-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
executable file
·393 lines (359 loc) · 15.6 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
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
import os
import shutil
import sys
import threading
import curses
import gc
import time
from random import choices
from itertools import chain
import numpy as np
import pandas as pd
import sklearn
import cv2
from tqdm import tqdm as T
from iterstrat.ml_stratifiers import MultilabelStratifiedKFold
from apex import amp
import torch, torchvision
from torchsummary import summary
from torch.utils.tensorboard import SummaryWriter
from torch import optim
from torch.nn import functional as F
from torch.utils.data import Dataset,DataLoader
from BanglaDataset import BanglaDataset
from utils import *
from metrics import *
from optimizers import Over9000
from augmentations.augmix import RandomAugMix
from augmentations.gridmask import GridMask
from model.seresnext import seresnext
from model.effnet import EfficientNetWrapper
from model.densenet import *
## This library is for augmentations .
from albumentations import (
PadIfNeeded,
HorizontalFlip,
VerticalFlip,
CenterCrop,
Crop,
Compose,
Transpose,
RandomRotate90,
ElasticTransform,
GridDistortion,
OpticalDistortion,
RandomSizedCrop,
Resize,
CenterCrop,
OneOf,
CLAHE,
RandomBrightnessContrast,
Cutout,
RandomGamma,
ShiftScaleRotate ,
GaussNoise,
Blur,
MotionBlur,
GaussianBlur,
Normalize,
)
n_fold = 5
fold = 1
SEED = 24
batch_size = 32
sz = 128
learning_rate = 1.25e-3
patience = 5
opts = ['normal', 'mixup', 'cutmix']
device = 'cuda:0'
apex = False
pretrained_model = 'se_resnext50_32x4d'
# pretrained_model = 'densenet121'
# pretrained_model = 'efficientnet-b4'
model_name = '{}_trial_stage1_fold_{}'.format(pretrained_model, fold)
model_dir = 'model_dir'
history_dir = 'history_dir'
tb_dir = 'runs_seresnext50'
imagenet_stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
load_model = True
history = pd.DataFrame()
prev_epoch_num = 0
n_epochs = 210
valid_recall = 0.0
best_valid_recall = 0.0
best_valid_loss = np.inf
np.random.seed(SEED)
os.makedirs(model_dir, exist_ok=True)
os.makedirs(history_dir, exist_ok=True)
if os.path.exists(tb_dir):
try:
shutil.rmtree(tb_dir)
except OSError as e:
print("Error: {} : {}".format(tb_dir, e.strerror))
def launchTensorBoard():
os.system('tensorboard --logdir ./ --port 9999 --host 0.0.0.0')
return
try:
t = threading.Thread(target=launchTensorBoard, args=([]))
t.start()
except:
pass
writer = SummaryWriter(tb_dir)
train_aug =Compose([
ShiftScaleRotate(p=0.9,border_mode= cv2.BORDER_CONSTANT, value=[0, 0, 0], scale_limit=0.25),
OneOf([
Cutout(p=0.3, max_h_size=sz//16, max_w_size=sz//16, num_holes=10, fill_value=0),
GridMask(num_grid=7, p=0.7, fill_value=0)
], p=0.20),
RandomAugMix(severity=1, width=1, alpha=1., p=0.05),
# OneOf([
# ElasticTransform(p=0.1, alpha=1, sigma=50, alpha_affine=30,border_mode=cv2.BORDER_CONSTANT,value =0),
# GridDistortion(distort_limit =0.05 ,border_mode=cv2.BORDER_CONSTANT,value =0, p=0.1),
# OpticalDistortion(p=0.1, distort_limit= 0.05, shift_limit=0.2,border_mode=cv2.BORDER_CONSTANT,value =0)
# ], p=0.3),
OneOf([
GaussNoise(var_limit=0.01),
Blur(),
GaussianBlur(blur_limit=3),
RandomGamma(p=0.8),
], p=0.4)
# Normalize()
]
)
# Normalize([0.0692], [0.2051])]
# val_aug = Compose([Normalize([0.0692], [0.2051])])
val_aug = Compose([Normalize()])
train_df = pd.read_csv('data/train.csv')
# train_pseudo_df = pd.read_csv('data/train_and_pseudo.csv')
nunique = list(train_df.nunique())[1:-1]
train_df['id'] = train_df['image_id'].apply(lambda x: int(x.split('_')[1]))
