-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
291 lines (260 loc) · 11.6 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
"""Train per-sample"""
import os
import argparse
import numpy as np
import scipy.io as sio
from copy import deepcopy
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from torch.optim.lr_scheduler import MultiStepLR
from torch.utils.data.sampler import WeightedRandomSampler
import torch.nn.functional as F
from algs import *
from config import *
from utils import *
def main():
parser = argparse.ArgumentParser()
# Basic settings
parser.add_argument('--dataset', type=str, required=True,
choices=['celeba', 'cifar10', 'cifar100', 'compas'])
parser.add_argument('--data_root', type=str)
parser.add_argument('--device', type=str)
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--save_file', type=str)
parser.add_argument('--load_file', type=str)
parser.add_argument('--load_warmup', type=str)
parser.add_argument('--n_val', type=int, help='Size of the validation set')
parser.add_argument('--download', default=False, action='store_true')
parser.add_argument('--verbose', default=False, action='store_true')
# Training settings
parser.add_argument('--alg', type=str, help='Training algorithm', required=True,
choices=['uniform', 'adalpboost', 'lpboost'])
parser.add_argument('--width', type=int, help='Width of Wide ResNet')
parser.add_argument('--epochs', type=int, help='Number of training epochs')
parser.add_argument('--iters_per_epoch', default=100, type=int)
parser.add_argument('--batch_size', type=int)
parser.add_argument('--lr', type=float)
parser.add_argument('--wd', type=float)
parser.add_argument('--scheduler', type=str)
parser.add_argument('--alpha', default=0.01, type=float)
parser.add_argument('--beta', type=float)
parser.add_argument('--eta', default=1.0, type=float)
parser.add_argument('--warmup', default=0, type=int)
parser.add_argument('--num_workers', type=int)
parser.add_argument('--pin_memory', action='store_true')
args = parser.parse_args()
populate_config(args.dataset, args)
print('Dataset: {}'.format(args.dataset))
print('Validation set size: {}'.format(args.n_val))
print('Training algorithm: {}'.format(args.alg))
print('Width: {}'.format(args.width))
print('Batch size: {}'.format(args.batch_size))
print('Epochs: {}'.format(args.epochs))
print('Iterations per epoch: {}'.format(args.iters_per_epoch))
print('Warmup epochs: {}'.format(args.warmup))
print('lr: {}'.format(args.lr))
print('wd: {}'.format(args.wd))
print('alpha: {}'.format(args.alpha))
print('beta: {}'.format(args.beta))
print('eta: {}'.format(args.eta))
device = args.device
if args.save_file is not None:
d = os.path.dirname(os.path.abspath(args.save_file))
if not os.path.isdir(d):
os.makedirs(d)
# Prepare dataset
dataset_train, dataset_test_train, dataset_valid, \
dataset_test, model, label_id = get_dataset(args)
# Build model
model = model.to(device)
# Fix seed for reproducibility
if args.seed is not None:
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.use_deterministic_algorithms(True)
cudnn.benchmark = False
else:
cudnn.benchmark = True
trainloader = DataLoader(dataset_train, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers,
pin_memory=args.pin_memory)
test_trainloader = DataLoader(dataset_test_train, batch_size=args.batch_size,
shuffle=False, num_workers=args.num_workers,
pin_memory=args.pin_memory) # only for test
testloader = DataLoader(dataset_test, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=args.pin_memory)
validloader = DataLoader(dataset_valid, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=args.pin_memory)
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr,
momentum=0.9, weight_decay=args.wd)
criterion = get_criterion(args.dataset)
scheduler = None
if args.scheduler is not None:
milestones = args.scheduler.split(',')
milestones = [int(s) for s in milestones]
print('scheduler: {}'.format(milestones))
scheduler = MultiStepLR(optimizer, milestones=milestones, gamma=0.1)
print('alpha: {}'.format(args.alpha))
print('eta: {}'.format(args.eta))
print('Num workers: {}'.format(args.num_workers))
print('Pin memory: {}'.format(args.pin_memory))
print('Seed: {}'.format(args.seed))
# Training
# 1. Warm up with ERM
warmup = False
if args.load_warmup is None:
for epoch in range(args.warmup):
print('=== Warmup (epoch={}) ==='.format(epoch + 1))
timed_run(erm, model, trainloader, optimizer,
criterion, None, device, 0)
timed_run(test, model, testloader, criterion, device, label_id)
warmup = True
warmup_state = {
'model': deepcopy(model.state_dict()),
'optimizer': deepcopy(optimizer.state_dict()),
}
else:
warmup = True
print('==> Loading warmup state from {}...'.format(args.load_warmup))
with open(args.load_warmup, 'rb') as f:
warmup_state = torch.load(f)
model.load_state_dict(warmup_state['model'])
optimizer.load_state_dict(warmup_state['optimizer'])
# 2. Boosting
val_avg_acc = []
val_avg_loss = []
val_correct = []
val_loss = []
test_avg_acc = []
test_avg_loss = []
test_correct = []
test_loss = []
train_avg_acc = []
train_avg_loss = []
train_correct = []
train_loss = []
train_sample_weights = []
obj_value = []
start_epoch = 0
sample_accuracy_history = np.zeros((0, len(dataset_train)), dtype=np.float)
sample_losses_history = np.zeros((0, len(dataset_train)), dtype=np.float)
