-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
237 lines (183 loc) · 8.7 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
import sys
from datetime import datetime
import torch
from torch.utils import data
from torch.utils.data import dataloader
from tqdm import tqdm
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from utils import *
from dataloader import *
from model import *
from config import *
'''
training and validation for one epoch
'''
def train_valid_one_epoch(model, loaders, writers, criterion, optimizer, steps):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
ngram = 4
kpi = {
phase: {
'loss': [],
'bleu': { '{}-gram'.format(i): 0
for i in range(1, 1+ngram) } # generate dict for each n_gram
}
for phase in ['train', 'valid']
}
for phase in ['train', 'valid']: # iterate for training and validation phases
if phase == 'train':
model.train()
elif loaders['valid']:
model.eval()
else: # if validation loader is not defined
continue
# iterate over batches
for idx, batch in tqdm(
enumerate(loaders[phase]), total=len(loaders[phase]), desc=phase
):
(images, captions, _) = batch
# images shape [batch, 3, 224, 224]
# captions shape [max sentence length, batch, 5]
max_length, batch_size, _ = captions.shape
captions = captions.reshape(max_length, batch_size*5)
images, captions = images.to(device), captions.to(device) # move itmes to gpu
images = torch.repeat_interleave(images, repeats=5, dim=0) # [a, b, c] -> [aaaaa, bbbbb, ccccc].T
# images shape [batch*5, 3, 224, 224]
# captions shape [max sentence length, batch*5]
with torch.set_grad_enabled(phase=='train'):
outputs = model(images, captions[:-1])
loss = criterion(outputs.reshape(-1, outputs.shape[2]), captions.reshape(-1)) # calculate loss
kpi[phase]['loss'].append(loss.item()) # register this step loss
writers[phase].add_scalar("steps loss", loss.item(), global_step=steps[phase])
bleu_dict = bleu_score_(model, batch, loaders['train'].dataset)
for n_gram in bleu_dict:
kpi[phase]['bleu'][n_gram] += bleu_dict[n_gram]
if phase == 'train':
model.train()
optimizer.zero_grad()
loss.backward()
optimizer.step()
steps[phase] += 1
del images, captions, loss
torch.cuda.empty_cache()
return kpi, steps
'''
training and validation for numerous of epochs
'''
def train_valid_epochs(model, loaders, writers, num_epochs, criterion, optimizer, scheduler, steps, run_path, prev_valid_loss=float('inf'), start_epoch=0):
model_save_path = '{}/models/model_checkpoint.pth'.format(run_path)
best_valid_loss = prev_valid_loss
start_epoch += 1
end_epoch = start_epoch + num_epochs
# iterate over the epochs
for epoch in range(start_epoch, end_epoch):
print('Epoch {:3d} of {}:'.format(epoch, end_epoch-1), flush=True)
# train and validation for one epoch
kpi, steps = train_valid_one_epoch(model, loaders, writers, criterion, optimizer, steps)
# pretty printing training and validation results
print_str = ''
for phase in ['train', 'valid']:
if loaders[phase]:
loss = sum(kpi[phase]['loss']) / len(loaders[phase])
print_str += '{}:\tloss={:.5f}'.format(phase, loss)
for n_gram in kpi[phase]['bleu']:
bleu = kpi[phase]['bleu'][n_gram] / len(loaders[phase])
print_str += '\t{}={:.5f}'.format(n_gram, bleu)
writers[phase].add_scalar(n_gram, bleu, epoch)
print_str += '\n'
writers[phase].add_scalar('loss', loss, epoch)
epoch_val_loss = sum(kpi['valid']['loss']) / len(loaders['valid'])
if scheduler is not None:
try:
lr = scheduler.get_last_lr()[0]
except:
lr = [ group['lr'] for group in optimizer.param_groups ][0]
writers['train'].add_scalar('lr', lr, epoch)
scheduler.step(epoch_val_loss)
