forked from haitongli/knowledge-distillation-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
443 lines (350 loc) · 18.1 KB
/
train.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
"""Main entrance for train/eval with/without KD on CIFAR-10"""
import argparse
import logging
import os
import time
import math
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR
from torch.autograd import Variable
from tqdm import tqdm
import utils
import model.net as net
import model.data_loader as data_loader
import model.resnet as resnet
import model.wrn as wrn
import model.densenet as densenet
import model.resnext as resnext
import model.preresnet as preresnet
from evaluate import evaluate, evaluate_kd
parser = argparse.ArgumentParser()
# parser.add_argument('--data_dir', default='data/64x64_SIGNS', help="Directory for the dataset")
parser.add_argument('--model_dir', default='experiments/base_model',
help="Directory containing params.json")
parser.add_argument('--restore_file', default=None,
help="Optional, name of the file in --model_dir \
containing weights to reload before training") # 'best' or 'train'
def train(model, optimizer, loss_fn, dataloader, metrics, params):
"""Train the model on `num_steps` batches
Args:
model: (torch.nn.Module) the neural network
optimizer: (torch.optim) optimizer for parameters of model
loss_fn:
dataloader:
metrics: (dict)
params: (Params) hyperparameters
"""
# set model to training mode
model.train()
# summary for current training loop and a running average object for loss
summ = []
loss_avg = utils.RunningAverage()
# Use tqdm for progress bar
with tqdm(total=len(dataloader)) as t:
for i, (train_batch, labels_batch) in enumerate(dataloader):
# move to GPU if available
if params.cuda:
train_batch, labels_batch = train_batch.cuda(async=True), \
labels_batch.cuda(async=True)
# convert to torch Variables
train_batch, labels_batch = Variable(train_batch), Variable(labels_batch)
# compute model output and loss
output_batch = model(train_batch)
loss = loss_fn(output_batch, labels_batch)
# clear previous gradients, compute gradients of all variables wrt loss
optimizer.zero_grad()
loss.backward()
# performs updates using calculated gradients
optimizer.step()
# Evaluate summaries only once in a while
if i % params.save_summary_steps == 0:
# extract data from torch Variable, move to cpu, convert to numpy arrays
output_batch = output_batch.data.cpu().numpy()
labels_batch = labels_batch.data.cpu().numpy()
# compute all metrics on this batch
summary_batch = {metric:metrics[metric](output_batch, labels_batch)
for metric in metrics}
summary_batch['loss'] = loss.data[0]
summ.append(summary_batch)
# update the average loss
loss_avg.update(loss.data[0])
t.set_postfix(loss='{:05.3f}'.format(loss_avg()))
t.update()
# compute mean of all metrics in summary
metrics_mean = {metric:np.mean([x[metric] for x in summ]) for metric in summ[0]}
metrics_string = " ; ".join("{}: {:05.3f}".format(k, v) for k, v in metrics_mean.items())
logging.info("- Train metrics: " + metrics_string)
def train_and_evaluate(model, train_dataloader, val_dataloader, optimizer,
loss_fn, metrics, params, model_dir, restore_file=None):
"""Train the model and evaluate every epoch.
Args:
model: (torch.nn.Module) the neural network
params: (Params) hyperparameters
model_dir: (string) directory containing config, weights and log
restore_file: (string) - name of file to restore from (without its extension .pth.tar)
"""
# reload weights from restore_file if specified
if restore_file is not None:
restore_path = os.path.join(args.model_dir, args.restore_file + '.pth.tar')
logging.info("Restoring parameters from {}".format(restore_path))
utils.load_checkpoint(restore_path, model, optimizer)
best_val_acc = 0.0
# learning rate schedulers for different models:
if params.model_version == "resnet18":
scheduler = StepLR(optimizer, step_size=150, gamma=0.1)
# for cnn models, num_epoch is always < 100, so it's intentionally not using scheduler here
elif params.model_version == "cnn":
scheduler = StepLR(optimizer, step_size=100, gamma=0.2)
for epoch in range(params.num_epochs):
scheduler.step()
# Run one epoch
logging.info("Epoch {}/{}".format(epoch + 1, params.num_epochs))
# compute number of batches in one epoch (one full pass over the training set)
train(model, optimizer, loss_fn, train_dataloader, metrics, params)
# Evaluate for one epoch on validation set
val_metrics = evaluate(model, loss_fn, val_dataloader, metrics, params)
val_acc = val_metrics['accuracy']
is_best = val_acc>=best_val_acc
# Save weights
utils.save_checkpoint({'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optim_dict' : optimizer.state_dict()},
is_best=is_best,
checkpoint=model_dir)
# If best_eval, best_save_path
if is_best:
logging.info("- Found new best accuracy")
best_val_acc = val_acc
# Save best val metrics in a json file in the model directory
best_json_path = os.path.join(model_dir, "metrics_val_best_weights.json")
utils.save_dict_to_json(val_metrics, best_json_path)
# Save latest val metrics in a json file in the model directory
last_json_path = os.path.join(model_dir, "metrics_val_last_weights.json")
utils.save_dict_to_json(val_metrics, last_json_path)
# Defining train_kd & train_and_evaluate_kd functions
def train_kd(model, teacher_model, optimizer, loss_fn_kd, dataloader, metrics, params):
"""Train the model on `num_steps` batches
Args:
model: (torch.nn.Module) the neural network
