def main():
# Scale and initialize the parameters
best_prec1 = 0
configs.TRAIN.epochs = int(
math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items():
print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
if configs.pretrained:
print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](pretrained=True)
else:
print("=> creating model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch]()
# Wrap the model into DataParallel
model.cuda()
# reverse mapping
param_to_moduleName = {}
for m in model.modules():
for p in m.parameters(recurse=False):
param_to_moduleName[p] = str(type(m).__name__)
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
group_decay = [
p for p in model.parameters()
if 'BatchNorm' not in param_to_moduleName[p]
]
group_no_decay = [
p for p in model.parameters() if 'BatchNorm' in param_to_moduleName[p]
]
groups = [
dict(params=group_decay),
dict(params=group_no_decay, weight_decay=0)
]
optimizer = torch.optim.SGD(groups,
configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
if configs.TRAIN.half and not configs.evaluate:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
model = torch.nn.DataParallel(model)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
# Initiate data loaders
traindir = os.path.join(configs.data, 'train')
valdir = os.path.join(configs.data, 'val')
resize_transform = []
if configs.DATA.img_size > 0:
resize_transform = [transforms.Resize(configs.DATA.img_size)]
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose(resize_transform + [
transforms.RandomResizedCrop(configs.DATA.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
]))
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=configs.DATA.batch_size,
shuffle=True,
num_workers=configs.DATA.workers,
pin_memory=True,
sampler=None)
normalize = transforms.Normalize(mean=configs.TRAIN.mean,
std=configs.TRAIN.std)
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(
valdir,
transforms.Compose(resize_transform + [
transforms.CenterCrop(configs.DATA.crop_size),
transforms.ToTensor(),
])),
batch_size=configs.DATA.batch_size,
shuffle=False,
num_workers=configs.DATA.workers,
pin_memory=True)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0],
pgd_param[1], configs, logger)
validate(val_loader, model, criterion, configs, logger)
return
lr_schedule = lambda t: np.interp([t], configs.TRAIN.lr_epochs, configs.
TRAIN.lr_values)[0]
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, lr_schedule,
configs.TRAIN.half)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion, configs, logger)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': configs.TRAIN.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, os.path.join('trained_models',
f'{configs.output_name}'), epoch + 1)
def train_free():
# Scale and initialize the parameters
best_prec1 = 0
configs.TRAIN.epochs = int(
math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items():
print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
if configs.pretrained:
print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](pretrained=True)
else:
print("=> creating model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch]()
# Wrap the model into DataParallel
model = torch.nn.DataParallel(model).cuda()
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
# Optimizer:
optimizer = torch.optim.SGD(model.parameters(),
configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
# setup data loader
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(configs.DATA.cifar10_mean,
configs.DATA.cifar10_std)
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(configs.DATA.cifar10_mean,
configs.DATA.cifar10_std)
])
train_dataset = torchvision.datasets.CIFAR10(root='../data',
train=True,
download=True,
transform=transform_train)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=configs.DATA.batch_size,
shuffle=True,
num_workers=configs.DATA.workers,
pin_memory=True,
sampler=None)
testset = torchvision.datasets.CIFAR10(root='../data',
train=False,
download=True,
transform=transform_test)
val_loader = torch.utils.data.DataLoader(
testset,
batch_size=configs.DATA.batch_size,
shuffle=False,
num_workers=configs.DATA.workers,
pin_memory=True)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0],
pgd_param[1], configs, logger)
validate(val_loader, model, criterion, configs, logger)
return
early_stopping = EarlyStopping(patience=15, verbose=True)
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
adjust_learning_rate(configs.TRAIN.lr, optimizer, epoch,
configs.ADV.n_repeats)
