import warnings
warnings.filterwarnings(action='ignore')
if __name__ == '__main__':
opt = parse_opts()
# Detect devices
use_cuda = torch.cuda.is_available() # check if GPU exists
device = torch.device("cuda" if use_cuda else "cpu") # use CPU or GPU
train_loader, valid_loader = dataLoadFunc(opt)
StudentModel = StudentModel(opt)
StudentModel = StudentModel.to(device)
smartModel = None
perturbation_model = None
if opt.smart_model:
opt.is_smart_model = True
smartModel = SmartModel(opt)
smartModel = smartModel.to(device)
# Parallelize model to multiple GPUs
if torch.cuda.device_count() > 1:
print("Using", torch.cuda.device_count(), "GPUs!")
StudentModel = nn.DataParallel(StudentModel)
parms = list(StudentModel.module.parameters())
if opt.isSource:
smartModel = nn.DataParallel(smartModel)
bridge = bridges.to(args.device)
t_model.eval()
elif args.task=='student':
model = StudentModel(params=args, pretrained_embedding=torch.tensor(pretrained_embedding).float())
param_groups = [{'params': model.parameters(), 'lr': args.learning_rate}]
elif args.task=='teacher':
model = TeacherModel(params=args, pretrained_embedding=torch.tensor(pretrained_embedding).float())
param_groups = [{'params': model.parameters(), 'lr': args.learning_rate}]
if args.classify_loss:
classifier = PathClassifier(params=args)
param_groups.append({'params': classifier.parameters(),'lr':args.learning_rate})
classifiers = classifier.to(args.device)
optimizer = SGD(param_groups, lr=args.learning_rate)
scheduler = lr_scheduler.StepLR(optimizer, step_size=1, gamma=args.gamma)
model = model.to(args.device)
total_step = 0
eval_result = {}
accum_train_link_loss, accum_train_label_loss = 0, 0
accum_distill_loss, accum_classify_loss = 0, 0
accum_eval_loss = 0
scheduler_step = 0
best_eval_result = None
stop_sign=0
for epoch in range(args.epoches):
print('{} epoch training..'.format(epoch + 1))
print('dialogue model learning rate {:.4f}'.format(optimizer.param_groups[0]['lr']))
model.train()
for batch in tqdm(train_dataloader):