sample_weights_history = np.zeros((0, len(dataset_train)), dtype=np.float)
if args.load_file is not None:
# Load mat file
print('==> Loading training history from {}...'.format(args.load_file))
mat = sio.loadmat(args.load_file)
num_epochs = mat['test_avg_acc'].shape[1]
test_avg_acc = [mat['test_avg_acc'][0,i] for i in range(num_epochs)]
test_avg_loss = [mat['test_avg_loss'][0,i] for i in range(num_epochs)]
test_correct = [mat['test_correct'][i,:] for i in range(num_epochs)]
test_loss = [mat['test_loss'][i,:] for i in range(num_epochs)]
val_avg_acc = [mat['val_avg_acc'][0,i] for i in range(num_epochs)]
val_avg_loss = [mat['val_avg_loss'][0,i] for i in range(num_epochs)]
val_correct = [mat['val_correct'][i,:] for i in range(num_epochs)]
val_loss = [mat['val_loss'][i,:] for i in range(num_epochs)]
train_avg_acc = [mat['train_avg_acc'][0,i] for i in range(num_epochs)]
train_avg_loss = [mat['train_avg_loss'][0,i] for i in range(num_epochs)]
train_correct = [mat['train_correct'][i,:] for i in range(num_epochs)]
train_loss = [mat['train_loss'][i,:] for i in range(num_epochs)]
train_sample_weights = [mat['train_sample_weights'][i,:] for i in range(num_epochs)]
sample_accuracy_history = mat['train_correct']
sample_losses_history = mat['train_loss']
sample_weights_history = mat['train_sample_weights']
start_epoch = num_epochs if warmup else num_epochs - 1
print('Starting from epoch {}.'.format(start_epoch + 1))
def test_epoch(epoch):
print('=== Validation (epoch={}) ==='.format(epoch))
a, b, c, d = timed_run(test, model, validloader, criterion, device, label_id)
val_avg_acc.append(a)
val_avg_loss.append(b)
val_correct.append(c)
val_loss.append(d)
print('=== Test (epoch={}) ==='.format(epoch))
a, b, c, d = timed_run(test, model, testloader, criterion, device, label_id)
test_avg_acc.append(a)
test_avg_loss.append(b)
test_correct.append(c)
test_loss.append(d)
print('=== Test over Train set (epoch={}) ==='.format(epoch))
a, b, c, d = timed_run(test, model, test_trainloader, criterion, device, label_id)
train_avg_acc.append(a)
train_avg_loss.append(b)
train_correct.append(c)
train_loss.append(d)
return c, d
if warmup and args.load_file is None:
print('==> Testing warmup model...')
test_epoch(0)
min_weighted_avg_acc = 1.0
for epoch in range(start_epoch, args.epochs):
print('===Train(epoch={})==='.format(epoch + 1))
sample_weights = get_sample_weights(args, sample_accuracy_history,
sample_losses_history, sample_weights_history,
obj_value, np.array(val_correct))
print('Weight max: {}'.format(sample_weights.max()))
train_sample_weights.append(sample_weights)
sample_weights_history = np.concatenate((sample_weights_history,
sample_weights.reshape(1, len(dataset_train))))
sweight = torch.tensor(sample_weights).to(device)
sampler = WeightedRandomSampler(sweight, args.iters_per_epoch * args.batch_size,
replacement=True)
trainloader = DataLoader(dataset_train, batch_size=args.batch_size,
sampler=sampler, num_workers=args.num_workers,
pin_memory=args.pin_memory)
timed_run(erm, model, trainloader, optimizer, criterion,
scheduler, device, args.iters_per_epoch)
sample_accuracy, sample_losses = test_epoch(epoch + 1)
sample_accuracy_history = np.concatenate(
(sample_accuracy_history, sample_accuracy.reshape(1, len(dataset_train))))
sample_losses_history = np.concatenate(
(sample_losses_history, sample_losses.reshape(1, len(dataset_train))))
weighted_avg_acc = sample_weights @ sample_accuracy
print('Weighted average accuracy: {}'.format(weighted_avg_acc))
min_weighted_avg_acc = min(min_weighted_avg_acc, weighted_avg_acc)
model.load_state_dict(warmup_state['model'])
optimizer.load_state_dict(warmup_state['optimizer'])
if scheduler is not None:
scheduler.last_epoch = 0
# Final result
print('==> Training complete.')
print('Min Weighted Average Acc: {}'.format(min_weighted_avg_acc))
# Save the results
if args.save_file is not None:
print('==>Saving training history to {}...'.format(args.save_file))
mat = {
'test_avg_acc': np.array(test_avg_acc),
'test_avg_loss': np.array(test_avg_loss),
'test_correct': np.array(test_correct),
'test_loss': np.array(test_loss),
'val_avg_acc': np.array(val_avg_acc),
'val_avg_loss': np.array(val_avg_loss),
'val_correct': np.array(val_correct),
'val_loss': np.array(val_loss),
'train_avg_acc': np.array(train_avg_acc),
'train_avg_loss': np.array(train_avg_loss),
'train_correct': np.array(train_correct),
'train_loss': np.array(train_loss),
'train_sample_weights': np.array(train_sample_weights),
}
if len(obj_value) > 0:
mat['obj_value'] = np.array(obj_value)
sio.savemat(args.save_file, mat)
print('Done.')
def get_sample_weights(args, sample_accuracy_history,
sample_losses_history, sample_weights_history,
obj_value, val_history=None):
# Input shape: each history - (epoch, num_samples)
# Output: group_weights size: (num_samples,)
if len(sample_accuracy_history) == 0:
return uniform(sample_losses_history)
if args.alg == 'uniform':
return uniform(sample_losses_history)
elif args.alg == 'adalpboost':
return adalpboost(sample_accuracy_history, sample_weights_history, args.eta)
elif args.alg == 'lpboost':
return lpboost(sample_accuracy_history, obj_value, args.alpha, args.beta, args.verbose)
else:
raise NotImplementedError
if __name__ == '__main__':
main()