Multimodel_setup / temp_DRCT_reference /scripts /train_contrastive.py
Thiru1122's picture
feat(deepfake-detection): comprehensive cross-model evaluation, parity audits, and Phase 7 v2 reference preprocessing
151f388
Raw History Blame Contribute Delete
16.6 kB
import argparse
import warnings
import sys
import os
def get_parser():
parser = argparse.ArgumentParser(description="AIGCDetection @cby Training")
parser.add_argument("--model_name", default='efficientnet-b0', help="Setting the model name", type=str)
parser.add_argument("--embedding_size", default=1024, help="Setting the embedding_size", type=int)
parser.add_argument("--pre_layer", default=None, help="Setting the pre_layer: srm or dct", type=str)
parser.add_argument("--num_classes", default=2, help="Setting the num classes", type=int)
parser.add_argument('--freeze_extractor', action='store_true', help='Whether to freeze extractor?')
parser.add_argument("--model_path", default=None, help="Setting the model path", type=str)
parser.add_argument('--no_strict', action='store_true', help='Whether to load model without strict?')
parser.add_argument("--root_path", default='/disk4/chenby/dataset/MSCOCO',
help="Setting the root path for dataset loader", type=str)
parser.add_argument("--fake_root_path", default='/disk4/chenby/dataset/AIGC_MSCOCO',
help="Setting the fake root path for dataset loader", type=str)
parser.add_argument('--is_dire', action='store_true', help='Whether to using DIRE?')
parser.add_argument('--post_aug_mode', default=None, help='Stetting the post aug mode during test phase.')
parser.add_argument('--save_txt', default=None, help='Stetting the save_txt path.')
parser.add_argument("--fake_indexes", default='1',
help="Setting the fake indexes, multi class using '1,2,3,...' ", type=str)
parser.add_argument("--dataset_name", default='MSCOCO', help="Setting the dataset name", type=str)
parser.add_argument("--device_id", default='0',
help="Setting the GPU id, multi gpu split by ',', such as '0,1,2,3'", type=str)
parser.add_argument("--input_size", default=224, help="Image input size", type=int)
parser.add_argument('--is_crop', action='store_true', help='Whether to crop image?')
parser.add_argument('--is_spectrum', action='store_true', help='Whether to using spectrum?')
parser.add_argument("--batch_size", default=64, help="Setting the batch size", type=int)
parser.add_argument("--epoch_start", default=0, help="Setting the epoch start", type=int)
parser.add_argument("--num_epochs", default=50, help="Setting the num epochs", type=int)
parser.add_argument("--num_workers", default=4, help="Setting the num workers", type=int)
parser.add_argument('--is_warmup', action='store_true', help='Whether to using lr warmup')
parser.add_argument("--lr", default=1e-3, help="Setting the learning rate", type=float)
parser.add_argument("--save_flag", default='', help="Setting the save flag", type=str)
parser.add_argument("--sampler_mode", default='', help="Setting the sampler mode", type=str)
parser.add_argument('--is_test', action='store_true', help='Whether to predict the test set?')
parser.add_argument('--is_amp', action='store_true', help='Whether to using amp autocast(使用混合精度加速)?')
parser.add_argument("--inpainting_dir", default='full_inpainting', help="rec_image dir", type=str)
parser.add_argument("--threshold", default=0.5, help="Setting the valid or testing threshold.", type=float)
# contrastive learning params
parser.add_argument('--alpha', default=0.3, type=float, help="Setting the alpha for contrastive learning")
parser.add_argument('--pos_margin', default=0.0, type=float)
parser.add_argument('--neg_margin', default=1.0, type=float)
parser.add_argument('--tau', default=0.5, type=float)
parser.add_argument('--loss_name', default='ContrastiveLoss', type=str)
parser.add_argument('--use_miner', action='store_true', help='Whether to using miner')
parser.add_argument('--memory_size', default=None, type=int)
parser.add_argument("opts", help="Modify config options using the command-line", default=None,
nargs=argparse.REMAINDER)
args = parser.parse_args()
return args
warnings.filterwarnings("ignore")
sys.path.append('..')
