-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
331 lines (281 loc) · 13.9 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
import os
import glob
import random
import numpy as np
import torch
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torchvision.transforms as transforms
from torch.utils.data.sampler import Sampler
import torch.backends.cudnn as cudnn
from torch.utils.tensorboard import SummaryWriter
from torch.autograd import Variable
from tqdm import tqdm
from shutil import copyfile
from datetime import datetime
from data_loader import foodSpaceLoader, error_catching_loader, my_collate, collate_fn
from args import get_parser
from models import FoodSpaceNet as FoodSpaceNet
from utils import PadToSquareResize, AverageMeter, worker_init_fn
from loss_func import calculate_loss
from eval import validate
# =============================================================================
parser = get_parser()
opts = parser.parse_args()
opts.gpu = list(map(int, opts.gpu.split(',')))
print('Using GPU(s): ' + ','.join([str(x) for x in opts.gpu]))
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join([str(x) for x in opts.gpu])
print(torch.cuda.is_available())
# =============================================================================
ranking_loss = torch.nn.MarginRankingLoss(margin=0.3)
class SubsetSequentialSampler(Sampler):
r"""Samples elements randomly from a given list of indices, without replacement.
Arguments:
indices (sequence): a sequence of indices
"""
def __init__(self, indices):
self.indices = indices
def __iter__(self):
return (self.indices[i] for i in range(len(self.indices)))
def __len__(self):
return len(self.indices)
def main():
random.seed(opts.seed)
torch.manual_seed(opts.seed)
torch.cuda.manual_seed(opts.seed)
torch.cuda.manual_seed_all(opts.seed)
print(torch.cuda.is_available())
print(torch.cuda.current_device())
print(torch.cuda.get_device_name(0))
iter = 0
# Track results on tensorboard
timestamp = datetime.now().strftime("%Y%m%d-%H%M%S")
print('Saving to: ' + timestamp)
logdir = "tensorboard/" + timestamp + '__' + os.path.basename(__file__).replace('.py', '') + "/"
logdir = os.path.join(os.path.dirname(__file__), logdir)
modeldir = logdir + "models/"
opts.snapshots = modeldir
if not os.path.isdir("tensorboard/"): os.mkdir("tensorboard/")
if not os.path.isdir(logdir): os.mkdir(logdir)
if not os.path.isdir(modeldir): os.mkdir(modeldir)
copyfile('models.py', os.path.join(logdir,'models.py'))
train_writer = SummaryWriter(logdir + '/train')
train_writer_text = open(os.path.join(logdir, 'train.txt'), 'a')
test_writer_text = open(os.path.join(logdir, 'test.txt'), 'a')
with open(os.path.join(logdir, 'opts.txt'), 'w') as file:
for arg in vars(opts):
file.write(arg + ': ' + str(getattr(opts, arg)) + '\n')
model = FoodSpaceNet(opts)
model.cuda()
# data preparation, loaders
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# preparing the training loader
train_data = foodSpaceLoader(opts.img_path,
transforms.Compose([
transforms.RandomChoice([
PadToSquareResize(resize=256, padding_mode='random'),
transforms.Resize((256, 256))]),
transforms.RandomRotation(10),
transforms.RandomCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize]),
data_path=opts.data_path,
partition='train',
maxInsts=opts.maxInsts)
print('Training loader prepared.')
# preparing the valitadion loader
val_data = foodSpaceLoader(opts.img_path,
transforms.Compose([
PadToSquareResize(resize=256, padding_mode='reflect'),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize]),
data_path=opts.data_path,
partition='val',
loader=error_catching_loader)
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=opts.batch_size, shuffle=False,
# sampler=SubsetSequentialSampler(np.arange(1000)),
num_workers=opts.workers, pin_memory=True)
print('Validation loader prepared.')