# train_pseudo_df['id'] = train_pseudo_df['image_id'].apply(lambda x: int(x.split('_')[1]))
X, y = train_df[['id', 'grapheme_root', 'vowel_diacritic', 'consonant_diacritic']].values[:,0], train_df.values[:,1:]
train_df['fold'] = np.nan
train_df= train_df.sample(frac=1, random_state=SEED).reset_index(drop=True)
#split data
mskf = MultilabelStratifiedKFold(n_splits=n_fold, random_state=SEED)
for i, (_, test_index) in enumerate(mskf.split(X, y)):
train_df.iloc[test_index, -1] = i
train_df['fold'] = train_df['fold'].astype('int')
idxs = [i for i in range(len(train_df))]
train_idx = []
val_idx = []
model = seresnext(nunique, pretrained_model).to(device)
# model = Dnet(nunique).to(device)
# model = EfficientNetWrapper(pretrained_model).to(device)
# print(summary(model, (3, 128, 128)))
writer.add_graph(model, torch.FloatTensor(np.random.randn(1, 1, 137, 236)).cuda())
# writer.close()
# For stratified split
for i in T(range(len(train_df))):
if train_df.iloc[i]['fold'] == fold: val_idx.append(i)
else: train_idx.append(i)
# train_idx = idxs[:int((n_fold-1)*len(idxs)/(n_fold))]
# train_idx = np.load('train_pseudo_idxs.npy')
# val_idx = idxs[int((n_fold-1)*len(idxs)/(n_fold)):]
train_ds = BanglaDataset(train_df, 'data/numpy_format', train_idx, aug=train_aug)
train_loader = DataLoader(train_ds,batch_size=batch_size, shuffle=True)
valid_ds = BanglaDataset(train_df, 'data/numpy_format', val_idx, aug=None)
valid_loader = DataLoader(valid_ds, batch_size=batch_size, shuffle=True)
writer = SummaryWriter(tb_dir)
## This function for train is copied from @hanjoonchoe
## We are going to train and track accuracy and then evaluate and track validation accuracy
def train(epoch,history):
t1 = time.time()
model.train()
losses = []
accs = []
acc= 0.0
total = 0.0
running_loss = 0.0
grapheme_root_out=0.0
vowel_diacritic_out=0.0
consonant_diacritic_out=0.0
running_acc = 0.0
running_recall = 0.0
rate = 1
if epoch<30:
rate = 1
elif epoch>=30 and rate>0.65:
rate = np.exp(-(epoch-30)/60)
else:
rate = 0.65
for idx, (inputs,labels1,labels2,labels3) in enumerate(train_loader):
inputs = inputs.to(device)
labels1 = labels1.to(device)
labels2 = labels2.to(device)
labels3 = labels3.to(device)
total += len(inputs)
choice = choices(opts, weights=[0.20, 0.30, 0.50])
# print(torch.max())
# denormalize = UnNormalize(*imagenet_stats)
# print(torch.max(denormalize(inputs)))
writer.add_images('my_image', inputs, 0)
optimizer.zero_grad()
if choice[0] == 'normal':
outputs1,outputs2,outputs3 = model(inputs.float())
loss1 = 0.7*criterion(outputs1,labels1)
loss2 = 0.20* criterion(outputs2,labels2)
loss3 = 0.10*criterion(outputs3,labels3)
loss = loss1 + loss2 + loss3
running_loss += loss.item()
elif choice[0] == 'mixup':
inputs, targets = mixup(inputs, labels1, labels2, labels3, np.random.uniform(0.8, 1.0))
outputs1, outputs2, outputs3 = model(inputs.float())
loss1, loss2, loss3 = mixup_criterion(outputs1,outputs2,outputs3, targets, rate=rate)
loss = 0.7*loss1 + 0.20*loss2 + 0.10*loss3
running_loss += loss.item()
elif choice[0] == 'cutmix':
inputs, targets = cutmix(inputs, labels1, labels2, labels3, np.random.uniform(0.8, 1.0))
outputs1, outputs2, outputs3 = model(inputs.float())
loss1, loss2, loss3 = cutmix_criterion(outputs1,outputs2,outputs3, targets, rate=rate)
loss = 0.7*loss1 + 0.20*loss2 + 0.10*loss3
running_loss += loss.item()
grapheme_root_out += (outputs1.argmax(1)==labels1).float().mean()
vowel_diacritic_out += (outputs2.argmax(1)==labels2).float().mean()
consonant_diacritic_out += (outputs3.argmax(1)==labels3).float().mean()
if apex:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
optimizer.zero_grad()
acc = running_acc/total
# scheduler.step()
elapsed = int(time.time() - t1)
eta = int(elapsed / (idx+1) * (len(train_loader)-(idx+1)))
lr = None