args = get_parser()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.device_id)
import torch
import torch.nn as nn
import torch.optim as optim
from catalyst.data import BalanceClassSampler
import time
from torch.utils.data import DataLoader
from torch.cuda.amp import autocast, GradScaler
from tqdm import tqdm
import gc
from sklearn.metrics import roc_auc_score, recall_score, precision_score, accuracy_score, f1_score
import pytorch_warmup as warmup
from utils.utils import Logger, AverageMeter, calculate_fnr
from network.models import get_models
from data.dataset import AIGCDetectionDataset, CLASS2LABEL_MAPPING, GenImage_LIST
from utils.losses import LabelSmoothing, CombinedLoss, FocalLoss
from data.transform import create_train_transforms, create_val_transforms
def merge_tensor(img, label, is_train=True):
def shuffle_tensor(img, label):
indices = torch.randperm(img.size(0))
return img[indices], label[indices]
if isinstance(img, list) and isinstance(label, list):
img, label = torch.cat(img, dim=0), torch.cat(label, dim=0)
if is_train:
img, label = shuffle_tensor(img, label)
return img, label
def eval_model(model, epoch, eval_loader, is_save=True, threshold=0.5, alpha=0.5,
save_txt=None):
model.eval()
losses = AverageMeter()
accuracies = AverageMeter()
eval_process = tqdm(eval_loader)
labels = []
outputs = []
with torch.no_grad():
for i, (img, label) in enumerate(eval_process):
img, label = merge_tensor(img, label, is_train=False)
if i > 0 and i % 1 == 0:
eval_process.set_description("Epoch: %d, Loss: %.4f, Acc: %.4f" %
(epoch, losses.avg, accuracies.avg))
img, label = img.cuda(), label.cuda()
y_pred, embeddings = model(img, return_feature=True)
y_pred = nn.Softmax(dim=1)(y_pred)
loss = (1 - alpha) * criterion(y_pred, label) + alpha * contrastive_loss(embeddings, label)
outputs.append(1 - y_pred[:, 0])
labels.append(label)
acc = (torch.max(y_pred.detach(), 1)[1] == label).sum().item() / img.size(0)
losses.update(loss.item(), img.size(0))
accuracies.update(acc, img.size(0))
outputs = torch.cat(outputs, dim=0).cpu().numpy()
labels = torch.cat(labels, dim=0).cpu().numpy()
labels[labels > 0] = 1
auc = roc_auc_score(labels, outputs)
recall = recall_score(labels, outputs > threshold)
precision = precision_score(labels, outputs > threshold)
binary_acc = accuracy_score(labels, outputs > threshold)
f1 = f1_score(labels, outputs > threshold)
fnr = calculate_fnr(labels, outputs > threshold)
print(f'AUC:{auc}-Recall:{recall}-Precision:{precision}-BinaryAccuracy:{binary_acc}, f1: {f1}, fnr:{fnr}')
if is_save:
train_logger.log(phase="val", values={
'epoch': epoch,
'loss': format(losses.avg, '.4f'),
'acc': format(accuracies.avg, '.4f'),
'lr': optimizer.param_groups[0]['lr']
})
print("Val:\t Loss:{0:.4f} \t Acc:{1:.4f}".format(losses.avg, accuracies.avg))
if save_txt is not None:
return binary_acc, auc, recall, precision, f1, fnr
return accuracies.avg
def train_model(model, criterion, optimizer, epoch, scaler=None, alpha=0.5):
model.train()
losses = AverageMeter()
accuracies = AverageMeter()
training_process = tqdm(train_loader)
for i, (XI, label) in enumerate(training_process):
XI, label = merge_tensor(XI, label, is_train=True)
optimizer.zero_grad()
current_lr = optimizer.param_groups[0]['lr']
if i > 0 and i % 1 == 0:
training_process.set_description(
"Epoch: %d, LR: %.8f, Loss: %.4f, Acc: %.4f" % (
epoch, current_lr, losses.avg, accuracies.avg))
x, label = XI.cuda(), label.cuda()
# Forward pass: Compute predicted y by passing x to the model
if scaler is None:
y_pred, embeddings = model(x, return_feature=True)
# Compute and print loss
loss = (1-alpha) * criterion(y_pred, label) + alpha * contrastive_loss(embeddings, label)
acc = (torch.max(y_pred.detach(), 1)[1] == label).sum().item() / x.size(0)
losses.update(loss.item(), x.size(0))
accuracies.update(acc, x.size(0))
loss.backward()
optimizer.step()
else:
with autocast():
y_pred, embeddings = model(x, return_feature=True)
# Compute and print loss
loss = (1-alpha) * criterion(y_pred, label) + alpha * contrastive_loss(embeddings, label)
acc = (torch.max(y_pred.detach(), 1)[1] == label).sum().item() / x.size(0)
losses.update(loss.item(), x.size(0))
accuracies.update(acc, x.size(0))
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
if args.is_warmup:
with warmup_scheduler.dampening():
scheduler.step()
if not args.is_warmup:
scheduler.step()
train_logger.log(phase="train", values={
'epoch': epoch,
'loss': format(losses.avg, '.4f'),
'acc': format(accuracies.avg, '.4f'),
'lr': optimizer.param_groups[0]['lr']
})
print("Train:\t Loss:{0:.4f} \t Acc:{1:.4f}".format(losses.avg, accuracies.avg))
# 垃圾回收
del losses, accuracies
gc.collect()
# python train.py --device_id=0 --model_name=efficientnet-b0 --input_size=224 --batch_size=48 --fake_indexes=1 --is_amp --save_flag=
if __name__ == '__main__':
batch_size = args.batch_size * torch.cuda.device_count()
writeFile = f"../output/{args.dataset_name}/{args.fake_indexes.replace(',', '_')}/" \
f"{args.model_name.split('/')[-1]}_{args.input_size}{args.save_flag}/logs"
store_name = writeFile.replace('/logs', '/weights')
print(
f'Using gpus:{args.device_id},batch size:{batch_size},gpu_count:{torch.cuda.device_count()},num_classes:{args.num_classes}')
# Load model
model = get_models(model_name=args.model_name, num_classes=args.num_classes,
embedding_size=args.embedding_size, freeze_extractor=args.freeze_extractor)
if args.model_path is not None:
model.load_state_dict(torch.load(args.model_path, map_location='cpu'), strict=not args.no_strict)
print('Model found in {}'.format(args.model_path))
else:
print('No model found, initializing random model.')