# creating different parameter groups
# model.instNet_.doc_encoder.atten_layer_sent.u
# model.instNet_.doc_encoder.sent_encoder.atten_layer.u
# model.attentionVisual.w_context_vector
# vision_params = list(map(id, model.visionMLP.parameters()))
vision_params = [kv[1] for kv in model.named_parameters() if kv[0].split('.')[0] in ['visionMLP']]
attention_params = [kv[1] for kv in model.named_parameters() if 'atten' in kv[0]]
# embs_params = list(map(id, model.embs.parameters()))
# base_params = filter(lambda p: id(p) not in vision_params+embs_params, model.parameters())
base_params = [kv[1] for kv in model.named_parameters() if (kv[0].split('.')[0] not in ['visionMLP','embs'] and 'atten' not in kv[0])]
optimizer = torch.optim.Adam([
{'params': base_params},
{'params': vision_params, 'lr': opts.lr*opts.freeVision },
{'params': model.embs.parameters(), 'lr': opts.lr*opts.freeEmbs },
{'params': attention_params, 'lr': 0 }
], lr=opts.lr*opts.freeRecipe)
if len(opts.gpu) > 1:
model = torch.nn.DataParallel(model)
if opts.resume:
if os.path.isfile(opts.resume):
print("=> loading checkpoint '{}'".format(opts.resume))
checkpoint = torch.load(opts.resume)
opts.start_epoch = checkpoint['epoch']
best_val = checkpoint['best_val']
best_rec = checkpoint['best_val']
model.load_state_dict(checkpoint['state_dict'])
optimizer = checkpoint['optimizer']
iter = len(train_data)/opts.batch_size * opts.start_epoch
print("=> loaded checkpoint '{}' (epoch {})"
.format(opts.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(opts.resume))
best_val = float('inf')
best_rec = 0
else:
best_val = float('inf')
best_rec = 0
since_improvement = 0
print('There are %d parameter groups' % len(optimizer.param_groups))
print('\tInitial base params lr: %f' % optimizer.param_groups[0]['lr'])
print('\tInitial vision params lr: %f' % optimizer.param_groups[1]['lr'])
print('\tInitial word2vec params lr: %f' % optimizer.param_groups[2]['lr'])
cudnn.benchmark = True
warmup_lr = np.linspace(opts.lr * 0.01, opts.lr, opts.warmup)
for epoch in range(opts.start_epoch, opts.epochs):
print('Started epoch {}/{}'.format(epoch+1, opts.epochs))
inds = np.arange(len(train_data)).tolist()
random.shuffle(inds)
inds = np.asanyarray(inds)
print('\tStarted training...')
# preparing the training loader
train_loader = torch.utils.data.DataLoader(train_data, batch_size=opts.batch_size,
shuffle=False,
sampler=SubsetSequentialSampler(inds),
num_workers=opts.workers,
pin_memory=True,
worker_init_fn=worker_init_fn,
collate_fn=my_collate)
if epoch < opts.warmup and opts.warmup >= 2:
opts.lr = warmup_lr[epoch]
optimizer.param_groups[0]['lr'] = opts.lr * opts.freeRecipe
optimizer.param_groups[1]['lr'] = opts.lr * opts.freeVision
optimizer.param_groups[2]['lr'] = opts.lr * opts.freeEmbs
iter = train(train_loader, train_data, model, optimizer, epoch, iter, train_writer, train_writer_text)
if str(epoch + 1) in opts.patience.split(','):
state = {'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_val': best_val,
'val_medR': 0,
'optimizer': optimizer,
'since_improvement': since_improvement,
'freeVision': opts.freeVision,
'curr_val': 0,
'curr_recall': 0}
filename = opts.snapshots + 'model_preFreeWeigths_epoch{}.pth.tar'.format(state['epoch'])
torch.save(state, filename)
opts.freeVision = True
opts.freeRecipe = True
opts.freeEmbs = True
optimizer.param_groups[0]['lr'] = opts.lr * opts.freeRecipe
optimizer.param_groups[1]['lr'] = opts.lr * opts.freeVision
optimizer.param_groups[2]['lr'] = opts.lr * opts.freeEmbs
optimizer.param_groups[3]['lr'] = opts.lr
if str(epoch + 1) in opts.patienceLR.split(','):
