-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
215 lines (170 loc) · 9.59 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
import sys
sys.path.append('..')
import numpy as np
import os
import csv
import copy
import logging
from tqdm import tqdm
import math
import random
from collections import defaultdict
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
from torch.utils.data import DataLoader
from torch.cuda.amp import autocast, GradScaler
from torch.nn.utils import clip_grad_norm_
from dataset import TrainDataset, TestDataset, ApplyWeightedRandomSampler
from utils import AvgrageMeter, performances_cross_db, compute_video_score
from model import MixStyleResCausalModel
#import metric_utils
torch.autograd.set_detect_anomaly(True)
def save_checkpoint(save_path, epoch, model, loss, lr_scheduler, optimizer):
save_state = {'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'loss': loss,
'epoch': epoch}
torch.save(save_state, save_path)
def run_training(train_csv, test_csv, log_file, output_path, args):
train_dataset = TrainDataset(csv_file=train_csv, input_shape=args.input_shape)
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, sampler=ApplyWeightedRandomSampler(train_csv),
num_workers=4, pin_memory=True, drop_last=True)
test_dataset = TestDataset(csv_file=test_csv, input_shape=args.input_shape)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4, pin_memory=True)
checkpoint_save_dir = os.path.join('checkpoints/', args.prefix)
print('Checkpoints folders', checkpoint_save_dir)
if not os.path.isdir(checkpoint_save_dir):
os.makedirs(checkpoint_save_dir)
model = torch.nn.DataParallel(MixStyleResCausalModel(model_name=args.model_name, pretrained=args.pretrain, num_classes=args.num_classes, prob=args.prob))
model = model.cuda()
optimizer = torch.optim.SGD([
{'params': model.module.feature_extractor.parameters()},
{'params': model.module.classifier.parameters(), 'lr': float(args.lr[1])},
], lr=float(args.lr[0]), momentum=0.9, weight_decay=0.0005)
#lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.998)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[30,45], gamma=0.5)
#lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=15)
cen_criterion = torch.nn.CrossEntropyLoss().cuda()
scaler = GradScaler()
flooding_impactor = 0.001
for epoch in range(1, args.max_epoch+1):
if os.path.isfile(os.path.join(checkpoint_save_dir, '{}.pth'.format(epoch))):
model.load_state_dict(torch.load(os.path.join(checkpoint_save_dir, '{}.pth'.format(epoch))))
continue
else:
print('-------------- train ------------------------')
model.train()
loss_total = AvgrageMeter()
loss_1_total = AvgrageMeter()
loss_2_total = AvgrageMeter()
progress_bar = tqdm(train_loader)
for i, data in enumerate(progress_bar):
progress_bar.set_description('Epoch ' + str(epoch))
raw = data["images"].cuda()
labels = data["labels"].cuda()
with autocast():
output, cf_output = model(raw, labels=labels, cf=args.ops, norm=args.norm)
loss_1 = cen_criterion(output, labels.to(torch.int64)) * 2
loss_2 = cen_criterion((output - cf_output), labels.to(torch.int64))
loss_1 = (loss_1 - 0.001).abs() + 0.001
loss_2 = (loss_2 - 0.001).abs() + 0.001
loss = loss_1 + loss_2
#loss = (loss - flooding_impactor).abs() + flooding_impactor
loss_total.update(loss.data, raw.shape[0])
loss_1_total.update(loss_1.data, raw.shape[0])
loss_2_total.update(loss_2.data, raw.shape[0])
clip_grad_norm_(model.parameters(), max_norm=5, norm_type=2)
optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
progress_bar.set_postfix(
loss ='%.5f' % (loss_total.avg),
loss_1 ='%.5f' % (loss_1_total.avg),
loss_2 ='%.5f' % (loss_2_total.avg),
)
torch.save( model.state_dict(), os.path.join(checkpoint_save_dir, '{}.pth'.format(epoch)))
tqdm.write('Epoch: %d, Train: loss_total= %.4f, loss_1_total= %.4f, loss_2_total= %.4f, lr_1=%.6f, lr_2=%.6f \n' % (epoch, loss_total.avg, loss_1_total.avg, loss_2_total.avg, optimizer.param_groups[0]['lr'], optimizer.param_groups[1]['lr']))# , curr_lr[0]
