def build(self):
if self.config['model']['type'] != 'CNN':
raise ValueError('Model type is not valid.')
encoder, decoder = self.config['model']['name'].upper().split('_')
self.encoder = make_encoder(e_config[encoder])
self.decoder = make_decoder(d_config[decoder])
def build(self):
if self.config['model']['type'] != 'RNN':
raise ValueError('Model type is not valid.')
self.e_name, self.d_name = self.config['model']['name'].upper().split(
'_')
self.encoder = make_encoder(e_config[self.e_name], self.config).cuda()
self.decoder = make_decoder(d_config[self.d_name]).cuda()
def main():
global args, best_prec1, model, train_loader, val_loaders, output, target
args = parser.parse_args()
logging.basicConfig(
format="%(message)s",
handlers=[
# logging.FileHandler("{0}/{1}.log".format(args.log, sys.argv[0].replace('.py','') + datetime.now().strftime('_%H_%M_%d_%m_%Y'))),
logging.FileHandler("{0}/{1}.log".format(
args.log, sys.argv[0].replace('.py', '') + args.name)),
logging.StreamHandler()
],
level=logging.INFO)
for sigma in range(6, 16, 2):
# create model
if args.pretrained:
pass
# logging.info("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True)
else:
model_name = CONFIG['model']['name']
en_name = model_name.split('_')[0]
logging.info("=> creating model '{}'".format(en_name))
model = make_encoder(en_name)
model._initialize_weights()
# model.load_imagenet_weights()
# CONFIG['train']['user'] = [8]
# model = Encoder(CONFIG)
# model._initialize_weights()
# model.encoder.load_weights()
model.cuda()
# define loss function (criterion) and optimizer
#criterion = nn.BCEWithLogitsLoss().cuda()
#criterion = nn.BCELoss().cuda()
criterion = nn.KLDivLoss()
# optimizer = torch.optim.SGD(
# list(model.encoder[1].parameters()) +
# list(model.readout.parameters()),
# # model.parameters(),
# args.lr,
# momentum=args.momentum,
# weight_decay=args.weight_decay)
# for param in model.encoder[0].parameters():
# param.requires_grad = False
optimizer = torch.optim.Adam(
# list(model.encoder[1].parameters()) +
# list(model.readout.parameters()),
model.parameters(),
args.lr,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
cudnn.benchmark = True
if args.resume:
if os.path.isfile(args.resume):
logging.info("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
logging.info("=> no checkpoint found at '{}'".format(
args.resume))
train_duration = True if (args.name == 'TDW') else False
# eval_duration = True if (args.name.split('-')[1] == 'GTDW') else False
print(args.name, train_duration) #, eval_duration)
train_dataset = Saliency()
train_dataset.load('OSIE',
'train',
duration=train_duration,
split={
'train': 0.90,
'eval': 0.1
},
sigma=sigma)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=None)
# val_loaders = list()
# for name_idx, name in enumerate(CONFIG['eval']['dataset']):
# if name == 'OSIE':
# split={'train' :0.75 , 'eval' : 0.25}
# else:
# split={'train' :0.0 , 'eval' : 1.0}
# val_ds = Saliency()
# # val_ds.load(name, 'eval', split=split,
# # duration=eval_duration, sigma=sigma)
# val_ds.load(name, 'eval', split=split,
# duration=eval_duration, sigma=sigma)
# val_loaders.append(
# torch.utils.data.DataLoader(
# val_ds,
# batch_size=args.batch_size * 2, shuffle=False,
# num_workers=args.workers, pin_memory=True)
# )
if args.evaluate:
validate(val_loaders, model, criterion)
return
# if args.visualize:
# visualize(val_loader, model)
# return
for epoch in range(args.start_epoch, args.epochs):
try:
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, sigma)
# evaluate on validation set
# prec1 = validate(val_loaders, model, criterion, epoch)
# remember best prec@1 and save checkpoint
# is_best = prec1 > best_prec1
# best_prec1 = max(prec1, best_prec1)
#best_prec1 = 1
if epoch in [4, 9, 14]:
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': True,
'optimizer': optimizer.state_dict(),
}, True, sigma)
except Exception as x:
print(x)
def main():
global args, best_prec1, model, train_loader, val_loaders, output, target
args = parser.parse_args()
logging.basicConfig(format="%(message)s",
handlers=[
logging.FileHandler("{0}/{1}.log".format(
args.log, sys.argv[0].replace('.py', '') +
datetime.now().strftime('_%H_%M_%d_%m_%Y'))),
logging.StreamHandler()
],
level=logging.INFO)
# create model
if args.pretrained:
pass
# logging.info("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True)
else:
model_name = CONFIG['model']['name']
en_name = model_name.split('_')[0]
logging.info("=> creating model '{}'".format(en_name))
model = make_encoder(en_name)
model._initialize_weights()
# model.load_imagenet_weights()
# CONFIG['train']['user'] = [8]
# model = Encoder(CONFIG)
# model._initialize_weights()
# model.encoder.load_weights()
model.cuda()
# define loss function (criterion) and optimizer
#criterion = nn.BCEWithLogitsLoss().cuda()
criterion = nn.BCELoss().cuda()
