cfg = Config(args=args.__dict__, mode='Train') cfg.show_config() torch.backends.cudnn.benchmark = True train_dataset = Seg_dataset(cfg) train_loader = data.DataLoader(train_dataset, batch_size=cfg.bs, shuffle=True, num_workers=8, pin_memory=True, drop_last=False) if cfg.model == 'unet': model = UNet(input_channels=3).cuda() model.apply(model.weights_init_normal) else: model = DLASeg(cfg).cuda() model.train() if cfg.resume: resume_epoch = int(cfg.resume.split('.')[0].split('_')[1]) + 1 model.load_state_dict(torch.load('weights/' + cfg.resume), strict=True) print(f'Resume training with \'{cfg.resume}\'.') else: resume_epoch = 0 print('Training with ImageNet pre-trained weights.') criterion = nn.CrossEntropyLoss(ignore_index=255).cuda() if cfg.optim == 'sgd': optimizer = torch.optim.SGD(model.optim_parameters(),
def train(frame_num, layer_nums, input_channels, output_channels, discriminator_num_filters, bn=False, pretrain=False, generator_pretrain_path=None, discriminator_pretrain_path=None): generator = UNet(n_channels=input_channels, layer_nums=layer_nums, output_channel=output_channels, bn=bn) discriminator = PixelDiscriminator(output_channels, discriminator_num_filters, use_norm=False) generator = generator.cuda() discriminator = discriminator.cuda() flow_network = Network() flow_network.load_state_dict(torch.load(lite_flow_model_path)) flow_network.cuda().eval() adversarial_loss = Adversarial_Loss().cuda() discriminate_loss = Discriminate_Loss().cuda() gd_loss = Gradient_Loss(alpha, num_channels).cuda() op_loss = Flow_Loss().cuda() int_loss = Intensity_Loss(l_num).cuda() step = 0 if not pretrain: generator.apply(weights_init_normal) discriminator.apply(weights_init_normal) else: assert (generator_pretrain_path != None and discriminator_pretrain_path != None) generator.load_state_dict(torch.load(generator_pretrain_path)) discriminator.load_state_dict(torch.load(discriminator_pretrain_path)) step = int(generator_pretrain_path.split('-')[-1]) print('pretrained model loaded!') print('initializing the model with Generator-Unet {} layers,' 'PixelDiscriminator with filters {} '.format( layer_nums, discriminator_num_filters)) optimizer_G = torch.optim.Adam(generator.parameters(), lr=g_lr) optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=d_lr) writer = SummaryWriter(writer_path) dataset = img_dataset.ano_pred_Dataset(training_data_folder, frame_num) dataset_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True, num_workers=1, drop_last=True) test_dataset = img_dataset.ano_pred_Dataset(testing_data_folder, frame_num) test_dataloader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=True, num_workers=1, drop_last=True) for epoch in range(epochs): for (input, _), (test_input, _) in zip(dataset_loader, test_dataloader): # generator = generator.train() # discriminator = discriminator.train() target = input[:, -1, :, :, :].cuda() input = input[:, :-1, ] input_last = input[:, -1, ].cuda() input = input.view(input.shape[0], -1, input.shape[-2], input.shape[-1]).cuda() test_target = test_input[:, -1, ].cuda() test_input = test_input[:, :-1].view(test_input.shape[0], -1, test_input.shape[-2], test_input.shape[-1]).cuda() #------- update optim_G -------------- G_output = generator(input) pred_flow_esti_tensor = torch.cat([input_last, G_output], 1) gt_flow_esti_tensor = torch.cat([input_last, target], 1) flow_gt = batch_estimate(gt_flow_esti_tensor, flow_network) flow_pred = batch_estimate(pred_flow_esti_tensor, flow_network) g_adv_loss = adversarial_loss(discriminator(G_output)) g_op_loss = op_loss(flow_pred, flow_gt) g_int_loss = int_loss(G_output, target) g_gd_loss = gd_loss(G_output, target) g_loss = lam_adv * g_adv_loss + lam_gd * g_gd_loss + lam_op * g_op_loss + lam_int * g_int_loss optimizer_G.zero_grad() g_loss.backward() optimizer_G.step() train_psnr = psnr_error(G_output, target) #----------- update optim_D ------- optimizer_D.zero_grad() d_loss = discriminate_loss(discriminator(target), discriminator(G_output.detach())) #d_loss.requires_grad=True d_loss.backward() optimizer_D.step() #----------- cal psnr -------------- test_generator = generator.eval() test_output = test_generator(test_input) test_psnr = psnr_error(test_output, test_target).cuda() if step % 10 == 0: print("[{}/{}]: g_loss: {} d_loss {}".format( step, epoch, g_loss, d_loss)) print('\t gd_loss {}, op_loss {}, int_loss {} ,'.format( g_gd_loss, g_op_loss, g_int_loss)) print('\t train psnr{},test_psnr {}'.format( train_psnr, test_psnr)) writer.add_scalar('psnr/train_psnr', train_psnr, global_step=step) writer.add_scalar('psnr/test_psnr', test_psnr, global_step=step) writer.add_scalar('total_loss/g_loss', g_loss, global_step=step) writer.add_scalar('total_loss/d_loss', d_loss, global_step=step) writer.add_scalar('g_loss/adv_loss', g_adv_loss, global_step=step) writer.add_scalar('g_loss/op_loss', g_op_loss, global_step=step) writer.add_scalar('g_loss/int_loss', g_int_loss, global_step=step) writer.add_scalar('g_loss/gd_loss', g_gd_loss, global_step=step) writer.add_image('image/train_target', target[0], global_step=step) writer.add_image('image/train_output', G_output[0], global_step=step) writer.add_image('image/test_target', test_target[0], global_step=step) writer.add_image('image/test_output', test_output[0], global_step=step) step += 1 if step % 500 == 0: utils.saver(generator.state_dict(), model_generator_save_path, step, max_to_save=10) utils.saver(discriminator.state_dict(), model_discriminator_save_path, step, max_to_save=10) if step >= 2000: print('==== begin evaluate the model of {} ===='.format( model_generator_save_path + '-' + str(step))) auc = evaluate(frame_num=5, layer_nums=4, input_channels=12, output_channels=3, model_path=model_generator_save_path + '-' + str(step), evaluate_name='compute_auc') writer.add_scalar('results/auc', auc, global_step=step)
train_cfg = update_config(args, mode='train') train_cfg.print_cfg() generator = UNet(input_channels=12, output_channel=3).cuda() discriminator = PixelDiscriminator(input_nc=3).cuda() optimizer_G = torch.optim.Adam(generator.parameters(), lr=train_cfg.g_lr) optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=train_cfg.d_lr) if train_cfg.resume: generator.load_state_dict(torch.load(train_cfg.resume)['net_g']) discriminator.load_state_dict(torch.load(train_cfg.resume)['net_d']) optimizer_G.load_state_dict(torch.load(train_cfg.resume)['optimizer_g']) optimizer_D.load_state_dict(torch.load(train_cfg.resume)['optimizer_d']) print(f'Pre-trained generator and discriminator have been loaded.\n') else: generator.apply(weights_init_normal) discriminator.apply(weights_init_normal) print('Generator and discriminator are going to be trained from scratch.\n') assert train_cfg.flownet in ('lite', '2sd'), 'Flow net only supports LiteFlownet or FlowNet2SD currently.' if train_cfg.flownet == '2sd': flow_net = FlowNet2SD() flow_net.load_state_dict(torch.load('models/flownet2/FlowNet2-SD.pth')['state_dict']) else: flow_net = lite_flow.Network() flow_net.load_state_dict(torch.load('models/liteFlownet/network-default.pytorch')) flow_net.cuda().eval() # Use flow_net to generate optic flows, so set to eval mode. adversarial_loss = Adversarial_Loss().cuda() discriminate_loss = Discriminate_Loss().cuda()