optimizer: (torch.optim) optimizer for parameters of model
loss_fn_kd:
dataloader:
metrics: (dict)
params: (Params) hyperparameters
"""
# set model to training mode
model.train()
teacher_model.eval()
# summary for current training loop and a running average object for loss
summ = []
loss_avg = utils.RunningAverage()
# Use tqdm for progress bar
with tqdm(total=len(dataloader)) as t:
for i, (train_batch, labels_batch) in enumerate(dataloader):
# move to GPU if available
if params.cuda:
train_batch, labels_batch = train_batch.cuda(async=True), \
labels_batch.cuda(async=True)
# convert to torch Variables
train_batch, labels_batch = Variable(train_batch), Variable(labels_batch)
# compute model output, fetch teacher output, and compute KD loss
output_batch = model(train_batch)
# get one batch output from teacher_outputs list
with torch.no_grad():
output_teacher_batch = teacher_model(train_batch)
if params.cuda:
output_teacher_batch = output_teacher_batch.cuda(async=True)
loss = loss_fn_kd(output_batch, labels_batch, output_teacher_batch, params)
# clear previous gradients, compute gradients of all variables wrt loss
optimizer.zero_grad()
loss.backward()
# performs updates using calculated gradients
optimizer.step()
# Evaluate summaries only once in a while
if i % params.save_summary_steps == 0:
# extract data from torch Variable, move to cpu, convert to numpy arrays
output_batch = output_batch.data.cpu().numpy()
labels_batch = labels_batch.data.cpu().numpy()
# compute all metrics on this batch
summary_batch = {metric:metrics[metric](output_batch, labels_batch)
for metric in metrics}
summary_batch['loss'] = loss.data[0]
summ.append(summary_batch)
# update the average loss
loss_avg.update(loss.data[0])
t.set_postfix(loss='{:05.3f}'.format(loss_avg()))
t.update()
# compute mean of all metrics in summary
metrics_mean = {metric:np.mean([x[metric] for x in summ]) for metric in summ[0]}
metrics_string = " ; ".join("{}: {:05.3f}".format(k, v) for k, v in metrics_mean.items())
logging.info("- Train metrics: " + metrics_string)
def train_and_evaluate_kd(model, teacher_model, train_dataloader, val_dataloader, optimizer,
loss_fn_kd, metrics, params, model_dir, restore_file=None):
"""Train the model and evaluate every epoch.
Args:
model: (torch.nn.Module) the neural network
params: (Params) hyperparameters
model_dir: (string) directory containing config, weights and log
restore_file: (string) - file to restore (without its extension .pth.tar)
"""
# reload weights from restore_file if specified
if restore_file is not None:
restore_path = os.path.join(args.model_dir, args.restore_file + '.pth.tar')
logging.info("Restoring parameters from {}".format(restore_path))
utils.load_checkpoint(restore_path, model, optimizer)
best_val_acc = 0.0
# Tensorboard logger setup
# board_logger = utils.Board_Logger(os.path.join(model_dir, 'board_logs'))
# learning rate schedulers for different models:
if params.model_version == "resnet18_distill":
scheduler = StepLR(optimizer, step_size=150, gamma=0.1)
# for cnn models, num_epoch is always < 100, so it's intentionally not using scheduler here
elif params.model_version == "cnn_distill":
scheduler = StepLR(optimizer, step_size=100, gamma=0.2)
for epoch in range(params.num_epochs):
scheduler.step()
# Run one epoch
logging.info("Epoch {}/{}".format(epoch + 1, params.num_epochs))
# compute number of batches in one epoch (one full pass over the training set)
train_kd(model, teacher_model, optimizer, loss_fn_kd, train_dataloader,
metrics, params)
# Evaluate for one epoch on validation set
val_metrics = evaluate_kd(model, val_dataloader, metrics, params)
val_acc = val_metrics['accuracy']
is_best = val_acc>=best_val_acc
# Save weights
utils.save_checkpoint({'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optim_dict' : optimizer.state_dict()},
is_best=is_best,
checkpoint=model_dir)
# If best_eval, best_save_path
if is_best:
logging.info("- Found new best accuracy")
best_val_acc = val_acc
# Save best val metrics in a json file in the model directory
best_json_path = os.path.join(model_dir, "metrics_val_best_weights.json")
utils.save_dict_to_json(val_metrics, best_json_path)
# Save latest val metrics in a json file in the model directory
last_json_path = os.path.join(model_dir, "metrics_val_last_weights.json")
utils.save_dict_to_json(val_metrics, last_json_path)
# #============ TensorBoard logging: uncomment below to turn in on ============#
# # (1) Log the scalar values
# info = {
# 'val accuracy': val_acc
# }
# for tag, value in info.items():
# board_logger.scalar_summary(tag, value, epoch+1)
# # (2) Log values and gradients of the parameters (histogram)
# for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# board_logger.histo_summary(tag, value.data.cpu().numpy(), epoch+1)
# # board_logger.histo_summary(tag+'/grad', value.grad.data.cpu().numpy(), epoch+1)
if __name__ == '__main__':
# Load the parameters from json file
args = parser.parse_args()
json_path = os.path.join(args.model_dir, 'params.json')
assert os.path.isfile(json_path), "No json configuration file found at {}".format(json_path)
params = utils.Params(json_path)
# use GPU if available
params.cuda = torch.cuda.is_available()
# Set the random seed for reproducible experiments
random.seed(230)
torch.manual_seed(230)
if params.cuda: torch.cuda.manual_seed(230)
# Set the logger
utils.set_logger(os.path.join(args.model_dir, 'train.log'))
# Create the input data pipeline
logging.info("Loading the datasets...")