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model).cuda()
else:
model = model.cuda()
criterion = nn.CrossEntropyLoss()
# criterion = LabelSmoothing(smoothing=0.05).cuda(device_id)
contrastive_loss = CombinedLoss(loss_name=args.loss_name, embedding_size=args.embedding_size,
pos_margin=args.pos_margin, neg_margin=args.neg_margin, tau=args.tau,
memory_size=args.memory_size, use_miner=args.use_miner, num_classes=args.num_classes)
is_train = not args.is_test
if is_train:
if store_name and not os.path.exists(store_name):
os.makedirs(store_name)
train_logger = Logger(model_name=writeFile, header=['epoch', 'loss', 'acc', 'lr'])
# setting data loader
xdl = AIGCDetectionDataset(args.root_path, fake_root_path=args.fake_root_path, fake_indexes=args.fake_indexes, phase='train',
num_classes=args.num_classes, inpainting_dir=args.inpainting_dir, is_dire=args.is_dire,
transform=create_train_transforms(size=args.input_size, is_crop=args.is_crop)
)
sampler = BalanceClassSampler(labels=xdl.get_labels(), mode=args.sampler_mode) if args.sampler_mode != '' else None # "upsampling"
train_loader = DataLoader(xdl, batch_size=batch_size, shuffle=sampler is None, num_workers=args.num_workers, sampler=sampler)
train_dataset_len = len(xdl)
xdl_eval = AIGCDetectionDataset(args.root_path, fake_root_path=args.fake_root_path, fake_indexes=args.fake_indexes, phase='val',
num_classes=args.num_classes, inpainting_dir=args.inpainting_dir, is_dire=args.is_dire,
transform=create_val_transforms(size=args.input_size, is_crop=args.is_crop)
)
eval_loader = DataLoader(xdl_eval, batch_size=batch_size, shuffle=False, num_workers=args.num_workers)
eval_dataset_len = len(xdl_eval)
print('train_dataset_len:', train_dataset_len, 'eval_dataset_len:', eval_dataset_len)
# optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
optimizer = optim.AdamW(model.parameters(), lr=args.lr, weight_decay=4e-5)
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.9)
if not args.is_warmup:
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, 5)
else:
num_steps = train_dataset_len * args.num_epochs
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_steps)
warmup_scheduler = warmup.UntunedLinearWarmup(optimizer)
best_acc = 0.5 if args.epoch_start == 0 else eval_model(model, args.epoch_start - 1, eval_loader, is_save=False)
for epoch in range(args.epoch_start, args.num_epochs):
train_model(model, criterion, optimizer, epoch, scaler=GradScaler() if args.is_amp else None, alpha=args.alpha)
if epoch % 1 == 0 or epoch == args.num_epochs - 1:
acc = eval_model(model, epoch, eval_loader, alpha=args.alpha)
if best_acc < acc:
best_acc = acc
save_path = '{}/{}_acc{:.4f}.pth'.format(store_name, epoch, acc)
if torch.cuda.device_count() > 1:
torch.save(model.module.state_dict(), save_path)
else:
torch.save(model.state_dict(), save_path)
print(f'Current best acc:{best_acc}')
last_save_path = '{}/last_acc{:.4f}.pth'.format(store_name, acc)
if torch.cuda.device_count() > 1:
torch.save(model.module.state_dict(), last_save_path)
else:
torch.save(model.state_dict(), last_save_path)
else:
start = time.time()
epoch_start = 1
num_epochs = 1
xdl_test = AIGCDetectionDataset(args.root_path, fake_root_path=args.fake_root_path, fake_indexes=args.ake_indexes,
phase='test', num_classes=args.num_classes, is_dire=args.is_dire,
post_aug_mode=args.post_aug_mode, inpainting_dir=args.inpainting_dir,
transform=create_val_transforms(size=args.input_size, is_crop=args.is_crop)
)
test_loader = DataLoader(xdl_test, batch_size=batch_size, shuffle=False, num_workers=4)
test_dataset_len = len(xdl_test)
print('test_dataset_len:', test_dataset_len)
out_metrics = eval_model(model, epoch_start, test_loader, is_save=False, threshold=args.threshold, save_txt=args.save_txt)
print('Total time:', time.time() - start)
# 保存测试结果
if args.save_txt is not None:
os.makedirs(os.path.dirname(args.save_txt), exist_ok=True)
acc, auc, recall, precision, f1, fnr = out_metrics
with open(args.save_txt, 'a') as file:
if args.dataset_name == 'GenImage':
class_name = GenImage_LIST[int(args.fake_indexes) - 1]
else:
class_name = list(CLASS2LABEL_MAPPING.keys())[int(args.fake_indexes)]
result_str = f'model_path:{args.model_path}, post_aug_mode:{args.post_aug_mode}, class_name:{class_name}\n' \
f'acc:{acc:.4f}, auc:{auc:.4f}, recall:{recall:.4f}, precision:{precision:.4f}, ' \
f'f1:{f1:.4f}, fnr: {fnr}\n'
file.write(result_str)
print(f'The result was saved in {args.save_txt}')