for param_group in optimizer.param_groups:
lr = param_group['lr']
writer.add_scalar('Learning Rate', lr, epoch*len(train_loader)+idx)
writer.add_scalar('OHEM Rate', rate, epoch)
writer.add_scalar('Loss/train', running_loss/(idx+1), epoch*len(train_loader)+idx)
writer.add_scalar('Train Accuracy/Root', grapheme_root_out/(idx+1), epoch*len(train_loader)+idx)
writer.add_scalar('Train Accuracy/Vowel', vowel_diacritic_out/(idx+1), epoch*len(train_loader)+idx)
writer.add_scalar('Train Accuracy/Consonant', consonant_diacritic_out/(idx+1), epoch*len(train_loader)+idx)
if idx%1==0:
msg = 'Epoch: {} \t Progress: {}/{} \t Loss: {:.4f} \t Time: {}s \t ETA: {}s'.format(epoch,
idx, len(train_loader), running_loss/(idx+1), elapsed, eta)
print(msg, end='\r')
# \nAcc: \t Root {:.4f} \t Vowel {:.4f} \t Consonant {:.4f}
# , grapheme_root_out/(idx+1), vowel_diacritic_out/(idx+1), consonant_diacritic_out/(idx+1)
# stdscr.addstr(0, 0, msg)
# stdscr.refresh()
losses.append(running_loss/len(train_loader))
# accs.append(running_acc/(len(train_loader)*3))
total_train_recall = running_recall/len(train_loader)
torch.cuda.empty_cache()
gc.collect()
history.loc[epoch, 'train_loss'] = losses[0]
history.loc[epoch, 'Time'] = elapsed
return total_train_recall
def evaluate(epoch,history):
model.eval()
total = 0.0
running_loss = 0.0
running_acc = 0.0
grapheme_root_out=0.0
vowel_diacritic_out=0.0
consonant_diacritic_out=0.0
running_acc = 0.0
pred1= []
pred2= []
pred3 = []
lab1 = []
lab2 = []
lab3 = []
with torch.no_grad():
for idx, (inputs,labels1,labels2,labels3) in T(enumerate(valid_loader),total=len(valid_loader)):
inputs = inputs.to(device)
labels1 = labels1.to(device)
labels2 = labels2.to(device)
labels3 = labels3.to(device)
total += len(inputs)
outputs1,outputs2,outputs3 = model(inputs.float())
pred1.extend(torch.argmax(outputs1, dim=1).cpu().numpy())
pred2.extend(torch.argmax(outputs2, dim=1).cpu().numpy())
pred3.extend(torch.argmax(outputs3, dim=1).cpu().numpy())
lab1.extend(labels1.cpu().numpy())
lab2.extend(labels2.cpu().numpy())
lab3.extend(labels3.cpu().numpy())
loss1 = 0.70*criterion(outputs1,labels1)
loss2 = 0.20*criterion(outputs2,labels2)
loss3 = 0.10*criterion(outputs3,labels3)
running_loss += loss1.item()+loss2.item()+loss3.item()
grapheme_root_out += (outputs1.argmax(1)==labels1).float().mean()
vowel_diacritic_out += (outputs2.argmax(1)==labels2).float().mean()
consonant_diacritic_out += (outputs3.argmax(1)==labels3).float().mean()
recall_graph = sklearn.metrics.recall_score(pred1, lab1, average='macro')
recall_vowel = sklearn.metrics.recall_score(pred2, lab2, average='macro')
recall_consonant = sklearn.metrics.recall_score(pred3, lab3, average='macro')
scores = [recall_graph, recall_vowel, recall_consonant]
total_recall = np.average(scores, weights=[2, 1, 1])
writer.add_scalar('Loss/val', running_loss/(len(valid_loader)), epoch)
writer.add_scalar('Val Accuracy/Root', grapheme_root_out/(len(valid_loader)), epoch)
writer.add_scalar('Val Accuracy/Vowel', vowel_diacritic_out/(len(valid_loader)), epoch)
writer.add_scalar('Val Accuracy/Consonant', consonant_diacritic_out/(len(valid_loader)), epoch)
writer.add_scalar('Val Recall/Root', recall_graph, epoch)
writer.add_scalar('Val Recall/Vowel', recall_vowel, epoch)
writer.add_scalar('Val Recall/Consonant', recall_consonant, epoch)
writer.add_scalar('Val Recall/Total', total_recall, epoch)
msg = 'Loss: {:.4f} \n Acc: \t Root {:.4f} \t Vowel {:.4f} \t Consonant {:.4f} \nRecall: \t Root {:.4f} \t Vowel {:.4f} \t Consonant {:.4f} Total {:.4f}\n'.format(running_loss/(len(valid_loader)), grapheme_root_out/(len(valid_loader)), vowel_diacritic_out/(len(valid_loader)), consonant_diacritic_out/(len(valid_loader)), recall_graph, recall_vowel, recall_consonant, total_recall)