# train for one epoch
do_train_free(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1, stopped, early_stopping = validate(val_loader, model, criterion,
configs, logger,
early_stopping)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': configs.TRAIN.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, os.path.join('trained_models', configs.output_name))
if (stopped):
break
# Automatically perform PGD Attacks at the end of training
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0], pgd_param[1],
configs, logger)
def main():
# Scale and initialize the parameters
best_prec1 = 0
if not configs.full_epoch:
configs.TRAIN.epochs = int(math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items(): print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
# if configs.pretrained:
# print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
# model = models.__dict__[configs.TRAIN.arch](pretrained=True)
# else:
# print("=> creating model '{}'".format(configs.TRAIN.arch))
# model = models.__dict__[configs.TRAIN.arch]()
print('loading arma model: ')
if configs.model =='res':
model = ResNet_('ResNet18',True,'CIFAR10',0,3,3)
else:
model = VGG('VGG16',True,'CIFAR10',0,3,3)
# Wrap the model into DataParallel
model = torch.nn.DataParallel(model).cuda()
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
# Optimizer:
optimizer = torch.optim.SGD(model.parameters(), configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
#transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
#transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
train_dataset = datasets.CIFAR10('../data', train=True, download=True,transform=transform_train)
test_dataset = datasets.CIFAR10('../data', train=False, transform=transform_test)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=256, shuffle=True,
num_workers=16, pin_memory=True, drop_last=True)
val_loader = torch.utils.data.DataLoader(test_dataset,batch_size=256,\
shuffle=False,num_workers=16, pin_memory=True)
# Initiate data loaders
# traindir = os.path.join(configs.data, 'train')
# valdir = os.path.join(configs.data, 'val')
# train_dataset = datasets.ImageFolder(
# traindir,
# transforms.Compose([
# transforms.RandomResizedCrop(configs.DATA.crop_size),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# ]))
# train_loader = torch.utils.data.DataLoader(
# train_dataset, batch_size=configs.DATA.batch_size, shuffle=True,
# num_workers=configs.DATA.workers, pin_memory=True, sampler=None)
normalize = transforms.Normalize(mean=configs.TRAIN.mean,
std=configs.TRAIN.std)
# val_loader = torch.utils.data.DataLoader(
# datasets.ImageFolder(valdir, transforms.Compose([
# transforms.Resize(configs.DATA.img_size),
# transforms.CenterCrop(configs.DATA.crop_size),
# transforms.ToTensor(),
# ])),
# batch_size=configs.DATA.batch_size, shuffle=False,
# num_workers=configs.DATA.workers, pin_memory=True)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0], pgd_param[1], configs, logger)
validate(val_loader, model, criterion, configs, logger)
return
save_folder = os.path.join('trained_models', configs.output_name)
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
adjust_learning_rate(configs.TRAIN.lr, optimizer, epoch, \
configs.ADV.n_repeats,adjust_epoch_factor=configs.adj_lr_factor)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion, configs, logger)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint({
'epoch': epoch + 1,
'arch': configs.TRAIN.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer' : optimizer.state_dict(),
}, is_best, os.path.join('trained_models', configs.output_name))
# Automatically perform PGD Attacks at the end of training
logger.info(pad_str(' Performing PGD Attacks '))
f = open(save_folder+'/log.txt','w')
for pgd_param in configs.ADV.pgd_attack:
acc = validate_pgd(val_loader, model, criterion, pgd_param[0], pgd_param[1], configs, logger)
f.write(str(pgd_param),' ',acc,'\n')
f.close()
def main():
global name_all
global para_all
global mean_per_epoch_cle
global std_per_epoch_cle
global mean_per_epoch_adv
global std_per_epoch_adv
# Scale and initialize the parameters
best_prec1 = 0
configs.TRAIN.epochs = int(
math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items():
print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
if configs.pretrained:
print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](pretrained=True,
useFN=configs.FN,
useWN=configs.WN)
else:
print("=> creating model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](useFN=configs.FN,