log_file.write('Epoch: %d, Train: loss_total= %.4f, loss_1_total= %.4f, loss_2_total= %.4f, lr_2=%.6f, lr_2=%.6f \n' % (epoch, loss_total.avg, loss_1_total.avg, loss_2_total.avg, optimizer.param_groups[0]['lr'], optimizer.param_groups[1]['lr'])) #, curr_lr[0]
log_file.flush()
print ('------------ test 1 -------------------')
AUC_value, HTER_value = test_model(model, test_loader)
lr_scheduler.step()
#lr_scheduler.step(hter)
write_txt = 'Test: AUC=%.4f, HTER= %.4f \n' % (AUC_value, HTER_value)
tqdm.write(write_txt)
log_file.write(write_txt)
log_file.flush()
def test_model(model, data_loader, video_format=True):
model.eval()
raw_test_scores, gt_labels = [], []
raw_scores_dict = []
raw_test_video_ids = []
with torch.no_grad():
# for train
for i, data in enumerate(tqdm(data_loader)):
raw, labels, img_pathes = data["images"].cuda(), data["labels"], data["img_path"]
output = model(raw, cf=None)
raw_scores = output.softmax(dim=1)[:, 1].cpu().data.numpy()
#raw_scores = 1 - raw_scores
raw_test_scores.extend(raw_scores)
gt_labels.extend(labels.data.numpy())
for j in range(raw.shape[0]):
image_name = os.path.splitext(os.path.basename(img_pathes[j]))[0]
video_id = os.path.join(os.path.dirname(img_pathes[j]), image_name.rsplit('_', 1)[0])
raw_test_video_ids.append(video_id)
if video_format:
raw_test_scores, gt_labels, _ = compute_video_score(raw_test_video_ids, raw_test_scores, gt_labels)
raw_test_stats = [np.mean(raw_test_scores), np.std(raw_test_scores)]
raw_test_scores = ( raw_test_scores - raw_test_stats[0]) / raw_test_stats[1]
AUC_values, _, _, HTER_values = performances_cross_db(raw_test_scores, gt_labels)
return AUC_values, HTER_values
def set_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
np.random.seed(seed) # Numpy module.
random.seed(seed) # Python random module.
#torch.use_deterministic_algorithms(True, warn_only=True)
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
os.environ['PYTHONHASHSEED'] = str(seed)
if __name__ == "__main__":
torch.cuda.empty_cache()
#cudnn.benchmark = True
set_seed(seed=777)
parser = argparse.ArgumentParser(description='CF-PAD')
parser.add_argument("--prefix", default='CFPAD', type=str, help="description")
parser.add_argument("--model_name", default='resnet50d', type=str, help="model backbone")
parser.add_argument("--training_csv", type=str, help="csv contains training data")
parser.add_argument("--test_csv", type=str, help="csv contains test data")
parser.add_argument('--lr', type=list, help='Learning rate', default=[0.001, 0.01])
parser.add_argument("--input_shape", default=(224, 224), type=tuple, help="Neural Network input shape")
parser.add_argument("--max_epoch", default=50, type=int, help="maximum epochs")
parser.add_argument("--batch_size", default=128, type=int, help="train batch size")
parser.add_argument("--pretrain", default=True, type=lambda x: (str(x).lower() in ['true','1', 'yes']))
parser.add_argument("--num_classes", default=2, type=int, help="number of classes (bona fide and attack)")
parser.add_argument("--ops", default=['cs','dropout','replace'], type=str, nargs='*', help="operations for causality")
parser.add_argument("--norm", default=False, type=lambda x: (str(x).lower() in ['true','1', 'yes']))
parser.add_argument("--prob", default=0.2, type=float, help="probabilities of CF")
args = parser.parse_args()
#training_csv = '/data/mfang/FacePAD_DB/WholeFrames/CropFaceFrames/Protocols/casia.csv'
#test_csv = '/data/mfang/FacePAD_DB/WholeFrames/CropFaceFrames/Protocols/replayattack.csv'
logging_filename = os.path.join('logs/', '{}.txt'.format(args.prefix))
if not os.path.isdir('logs/'):
os.makedirs('logs/')
log_file = open(logging_filename, 'a')
log_file.write(f'train data: {args.training_csv} \n test data: {args.test_csv} \n causality ops: {args.ops}, prob: {args.prob}, norm feature: {args.norm} \n model_name: {args.model_name}, {args.pretrain}, lr: {args.lr}, prefix: {args.prefix}, bs: {args.batch_size} \n')
log_file.write(f"-------------------------------------------- \n")
log_file.flush()
run_training(train_csv=args.training_csv,
test_csv=args.test_csv,
log_file=log_file,
output_path='prediction_scores',
args=args)