rank=0, name="resume_from_task"
).item()
self.resume_from_epoch = hvd.broadcast(
torch.tensor(resume_from_epoch), root_rank=0, name="resume_from_epoch"
).item()
# Saving an initial checkpoint
save_checkpoint(
{
"task": 0,
"epoch": 0,
"model": self.args.backbone,
"state_dict": self.model.backbone.state_dict(),
# "optimizer_state_dict": self.model.optimizer_regime.state_dict(),
},
self.save_path,
is_initial=True,
dummy=hvd.rank() > 0,
)
logging.info("Initial checkpoint created")
def prepare_dataset(self):
defaults = {
"dataset": self.args.dataset,
"dataset_dir": self.args.dataset_dir,
"pin_memory": True,
"drop_last": True,
"use_dali": self.args.use_dali,
# https://github.com/horovod/horovod/issues/2053
"num_workers": self.args.dataloader_workers,
**literal_eval(self.args.tasksets_config),
}
self.train_data_regime = DataRegime(
hvd,
{
**defaults,
"split": "train",
"batch_size": self.args.batch_size * self.args.batches_per_allreduce,
},
)
logging.info(
"Created train data regime: %s", str(self.train_data_regime.config)
)
self.validate_data_regime = DataRegime(
hvd,
{
**defaults,
"split": "validate",
"batch_size": self.args.eval_batch_size
if self.args.eval_batch_size > 0
else self.args.batch_size,
},
)
logging.info(
"Created test data regime: %s", str(self.validate_data_regime.config)
)
return self.train_data_regime.total_num_classes
def run(self):
dl_metrics_path = path.join(self.save_path, "dl_metrics")
dl_metrics = ResultsLog(
dl_metrics_path,
title=f"DL metrics - {self.args.save_dir}",
dummy=hvd.rank() > 0 or hvd.local_rank() > 0,
)
tasks_metrics_path = path.join(self.save_path, "tasks_metrics")
tasks_metrics = ResultsLog(
tasks_metrics_path,
title=f"Tasks metrics - {self.args.save_dir}",
dummy=hvd.rank() > 0 or hvd.local_rank() > 0,
)
time_metrics_path = path.join(self.save_path, "time_metrics")
time_metrics = ResultsLog(
time_metrics_path,
title=f"Time metrics - {self.args.save_dir}",
dummy=hvd.rank() > 0 or hvd.local_rank() > 0,
)
final_checkpoint_info = train(
self.model,
self.train_data_regime,
self.validate_data_regime,
literal_eval(self.args.tasksets_config).get("epochs", self.args.epochs),
resume_from_task=self.resume_from_task,
resume_from_epoch=self.resume_from_epoch,
evaluate=self.args.evaluate,
log_interval=self.args.log_interval,
dl_metrics=dl_metrics,
tasks_metrics=tasks_metrics,
time_metrics=time_metrics,
)
save_checkpoint(
final_checkpoint_info,
self.save_path,
is_final=True,
dummy=hvd.rank() > 0,
)
def on_exit(sig, frame):
logging.info("Interrupted")
try:
wandb.finish()
except:
pass
os.system(
"kill $(ps aux | grep multiprocessing.spawn | grep -v grep | awk '{print $2}')"
)
sys.exit(0)
if __name__ == "__main__":
signal.signal(signal.SIGINT, on_exit)
signal.signal(signal.SIGTERM, on_exit)
main()