-
Notifications
You must be signed in to change notification settings - Fork 0
/
baseline.py
247 lines (232 loc) · 10.6 KB
/
baseline.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
from pathlib import Path
import argparse
import time
import matplotlib.pyplot as plt
import yaml
import logging
import torch
from torch import nn
from torch.optim import SGD, Adam, lr_scheduler
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from timm.models.resnet import resnet18, resnet50, resnet101, resnext50_32x4d, resnext101_32x8d, Bottleneck, BasicBlock
from timm.models.efficientnet import efficientnet_b2
from timm.models.swin_transformer import swin_base_patch4_window7_224, swin_base_patch4_window7_224_in22k, \
swin_base_patch4_window12_384_in22k, swin_large_patch4_window7_224_in22k, swin_small_patch4_window7_224
from timm.data import Mixup
from timm.loss import LabelSmoothingCrossEntropy
from torch.cuda import amp
import cv2
import numpy as np
from tqdm import tqdm
import pandas as pd
from utils import init_seeds, increment_path, MetricLogger, mixup_data, one_hot, warmup_cosine_schedule, CenterLoss
logging.basicConfig(format="%(message)s", level=logging.INFO)
logger = logging.getLogger(__name__)
logger.setLevel(level=logging.INFO)
DATA_TXT = {'train': 'train_qtcom.txt', 'val': 'val_qtcom.txt', 'test': 'test_qtcom.txt'}
DATA_DIR = {'train': 'Train_qtc', 'val': 'val', 'test': 'test_new'}
def parse_opt():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', default=32, type=int)
parser.add_argument('--img_size', default=224, type=int)
parser.add_argument('--num_classes', default=1000, type=int)
parser.add_argument('--lr', default=4e-5, type=float)
parser.add_argument('--warmup_epochs', default=0, type=float)
parser.add_argument('--label_smooth', default=0., type=float)
parser.add_argument('--weight_decay', default=1e-4, type=float)
parser.add_argument('--batch_size', default=32, type=int)
parser.add_argument('--num_workers', default=8, type=int)
parser.add_argument('--model', default='resnet50')
parser.add_argument('--device', default='cuda:0')
parser.add_argument('--adam', action='store_true')
parser.add_argument('--pretrained', action='store_true')
parser.add_argument('--bce_loss', action='store_true')
parser.add_argument('--center_loss', action='store_true')
parser.add_argument('--mix_up', default=0., type=float)
parser.add_argument('--cut_mix', default=0., type=float)
parser.add_argument('--save_dir', default='run/exp')
opt = parser.parse_intermixed_args()
return opt
class FoodDataset(Dataset):
def __init__(self, mode, infer_size=224): # mode 'train','val' or 'test'
self.mode = mode
if isinstance(infer_size,int):
infer_size = infer_size, infer_size
self.infer_size = infer_size
txt = DATA_TXT[mode]
direc = Path(DATA_DIR[mode]).absolute()
if mode in ['train', 'val']:
img_paths, labels = [], []
with open(txt) as f:
for line in f.readlines():
path, label = line.strip('\n').split()
img_paths.append(str(direc / path))
labels.append(label)
self.img_paths = img_paths
self.labels = labels
else:
img_paths = []
with open(txt) as f:
for line in f.readlines():
path = line.strip('\n')
img_paths.append(str(direc / path))
self.img_paths = img_paths
self.trans = None
self.trans = transforms.Compose([transforms.RandomAffine(degrees=10, translate=(0.2, 0.2),scale=(0.8, 1.5)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, ),
])
def __len__(self):
return len(self.img_paths)
def __getitem__(self, item):
img = cv2.imread(self.img_paths[item])
img=cv2.resize(img,self.infer_size)
img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB
img = torch.from_numpy(np.ascontiguousarray(img)) / 255.