# optimizer = torch.optim.SGD(
# list(model.encoder[1].parameters()) +
# list(model.readout.parameters()),
# # model.parameters(),
# args.lr,
# momentum=args.momentum,
# weight_decay=args.weight_decay)
# for param in model.encoder[0].parameters():
# param.requires_grad = False
optimizer = torch.optim.Adam(
# list(model.encoder[1].parameters()) +
# list(model.readout.parameters()),
model.parameters(),
args.lr,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
cudnn.benchmark = True
if args.resume:
if os.path.isfile(args.resume):
logging.info("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
logging.info("=> no checkpoint found at '{}'".format(args.resume))
# train_duration = True if (args.name.split('-')[0] == 'TDW') else False
# eval_duration = True if (args.name.split('-')[1] == 'GTDW') else False
# print(args.name, train_duration, eval_duration)
# train_dataset = Saliency()
# train_dataset.load('OSIE', 'train', duration=train_duration)
# train_loader = torch.utils.data.DataLoader(
# train_dataset, batch_size=args.batch_size, shuffle=False,
# num_workers=args.workers, pin_memory=True, sampler=None)
# val_loaders = list()
# for name_idx, name in enumerate(CONFIG['eval']['dataset']):
# # if name == 'OSIE':
# # split={'train' :0.75 , 'eval' : 0.25}
# # else:
# split={'train' :0.0 , 'eval' : 1.0}
# val_ds = Saliency()
# val_ds.load(name, 'eval', split=split,
# duration=eval_duration)
# val_loaders.append(
# torch.utils.data.DataLoader(
# val_ds,
# batch_size=args.batch_size * 2, shuffle=False,
# num_workers=args.workers, pin_memory=True)
# )
# if args.evaluate:
# validate(val_loaders, model, criterion)
# return
# if args.visualize:
# visualize(val_loader, model)
# return
# policies = [
# [
# 'TS-GTS-5.pth.tar',
# 'TS-GTS-10.pth.tar',
# 'TS-GTS-15.pth.tar'],
# [
# 'TS-GTDW-5.pth.tar',
# 'TS-GTDW-10.pth.tar',
# 'TS-GTDW-15.pth.tar'],
# [
# 'TDW-GTS-5.pth.tar',
# 'TDW-GTS-10.pth.tar',
# 'TDW-GTS-15.pth.tar'],
# [
# 'TDW-GTDW-5.pth.tar',
# 'TDW-GTDW-10.pth.tar',
# 'TDW-GTDW-15.pth.tar'],
# ]
policies = [
'TS-{0}-{1}.pth.tar',
'TDW-{0}-{1}.pth.tar',
]
for train_policy_idx, train_policy in enumerate(policies):
for eval_policy_idx, eval_policy in enumerate(['GTS', 'GTDW']):
for sigma in range(6, 15, 2):
val_loaders = list()
# model_name = '-'.join(policy[0].split('-')[:2])
# eval_duration = True if (model_name.split('-')[1] == 'GTDW') else False
# print('eval_duration: {0}'.format(eval_duration))
# model_name = '-'.join(policy[0].split('-')[:2])
eval_duration = True if (eval_policy == 'GTDW') else False
for name_idx, name in enumerate(CONFIG['eval']['dataset']):
# if name == 'OSIE':
# split={'train' :0.75 , 'eval' : 0.25}
# else:
split = {'train': 0.0, 'eval': 1.0}
val_ds = Saliency()
val_ds.load(name,
'eval',
split=split,
duration=eval_duration,
sigma=sigma)
val_loaders.append(
torch.utils.data.DataLoader(
val_ds,
batch_size=args.batch_size * 2,
shuffle=False,
num_workers=args.workers,
pin_memory=True))
for epoch in ([15]):
weight = train_policy.format(epoch, sigma)
logging.info('eval_duration: {0} - {1}'.format(
weight, eval_policy))
try:
weight = os.path.join(args.weights, weight)
# adjust_learning_rate(optimizer, epoch)
model = load_check_point(model, weight)
prec1 = validate(
val_loaders, model, criterion,
(train_policy_idx, eval_policy_idx, sigma, epoch))
with open('results/{0}.npy'.format('TOTAL'),
'wb') as f:
np.save(f, np.array(RESULTS))
except Exception as x:
print(x)
def main():
global args, best_prec1, model, train_dataset, val_loader, en_name
args = parser.parse_args()
logging.basicConfig(format="%(message)s",
handlers=[
logging.FileHandler("{0}/{1}.log".format(
args.log, sys.argv[0].replace('.py', '') +
datetime.now().strftime('_%H_%M_%d_%m_%Y'))),
logging.StreamHandler()
],
level=logging.INFO)
# create model
if args.pretrained:
pass
# logging.info("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True)
else:
model_name = CONFIG['model']['name']
en_name = model_name.split('_')[0]
logging.info("=> creating model '{}'".format(en_name))
model = make_encoder(e_config[en_name], CONFIG)
# model._initialize_weights()
model.cuda()
# define loss function (criterion) and optimizer
criterion = nn.BCELoss().cuda()
# optimizer = torch.optim.SGD(model.parameters(), args.lr,
# momentum=args.momentum,
# weight_decay=args.weight_decay)
optimizer = torch.optim.Adam(model.parameters(),
args.lr,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
cudnn.benchmark = True
# Data loading code
train_dataset = Saliency(CONFIG, 'train')
if args.resume:
if os.path.isfile(args.resume):
logging.info("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer'])
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
logging.info("=> no checkpoint found at '{}'".format(args.resume))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=None)
val_loader = torch.utils.data.DataLoader(Saliency(CONFIG, 'test'),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
if args.visualize:
visualize(val_loader, model)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best)