useWN=configs.WN)
# Wrap the model into DataParallel
model = torch.nn.DataParallel(model).cuda()
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
# Optimizer:
optimizer = torch.optim.SGD(model.parameters(),
configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
# Initiate data loaders
traindir = os.path.join(configs.data, 'train')
valdir = os.path.join(configs.data, 'val')
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(configs.DATA.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
]))
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=configs.DATA.batch_size,
shuffle=True,
num_workers=configs.DATA.workers,
pin_memory=True,
sampler=None)
normalize = transforms.Normalize(mean=configs.TRAIN.mean,
std=configs.TRAIN.std)
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Resize(configs.DATA.img_size),
transforms.CenterCrop(configs.DATA.crop_size),
transforms.ToTensor(),
])),
batch_size=configs.DATA.batch_size,
shuffle=False,
num_workers=configs.DATA.workers,
pin_memory=True)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0],
pgd_param[1], configs, logger)
validate(val_loader, model, criterion, configs, logger)
return
# calculate the gradients w.r.t conv and bn layers
if configs.print_gradients == True:
pattern_bn = re.compile(r'bn')
pattern_conv = re.compile(r'conv')
for name, param in model.named_parameters():
if pattern_bn.search(name) or pattern_conv.search(name):
name_all.append(name)
para_all[name] = param
mean_per_epoch_cle[name] = []
mean_per_epoch_adv[name] = []
std_per_epoch_cle[name] = []
std_per_epoch_adv[name] = []
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
adjust_learning_rate(configs.TRAIN.lr, optimizer, epoch,
configs.ADV.n_repeats)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion, configs, logger)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': configs.TRAIN.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, os.path.join('trained_models', configs.output_name))
# Automatically perform PGD Attacks at the end of training
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0], pgd_param[1],
configs, logger)
def main():
print(configs.TRAIN.epochs)
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Create weights folder
if not os.path.isdir(os.path.join('train_fgsm_output',
configs.output_name)):
os.makedirs(os.path.join('train_fgsm_output', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items():
print('{}: {}'.format(k, v))
logger.info(pad_str(''))
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
train_loader, test_loader = get_loaders(args.data_dir,
configs.DATA.batch_size,
configs.DATA.workers,
configs.DATA.crop_size)
epsilon = (args.epsilon / 255.) / std
alpha = (args.alpha / 255.) / std
pgd_alpha = (2 / 255.) / std
print("=> creating model '{}'".format(configs.TRAIN.arch))
# model = models.__dict__[configs.TRAIN.arch]()
model = WideResNet().to(device)
# model = PreActResNet18().cuda()
# resnet50 = models.resnet50()
# model = resnet18
# if(configs.load_weights):
# logger.info(pad_str("LOADING WEIGHTS"))
# model_path = "cifar_model_weights_30_epochs.pth"
# state_dict = torch.load(model_path)
# model.load_state_dict(state_dict)
# model = model.eval()
# Use GPU or CPU
model = model.to(device)
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
logger.info(pad_str("PGD-" + str(pgd_param[0])))
pgd_loss, pgd_acc = evaluate_pgd(test_loader, model, pgd_param[0],
10)
test_loss, test_acc = evaluate_standard(test_loader, model)
logger.info('Test Loss \t Test Acc \t PGD Loss \t PGD Acc')
logger.info('%.4f \t \t %.4f \t %.4f \t %.4f', test_loss, test_acc,
pgd_loss, pgd_acc)
return
model.train()
opt = torch.optim.SGD(model.parameters(),
lr=args.lr_max,
momentum=args.momentum,
weight_decay=args.weight_decay)
amp_args = dict(opt_level=args.opt_level,
loss_scale=args.loss_scale,
verbosity=False)
if args.opt_level == 'O2':
amp_args['master_weights'] = args.master_weights
model, opt = amp.initialize(model, opt, **amp_args)
criterion = nn.CrossEntropyLoss()
if args.delta_init == 'previous':
delta = torch.zeros(configs.DATA.batch_size, 3, 32, 32).cuda()
lr_steps = configs.TRAIN.epochs * len(train_loader)
if args.lr_schedule == 'cyclic':
scheduler = torch.optim.lr_scheduler.CyclicLR(
opt,
base_lr=args.lr_min,
max_lr=args.lr_max,
step_size_up=lr_steps / 2,
step_size_down=lr_steps / 2)
elif args.lr_schedule == 'multistep':
scheduler = torch.optim.lr_scheduler.MultiStepLR(
opt, milestones=[lr_steps / 2, lr_steps * 3 / 4], gamma=0.1)
# Training
prev_robust_acc = 0.