if self.mode == 'train':
if self.trans:
img = self.trans(img)
return img, torch.tensor(int(self.labels[item])).long()
elif self.mode == 'val':
return img, torch.tensor(int(self.labels[item])).long()
else:
return img
def train(opt):
init_seeds()
save_dir, device, epochs, lr, weight_decay, batch_size, num_workers,img_size = Path(
opt.save_dir), opt.device, opt.epochs, opt.lr, opt.weight_decay, opt.batch_size, opt.num_workers,opt.img_size
# save dir
save_dir = increment_path(save_dir)
w = save_dir / 'weights' # weights dir, /==.joinpath()
w.mkdir(parents=True, exist_ok=True) # make dir
last, best = w / 'last.pt', w / 'best.pt'
txt_res, img_res = save_dir / 'res.txt', save_dir / 'res.png'
submit = save_dir / 'submit.csv'
with open(save_dir / 'opt.yaml', 'w') as f:
yaml.safe_dump(vars(opt), f, sort_keys=False)
# logger
metric_logger = MetricLogger()
# model
model = eval(opt.model)(opt.pretrained,num_classes=opt.num_classes).to(device)
# model.load_state_dict(torch.load('run/exp12/weights/best.pt'))
# data
train_ds = FoodDataset('train',img_size)
val_ds = FoodDataset('val',img_size)
test_ds = FoodDataset('test',img_size)
train_dl = DataLoader(train_ds, batch_size, True, num_workers=num_workers,)
val_dl = DataLoader(val_ds, batch_size, False, num_workers=num_workers)
test_dl = DataLoader(test_ds, batch_size, False, num_workers=num_workers)
# optimizer
g0, g1, g2 = [], [], [] # optimizer parameter groups
for v in model.modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter) and v.bias.requires_grad: # bias
g2.append(v.bias)
if isinstance(v, nn.BatchNorm2d) and v.weight.requires_grad: # weight (no decay)
g0.append(v.weight)
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter) and v.weight.requires_grad: # weight (with decay)
g1.append(v.weight)
if opt.adam:
optimizer = Adam(g0+g2, lr=lr, betas=(0.9, 0.999)) # adjust beta1 to momentum
else:
optimizer = SGD(g0+g2, lr=lr, momentum=0.9)
optimizer.add_param_group({'params': g1, 'weight_decay': weight_decay}) # add g1 with weight_decay
logger.info(f"{'optimizer:'} {type(optimizer).__name__} with parameter groups "
f"{len(g0)} weight, {len(g1)} weight (with decay), {len(g2)} bias")
# scheduler
steps = epochs * len(train_dl)
warmup_steps = opt.warmup_epochs * len(train_dl)
scheduler = warmup_cosine_schedule(optimizer, warmup_steps, steps)
# loss func
if opt.bce_loss:
loss_func = nn.BCEWithLogitsLoss()
off_value = opt.label_smooth / opt.num_classes
on_value = 1. - opt.label_smooth + off_value
elif opt.center_loss:
loss_func = CenterLoss()
else:
loss_func = LabelSmoothingCrossEntropy(opt.label_smooth)
# train
best_acc_top1 = float('-inf')
scaler = amp.GradScaler()
for epoch in range(epochs):
model.train()
logger.info(f'epoch: {epoch + 1}/{epochs}')
logger.info('training:')
train_loss, val_loss, top1_acc, top5_acc = 0, 0, 0, 0
for i, (img, label) in tqdm(enumerate(train_dl),total=len(train_dl)):
img = img.to(device)
label = label.to(device)
with amp.autocast():
img, label1, label2, lam = mixup_data(img, label, opt.mix_up, opt.cut_mix)
pred = model(img)
if opt.bce_loss:
_label = lam * one_hot(label1, opt.num_classes, on_value, off_value) + (1 - lam) * \
one_hot(label2, opt.num_classes, on_value, off_value)
loss = loss_func(pred, _label)*opt.num_classes
else:
loss = lam*loss_func(pred, label1)+(1-lam)*loss_func(pred,label2)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
scheduler.step()
train_loss = (train_loss * i + loss.item()) / (i + 1)
logger.info('validating')
# val
with torch.no_grad():
model.eval()
for i, (img, label) in tqdm(enumerate(val_dl), total=len(val_dl)):
img = img.to(device)
label = label.to(device)
with amp.autocast():
pred = model(img)
if opt.bce_loss:
_label = one_hot(label, opt.num_classes)
loss = loss_func(pred, _label)*opt.num_classes
else:
loss = loss_func(pred, label)
_, arg = pred.topk(5, 1)
label = label.view(-1, 1)
_top1_acc = (label == arg[:, :1]).sum().item() / label.numel()
_top5_acc = (label == arg).any(1).sum().item() / label.numel()
val_loss = (val_loss * i + loss.item()) / (i + 1)
top1_acc = (top1_acc * i + _top1_acc) / (i + 1)
top5_acc = (top5_acc * i + _top5_acc) / (i + 1)
metric_logger.update(lr=scheduler.get_last_lr()[0], train_loss=train_loss, val_loss=val_loss, top1_acc=top1_acc,
top5_acc=top5_acc)
torch.save(model.state_dict(), last)
logger.info(str(metric_logger))
if top1_acc > best_acc_top1:
logger.info('changing best weight...')
torch.save(model.state_dict(), best)
best_acc_top1 = top1_acc
metric_logger.output_csv(txt_res)
metric_logger.plot(img_res, 3)
# submit
model.load_state_dict(torch.load(best))
arg = np.argmax(metric_logger.meters['top1_acc'])
top1_acc = metric_logger.meters['top1_acc'][arg]
top5_acc = metric_logger.meters['top5_acc'][arg]
logger.info(f'\nsubmit: top1_acc={top1_acc},top5_acc={top5_acc}')
with torch.no_grad():
model.eval()
submit_res = []
for img in tqdm(test_dl):
img = img.to(device)
pred = model(img)
_, arg = pred.topk(5, 1)
submit_res.append(arg)
submit_res = torch.cat(submit_res)
names = []
with open(DATA_TXT['test']) as f:
for line in f.readlines():
path = line.strip('\n')
names.append(path)
sub = pd.DataFrame()
sub['name'] = names
sub[['1', '2', '3', '4', '5']] = submit_res.cpu().numpy()
sub.to_csv(submit,index=False,header=False)
if __name__ == '__main__':
opt = parse_opt()
train(opt)