def main(): global args, best_prec_result True_loader, Fake_loader, Noise_loader, Noise_Test_loader, Noise_Sample_loader, Noise_Triple_loader, All_loader, Test_loader, chIn, clsN = dataset_selector( args.db) args.chIn = chIn args.clsN = clsN args.milestones = [80, 120] args.Dmilestones = [30, 60] state_info = utils.model_optim_state_info() state_info.model_init(args) state_info.model_cuda_init() if cuda: # os.environ["CUDA_VISIBLE_DEVICES"] = '0' print("USE", torch.cuda.device_count(), "GPUs!") state_info.weight_cuda_init() cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(args) train_Sample(args, state_info, Noise_Sample_loader, Noise_Test_loader)
def main(): global args, best_prec_result Train_loader, Test_loader, chIn, clsN = dataset_selector(args.db) AnchorSet = dataset.Cifar10_Sample(args) args.chIn = chIn args.clsN = clsN if not args.milestones: args.milestones = [250, 400] state_info = utils.model_optim_state_info() state_info.model_init(args) state_info.model_cuda_init() if cuda: # os.environ["CUDA_VISIBLE_DEVICES"] = '0' print("USE", torch.cuda.device_count(), "GPUs!") state_info.weight_cuda_init() cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(args) train_MEM(args, state_info, Train_loader, Test_loader, AnchorSet)
def main(): global args, best_prec_result start_epoch = 0 utils.default_model_dir = args.dir start_time = time.time() train_loader, test_loader, _, _ = dataset_selector(args.sd) state_info = utils.model_optim_state_info() state_info.model_init(args=args, num_class=10) state_info.model_cuda_init() # state_info.weight_init() state_info.optimizer_init(args) if cuda: print("USE", torch.cuda.device_count(), "GPUs!") cudnn.benchmark = True checkpoint = utils.load_checkpoint(utils.default_model_dir, is_last=False) if checkpoint: start_epoch = checkpoint['epoch'] + 1 best_prec_result = checkpoint['Best_Prec'] state_info.load_state_dict(checkpoint) check_selection(state_info, train_loader) check_selection(state_info, test_loader)
def main(): global args, best_prec_result start_epoch = 0 utils.default_model_dir = args.dir start_time = time.time() train_loader, test_loader, ch, wh = dataset_selector(args.dataset) sample = extract_sample(train_loader) state_info = utils.model_optim_state_info() state_info.model_init(Img=[ch, wh], H=args.h, latent_size=args.latent_size, num_class=10) state_info.model_cuda_init() state_info.weight_init() state_info.optimizer_init(args) if cuda: print("USE", torch.cuda.device_count(), "GPUs!") cudnn.benchmark = True state_info.learning_scheduler_init(args) for epoch in range(start_epoch, args.epoch): train(state_info, train_loader, epoch) test(state_info, test_loader, sample, epoch) state_info.learning_step() now = time.gmtime(time.time() - start_time) utils.print_log('{} hours {} mins {} secs for training'.format( now.tm_hour, now.tm_min, now.tm_sec))
def main(): global args, best_prec_result utils.default_model_dir = args.dir start_time = time.time() Source_train_loader, Source_test_loader = dataset_selector(args.sd) Target_train_loader, Target_test_loader = dataset_selector(args.td) Target_shuffle_loader, _ = dataset_selector(args.td) state_info = utils.model_optim_state_info() state_info.model_init() state_info.model_cuda_init() if cuda: # os.environ["CUDA_VISIBLE_DEVICES"] = '0' print("USE", torch.cuda.device_count(), "GPUs!") state_info.weight_cuda_init() cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(lr=args.lr, b1=args.b1, b2=args.b2, weight_decay=args.weight_decay) start_epoch = 0 checkpoint = utils.load_checkpoint(utils.default_model_dir) if not checkpoint: state_info.learning_scheduler_init(args) else: start_epoch = checkpoint['epoch'] + 1 best_prec_result = checkpoint['Best_Prec'] state_info.load_state_dict(checkpoint) state_info.learning_scheduler_init(args, load_epoch=start_epoch) realS_sample_iter = iter(Source_train_loader) realT_sample_iter = iter(Target_train_loader) realS_sample = to_var(realS_sample_iter.next()[0], FloatTensor) realT_sample = to_var(realT_sample_iter.next()[0], FloatTensor) for epoch in range(args.epoch): train(state_info, Source_train_loader, Target_train_loader, Target_shuffle_loader, epoch) prec_result = test(state_info, Source_test_loader, Target_test_loader, realS_sample, realT_sample, epoch) if prec_result > best_prec_result: best_prec_result = prec_result filename = 'checkpoint_best.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) filename = 'latest.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) state_info.learning_step() now = time.gmtime(time.time() - start_time) utils.print_log('{} hours {} mins {} secs for training'.format(now.tm_hour, now.tm_min, now.tm_sec))