start_train_time = time.time()
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
# Train
train(train_loader, model, criterion, epoch, epsilon, opt, alpha,
scheduler)
prec1 = validate(test_loader, model, criterion, configs, logger)
# prec1 = 100
if args.early_stop:
# Check current PGD robustness of model using random minibatch
X, y = first_batch
pgd_delta = attack_pgd(model, X, y, epsilon, pgd_alpha, 5, 1, opt)
with torch.no_grad():
output = model(
clamp(X + pgd_delta[:X.size(0)], lower_limit, upper_limit))
robust_acc = (output.max(1)[1] == y).sum().item() / y.size(0)
if robust_acc - prev_robust_acc < -0.2:
break
prev_robust_acc = robust_acc
best_state_dict = copy.deepcopy(model.state_dict())
train_time = time.time()
if not args.early_stop:
best_state_dict = model.state_dict()
torch.save(best_state_dict, os.path.join(args.out_dir, 'model.pth'))
logger.info('Total train time: %.4f minutes',
(train_time - start_train_time) / 60)
# Automatically perform PGD Attacks at the end of training
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(test_loader, model, criterion, pgd_param[0], pgd_param[1],
configs, logger)
# Evaluation
'''
def main():
global name_all
global para_all
global mean_per_epoch_cle
global std_per_epoch_cle
global mean_per_epoch_adv
global std_per_epoch_adv
# Scale and initialize the parameters
best_prec1 = 0
configs.TRAIN.epochs = int(math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items(): print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
if configs.pretrained:
print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](pretrained=True, useFN=configs.FN, useWN=configs.WN)
else:
print("=> creating model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](useFN=configs.FN, useWN=configs.WN)
# Wrap the model into DataParallel
model = torch.nn.DataParallel(model).cuda()
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
# Optimizer:
optimizer = torch.optim.SGD(model.parameters(), configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
# Initiate data loaders
normalize = transforms.Normalize(mean=configs.TRAIN.mean, std=configs.TRAIN.std)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
valdir = os.path.join(configs.data, 'val')
#fix random seed for save images and perturbation
#torch.manual_seed(1234)
#np.random.seed(1234)
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(valdir, transforms.Compose([
transforms.Resize(configs.DATA.img_size),
transforms.CenterCrop(configs.DATA.crop_size),
transforms.ToTensor(),
])),
batch_size=configs.DATA.batch_size, shuffle=False,
num_workers=configs.DATA.workers, pin_memory=True)
if configs.FN or configs.WN:
print('use HE for evaluate and save images!!!!!!!!!!')
HEs = True
else:
print('DO NOT use HE for evaluate and save images!!!!!!!!!!')
HEs = False
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(val_loader, model, criterion, pgd_param[0], pgd_param[1], configs, logger, save_image=True, HE=HEs)
validate(val_loader, model, criterion, configs, logger)
return
# If evaluate on ImageNet-C
if configs.eva_on_imagenet_c:
if configs.FN or configs.WN:
print('use HE for evaluate and save images!!!!!!!!!!')
else:
print('DO NOT use HE for evaluate and save images!!!!!!!!!!')