# fetch dataloaders, considering full-set vs. sub-set scenarios
if params.subset_percent < 1.0:
train_dl = data_loader.fetch_subset_dataloader('train', params)
else:
train_dl = data_loader.fetch_dataloader('train', params)
dev_dl = data_loader.fetch_dataloader('dev', params)
logging.info("- done.")
"""Based on the model_version, determine model/optimizer and KD training mode
WideResNet and DenseNet were trained on multi-GPU; need to specify a dummy
nn.DataParallel module to correctly load the model parameters
"""
if "distill" in params.model_version:
# train a 5-layer CNN or a 18-layer ResNet with knowledge distillation
if params.model_version == "cnn_distill":
model = net.Net(params).cuda() if params.cuda else net.Net(params)
optimizer = optim.Adam(model.parameters(), lr=params.learning_rate)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = net.metrics
elif params.model_version == 'resnet18_distill':
model = resnet.ResNet18().cuda() if params.cuda else resnet.ResNet18()
optimizer = optim.SGD(model.parameters(), lr=params.learning_rate,
momentum=0.9, weight_decay=5e-4)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = resnet.metrics
"""
Specify the pre-trained teacher models for knowledge distillation
Important note: wrn/densenet/resnext/preresnet were pre-trained models using multi-GPU,
therefore need to call "nn.DaraParallel" to correctly load the model weights
Trying to run on CPU will then trigger errors (too time-consuming anyway)!
"""
if params.teacher == "resnet18":
teacher_model = resnet.ResNet18()
teacher_checkpoint = 'experiments/base_resnet18/best.pth.tar'
teacher_model = teacher_model.cuda() if params.cuda else teacher_model
elif params.teacher == "wrn":
teacher_model = wrn.WideResNet(depth=28, num_classes=10, widen_factor=10,
dropRate=0.3)
teacher_checkpoint = 'experiments/base_wrn/best.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "densenet":
teacher_model = densenet.DenseNet(depth=100, growthRate=12)
teacher_checkpoint = 'experiments/base_densenet/best.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "resnext29":
teacher_model = resnext.CifarResNeXt(cardinality=8, depth=29, num_classes=10)
teacher_checkpoint = 'experiments/base_resnext29/best.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "preresnet110":
teacher_model = preresnet.PreResNet(depth=110, num_classes=10)
teacher_checkpoint = 'experiments/base_preresnet110/best.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
utils.load_checkpoint(teacher_checkpoint, teacher_model)
# Train the model with KD
logging.info("Experiment - model version: {}".format(params.model_version))
logging.info("Starting training for {} epoch(s)".format(params.num_epochs))
logging.info("First, loading the teacher model and computing its outputs...")
train_and_evaluate_kd(model, teacher_model, train_dl, dev_dl, optimizer, loss_fn_kd,
metrics, params, args.model_dir, args.restore_file)
# non-KD mode: regular training of the baseline CNN or ResNet-18
else:
if params.model_version == "cnn":
model = net.Net(params).cuda() if params.cuda else net.Net(params)
optimizer = optim.Adam(model.parameters(), lr=params.learning_rate)
# fetch loss function and metrics
loss_fn = net.loss_fn
metrics = net.metrics
elif params.model_version == "resnet18":
model = resnet.ResNet18().cuda() if params.cuda else resnet.ResNet18()
optimizer = optim.SGD(model.parameters(), lr=params.learning_rate,
momentum=0.9, weight_decay=5e-4)
# fetch loss function and metrics
loss_fn = resnet.loss_fn
metrics = resnet.metrics
# elif params.model_version == "wrn":
# model = wrn.wrn(depth=28, num_classes=10, widen_factor=10, dropRate=0.3)
# model = model.cuda() if params.cuda else model
# optimizer = optim.SGD(model.parameters(), lr=params.learning_rate,
# momentum=0.9, weight_decay=5e-4)
# # fetch loss function and metrics
# loss_fn = wrn.loss_fn
# metrics = wrn.metrics
# Train the model
logging.info("Starting training for {} epoch(s)".format(params.num_epochs))
train_and_evaluate(model, train_dl, dev_dl, optimizer, loss_fn, metrics, params,
args.model_dir, args.restore_file)