def main(): global args, best_prec_result start_epoch = 0 utils.default_model_dir = args.dir start_time = time.time() train_loader, test_loader, _, _ = dataset_selector(args.sd) state_info = utils.model_optim_state_info() state_info.model_init(args=args, num_class=10) state_info.model_cuda_init() # state_info.weight_init() state_info.optimizer_init(args) if cuda: print("USE", torch.cuda.device_count(), "GPUs!") cudnn.benchmark = True checkpoint = utils.load_checkpoint(utils.default_model_dir, is_last=True) if checkpoint: start_epoch = checkpoint['epoch'] + 1 best_prec_result = checkpoint['Best_Prec'] state_info.load_state_dict(checkpoint) for epoch in range(0, args.epoch): if epoch < 80: lr = args.lr elif epoch < 120: lr = args.lr * 0.1 else: lr = args.lr * 0.01 for param_group in state_info.optimizer.param_groups: param_group['lr'] = lr train(state_info, train_loader, epoch) prec_result = test(state_info, test_loader, epoch) if prec_result > best_prec_result: best_prec_result = prec_result filename = 'checkpoint_best.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) utils.print_log('Best Prec : {:.4f}'.format( best_prec_result.item())) filename = 'latest.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) now = time.gmtime(time.time() - start_time) utils.print_log('Best Prec : {:.4f}'.format(best_prec_result.item())) utils.print_log('{} hours {} mins {} secs for training'.format( now.tm_hour, now.tm_min, now.tm_sec)) print('done')
def main(): global args, best_prec_result train_labeled_dataset, train_unlabeled_dataset, test_dataset = dataset_selector() args.clsN = 10 args.milestones = [80,120] state_info = utils.model_optim_state_info() state_info.model_init(args) state_info.model_cuda_init() if cuda: # os.environ["CUDA_VISIBLE_DEVICES"] = '0' print("USE", torch.cuda.device_count(), "GPUs!") state_info.weight_cuda_init() cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(args) train_MEM(args, state_info, train_labeled_dataset, train_unlabeled_dataset, test_dataset)
def main(): global args, best_prec_result Train_loader, Test_loader, chIn, clsN = dataset_selector(args.db) args.chIn = chIn args.clsN = clsN # args.milestones = [150,225] state_info = utils.model_optim_state_info() state_info.model_init(args) if args.use_switch: utils.init_learning(state_info.model.module) if cuda: print("USE", torch.cuda.device_count(), "GPUs!") cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(args) train_Epoch(args, state_info, Train_loader, Test_loader)
def main(): global args, best_prec_result utils.default_model_dir = args.dir start_time = time.time() Source_train_loader, Source_test_loader = dataset_selector(args.sd) Target_train_loader, Target_test_loader = dataset_selector(args.td) state_info = utils.model_optim_state_info() state_info.model_init() state_info.model_cuda_init() if cuda: # os.environ["CUDA_VISIBLE_DEVICES"] = '0' print("USE", torch.cuda.device_count(), "GPUs!") state_info.weight_cuda_init() cudnn.benchmark = True else: print("NO GPU") state_info.optimizer_init(lr=args.lr, b1=args.b1, b2=args.b2, weight_decay=args.weight_decay) adversarial_loss = torch.nn.BCELoss() criterion = nn.CrossEntropyLoss().cuda() start_epoch = 0 utils.default_model_dir filename = 'latest.pth.tar' checkpoint = utils.load_checkpoint(utils.default_model_dir) if not checkpoint: pass else: start_epoch = checkpoint['epoch'] + 1 best_prec_result = checkpoint['Best_Prec'] state_info.load_state_dict(checkpoint) numEpochs = int( math.ceil( float(args.train_iters) / float(min(len(Source_train_loader), len(Target_train_loader))))) for epoch in range(numEpochs): # if epoch < 80: # learning_rate = args.lr # elif epoch < 120: # learning_rate = args.lr * 0.1 # else: # learning_rate = args.lr * 0.01 # for param_group in optimizer.param_groups: # param_group['lr'] = learning_rate train(state_info, Source_train_loader, Target_train_loader, criterion, adversarial_loss, epoch) prec_result = test(state_info, Source_test_loader, Target_test_loader, criterion, epoch) if prec_result > best_prec_result: best_prec_result = prec_result filename = 'checkpoint_best.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) if epoch % 5 == 0: filename = 'latest.pth.tar' utils.save_state_checkpoint(state_info, best_prec_result, filename, utils.default_model_dir, epoch) now = time.gmtime(time.time() - start_time) utils.print_log('{} hours {} mins {} secs for training'.format( now.tm_hour, now.tm_min, now.tm_sec))