state = {'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_val': best_val,
'val_medR': 0,
'optimizer': optimizer,
'since_improvement': since_improvement,
'freeVision': opts.freeVision,
'curr_val': 0,
'curr_recall': 0}
filename = opts.snapshots + 'model_preLR_{}-{}_epoch{}.pth.tar'.format(opts.lr,opts.lr*0.1, state['epoch'])
torch.save(state, filename)
opts.lr = opts.lr * 0.1
optimizer.param_groups[0]['lr'] = opts.lr * opts.freeRecipe
optimizer.param_groups[1]['lr'] = opts.lr * opts.freeVision
optimizer.param_groups[2]['lr'] = opts.lr * opts.freeEmbs
optimizer.param_groups[3]['lr'] = opts.lr
if (epoch+1) % opts.valfreq == 0:
val_medR, val_recall, val_meanR, val_meanDCG = validate(val_loader, model, opts)
if val_medR > best_val:
since_improvement += 1
else:
since_improvement = 0
cum_recall = sum([v for k, v in val_recall.items()])
state = {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_val': best_val,
'val_medR': val_medR,
'optimizer': optimizer,
'since_improvement': since_improvement,
'freeVision': opts.freeVision,
'curr_val': val_medR,
'curr_recall': val_recall
}
filename = opts.snapshots + 'model.pth.tar'
torch.save(state, filename)
if val_medR < best_val:
if glob.glob(opts.snapshots + 'model_BEST_VAL*'): os.remove(glob.glob(opts.snapshots + 'model_BEST_VAL*')[0])
filename = opts.snapshots + 'model_BEST_VAL_e%03d_v-%.3f_cr-%.4f.pth.tar' % (
state['epoch'], state['val_medR'], cum_recall)
torch.save(state, filename)
if cum_recall > best_rec:
if glob.glob(opts.snapshots + 'model_BEST_REC*'): os.remove(glob.glob(opts.snapshots + 'model_BEST_REC*')[0])
filename = opts.snapshots + 'model_BEST_REC_e%03d_v-%.3f_cr-%.4f.pth.tar' % (
state['epoch'], state['val_medR'], cum_recall)
torch.save(state, filename)
best_val = min(val_medR, best_val)
best_rec = max(cum_recall, best_rec)
train_writer.add_scalar("medR", val_medR, epoch)
train_writer.add_scalar("meanR", val_meanR, epoch)
train_writer.add_scalar("meanDCGR", val_meanDCG, epoch)
train_writer.add_scalar("recall@1", val_recall[1], epoch)
train_writer.add_scalar("recall@5", val_recall[5], epoch)
train_writer.add_scalar("recall@10", val_recall[10], epoch)
train_writer.add_scalar("recall@1-5-10", val_recall[10]+val_recall[5]+val_recall[1], epoch)
train_writer.flush()
test_writer_text.write(str(iter)+','+str(val_medR)+'\n')
test_writer_text.flush()
print('\t** Validation: %f (best) - %d (since_improvement)' % (best_val, since_improvement))
def train(train_loader, train_data, model, optimizer, epoch, iter, train_writer, train_writer_text):
losses = AverageMeter()
model.train()
for i, (input, rec_ids) in enumerate(tqdm(train_loader, total=len(train_loader))):
iter += 1
output = model(input, opts)
loss = calculate_loss(output, opts)
with torch.no_grad():
losses.update(loss.item(), input[0].size(0))
if i % 100 == 0:
train_writer.add_scalar("cos_loss", loss.item(), iter)
train_writer.flush()
train_writer_text.write(str(iter)+','+str(loss.item())+'\n')
train_writer_text.flush()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# if i==10:
# break
del loss
print('\t\tSubepoch: {0}\t'
'cos_loss {loss.val:.4f} ({loss.avg:.4f})\t'
'vision ({visionLR}) - recipe ({recipeLR}) - word2vec ({embsLR}) - attention ({attLR})\t'.format(
epoch+1, loss=losses,
visionLR=optimizer.param_groups[1]['lr'],
recipeLR=optimizer.param_groups[0]['lr'],
embsLR=optimizer.param_groups[2]['lr'],
attLR=optimizer.param_groups[3]['lr']))
return iter
def save_checkpoint(state):
filename = opts.snapshots + 'model.pth.tar'
torch.save(state, filename)
if __name__ == '__main__':
main()