logger.info(pad_str(' Performing evaluation on ImageNet-C'))
files_names = ['gaussian_noise', 'shot_noise', 'impulse_noise',#noise
'defocus_blur', 'glass_blur', 'motion_blur', 'zoom_blur', #blur
'frost', 'snow', 'fog', 'brightness', #whether
'contrast', 'elastic_transform', 'pixelate', 'jpeg_compression']#digital
for file_n in files_names:
print('Processing: ', file_n)
validate_ImagetNet_C(file_n, model, criterion, configs, logger)
return
def main():
# Scale and initialize the parameters
best_prec1 = 0
configs.TRAIN.epochs = int(
math.ceil(configs.TRAIN.epochs / configs.ADV.n_repeats))
configs.ADV.fgsm_step /= configs.DATA.max_color_value
configs.ADV.clip_eps /= configs.DATA.max_color_value
# Create output folder
if not os.path.isdir(os.path.join('trained_models', configs.output_name)):
os.makedirs(os.path.join('trained_models', configs.output_name))
# Log the config details
logger.info(pad_str(' ARGUMENTS '))
for k, v in configs.items():
print('{}: {}'.format(k, v))
logger.info(pad_str(''))
# Create the model
if configs.pretrained:
print("=> using pre-trained model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch](pretrained=True)
else:
print("=> creating model '{}'".format(configs.TRAIN.arch))
model = models.__dict__[configs.TRAIN.arch]()
# Use GPU or CPU
model = model.to(device)
# Criterion:
criterion = nn.CrossEntropyLoss().cuda()
# Optimizer:
optimizer = torch.optim.SGD(model.parameters(),
configs.TRAIN.lr,
momentum=configs.TRAIN.momentum,
weight_decay=configs.TRAIN.weight_decay)
# Resume if a valid checkpoint path is provided
if configs.resume:
if os.path.isfile(configs.resume):
print("=> loading checkpoint '{}'".format(configs.resume))
checkpoint = torch.load(configs.resume)
configs.TRAIN.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
configs.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(configs.resume))
# Data
print('==> Preparing data..')
trainloader, testloader = get_loaders("./data", configs.DATA.batch_size,
configs.DATA.workers,
configs.DATA.crop_size)
# If in evaluate mode: perform validation on PGD attacks as well as clean samples
if configs.evaluate:
logger.info(pad_str(' Performing PGD Attacks '))
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(testloader, model, criterion, pgd_param[0],
pgd_param[1], configs, logger)
validate(testloader, model, criterion, configs, logger)
return
start_train_time = time.time()
for epoch in range(configs.TRAIN.start_epoch, configs.TRAIN.epochs):
adjust_learning_rate(configs.TRAIN.lr, optimizer, epoch,
configs.ADV.n_repeats)
# train for one epoch
train(trainloader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(testloader, model, criterion, configs, logger)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': configs.TRAIN.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, os.path.join('trained_models', configs.output_name))
train_time = time.time()
print('Total train time: %.4f minutes',
(train_time - start_train_time) / 60)
# Automatically perform PGD Attacks at the end of training
logger.info(pad_str(' Performing PGD Attacks '))
'''
for pgd_param in configs.ADV.pgd_attack:
validate_pgd(testloader, model, criterion, pgd_param[0], pgd_param[1], configs, logger)
'''
train_time = time.time()
best_state_dict = model.state_dict()
torch.save(best_state_dict, os.path.join(args.out_dir, 'model.pth'))
logger.info('Total train time: %.4f minutes',
(train_time - start_train_time) / 60)
# Evaluation
model_test = models.__dict__[configs.TRAIN.arch]().to(device)
model_test.load_state_dict(best_state_dict)
model_test.float()
model_test.eval()
for pgd_param in configs.ADV.pgd_attack:
pgd_loss, pgd_acc = evaluate_pgd(testloader, model_test, pgd_param[0],
10)
test_loss, test_acc = evaluate_standard(testloader, model_test)
logger.info('Test Loss \t Test Acc \t PGD Loss \t PGD Acc')
logger.info('%.4f \t \t %.4f \t %.4f \t %.4f', test_loss, test_acc,
pgd_loss, pgd_acc)