# if validation loss is better, save model chechpoint
if epoch_val_loss < best_valid_loss:
best_valid_loss = epoch_val_loss
checkpoint = {
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'train_step': steps['train'],
'valid_step': steps['valid'],
'epoch': epoch,
}
save_checkpoint(checkpoint, model_save_path)
print_examples(model, loaders['train'].dataset)
print(print_str)
def test(model, loader, criterion, run_path):
save_path = '{}/test/test.csv'.format(run_path)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
ngram = 4
df = pd.DataFrame(columns=['image', 'prediction', 'loss', *['{}-gram'.format(i) for i in range(1, 1+ngram)] ])
model.eval()
with torch.no_grad():
# iterate over batches
for idx, batch in tqdm(
enumerate(loader), total=len(loader), desc='testing'
):
(image, caption, img_id) = batch
# images shape [batch, 3, 224, 224]
# captions shape [max sentence length, batch, 5]
max_length, batch_size, _ = caption.shape
caption = caption.reshape(max_length, batch_size*5)
image, caption = image.to(device), caption.to(device) # move itmes to gpu
words = ' '.join(model.caption(image, loader.dataset.vocab)) # caption before multiple the images to 5 each one
image = torch.repeat_interleave(image, repeats=5, dim=0) # [a, b, c] -> [aaaaa, bbbbb, ccccc].T
# images shape [batch*5, 3, 224, 224]
# captions shape [max sentence length, batch*5]
outputs = model(image, caption[:-1])
loss = criterion(outputs.reshape(-1, outputs.shape[2]), caption.reshape(-1)).cpu().detach().item() # calculate loss
bleu_dict = bleu_score_(model, batch, loader.dataset)
df = df.append({ 'image': img_id[0], 'prediction': words, 'loss': loss, **bleu_dict }, ignore_index=True)
del image, caption, loss, words
torch.cuda.empty_cache()
df. to_csv(save_path)
'''
train model by a predefined configuration
'''
def train():
np.random.seed(42)
torch.manual_seed(42)
# generate running environment
datetime_srt = datetime.today().strftime("%d-%m-%y_%H:%M")
run_path = os.path.join(sys.path[0], 'runs', datetime_srt)
print('Generating running environment')
create_env(run_path)
torch.backends.cudnn.benchmark = True
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# generate training and validation dataloaders
print('Generating loaders')
loaders = get_loaders(train_size=CFG.train_size, batch_size=CFG.batch_size)
# update vocabulary size in config file
CFG.vocab_size = len(loaders['train'].dataset.vocab)
CFG.save(run_path)
# generate training and validation writers
print('Generating writers')
writers = get_writers(run_path, CFG.model_name)
steps = {
'train': 0,
'valid': 0
}
# initialize model, loss, optimizer, scheduler
print('Generating model')
model = EncoderDecoder(CFG.embed_size, CFG.hidden_size, CFG.vocab_size, CFG.lstm_num_layers, pretrained=CFG.pretrained, train_backbone=CFG.train_backbone, drop_prob=CFG.drop_rate).to(device)
criterion = CFG.criterion(ignore_index=loaders['train'].dataset.vocab.stoi['<PAD>'], **CFG.criterion_dict)
optimizer = CFG.optimizer(model.parameters(), **CFG.optimizer_dict)
scheduler = CFG.scheduler(optimizer, **CFG.scheduler_dict) if CFG.scheduler else None
start_epoch = 0
if CFG.load_model:
steps, end_epoch = load_checkpoint(torch.load(CFG.model_path), model, optimizer)
start_epoch = end_epoch + 1
train_valid_epochs(model, loaders, writers, CFG.num_epochs, criterion, optimizer, scheduler, steps, run_path, start_epoch=start_epoch)
test(model, loaders['test'], criterion, run_path)
if __name__ == "__main__":
np.random.seed(CFG.seed)
torch.manual_seed(CFG.seed)
torch.cuda.manual_seed(CFG.seed)
train()