print(msg)
lr_reduce_scheduler.step(running_loss)
history.loc[epoch, 'valid_loss'] = running_loss/(len(valid_loader))
history.loc[epoch, 'valid_grapheme_recall'] = recall_graph
history.loc[epoch, 'valid_vowel_recall'] = recall_vowel
history.loc[epoch, 'valid_conso_recall'] = recall_consonant
history.loc[epoch, 'valid_recall'] = total_recall
history.to_csv(os.path.join(history_dir, 'history_{}.csv'.format(model_name)), index=False)
return running_loss/(len(valid_loader)), total_recall
plist = [
{'params': model.backbone.layer0.parameters(), 'lr': learning_rate/50},
{'params': model.backbone.layer1.parameters(), 'lr': learning_rate/50},
{'params': model.backbone.layer2.parameters(), 'lr': learning_rate/50},
{'params': model.backbone.layer3.parameters(), 'lr': learning_rate/50},
{'params': model.backbone.layer4.parameters(), 'lr': learning_rate/50}
]
# plist = [
# {"params": model.head1.parameters(), "lr": learning_rate},
# {"params": model.head2.parameters(), "lr": learning_rate},
# {"params": model.head3.parameters(), "lr": learning_rate},
# # {"params": model.backbone.extract_features.parameters(), "lr": learning_rate/100}
# ]
# optimizer = Over9000(plist, lr=learning_rate, weight_decay=1e-3)
optimizer = optim.Adam(plist, lr=learning_rate)
# scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer, learning_rate, total_steps=None, epochs=n_epochs, steps_per_epoch=3348, pct_start=0.0,
# anneal_strategy='cos', cycle_momentum=True,base_momentum=0.85, max_momentum=0.95, div_factor=100.0)
lr_reduce_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=patience, verbose=True, threshold=1e-4, threshold_mode='rel', cooldown=0, min_lr=1e-7, eps=1e-08)
criterion = nn.CrossEntropyLoss()
if load_model:
tmp = torch.load(os.path.join(model_dir, model_name+'_loss.pth'))
model.load_state_dict(tmp['model'])
optimizer.load_state_dict(tmp['optim'])
best_valid_recall = tmp['recall']
prev_epoch_num = tmp['epoch']
best_valid_loss = tmp['best_loss']
del tmp
print('Model Loaded!')
if apex:
amp.initialize(model, optimizer, opt_level='O1')
for epoch in range(prev_epoch_num, n_epochs):
torch.cuda.empty_cache()
print(gc.collect())
# stdscr = curses.initscr()
train_recall = train(epoch,history)
valid_loss, valid_recall = evaluate(epoch,history)
if valid_recall > best_valid_recall:
print(f'Validation recall has increased from: {best_valid_recall:.4f} to: {valid_recall:.4f}. Saving checkpoint')
best_state = {'model': model.state_dict(), 'optim': optimizer.state_dict(), 'scheduler': lr_reduce_scheduler.state_dict(), 'loss':valid_loss, 'best_recall':valid_recall, 'epoch':epoch}
# torch.save(best_state, model_name+'.pth')
torch.save(best_state, os.path.join(model_dir, model_name+'_rec.pth'))
torch.save(model.state_dict(), os.path.join(model_dir, '{}_model_weights_best_recall.pth'.format(model_name))) ## Saving model weights based on best validation accuracy.
best_valid_recall = valid_recall ## Set the new validation Recall score to compare with next epoch
if valid_loss<best_valid_loss:
print(f'Validation loss has decreased from: {best_valid_loss:.4f} to: {valid_loss:.4f}. Saving checkpoint')
best_state = {'model': model.state_dict(), 'optim': optimizer.state_dict(), 'scheduler': lr_reduce_scheduler.state_dict(), 'recall':valid_recall, 'best_loss':valid_loss, 'epoch':epoch}
torch.save(best_state, os.path.join(model_dir, model_name+'_loss.pth'))
torch.save(model.state_dict(), os.path.join(model_dir, '{}_model_weights_best_loss.pth'.format(model_name))) ## Saving model weights based on best validation accuracy.
best_valid_loss = valid_loss ## Set the new validation Recall score to compare with next epoch