Failed lstm experiment

natanielruiz

2017-09-27 43416c4717d2430c3e11f042294d12b781fee2e1

 code/train.py

@@ -46,6 +46,9 @@
    parser.add_argument('--output_string', dest='output_string', help='String appended to output snapshots.', default = '', type=str)
    parser.add_argument('--alpha', dest='alpha', help='Regression loss coefficient.',
          default=0.001, type=float)
    parser.add_argument('--iter_ref', dest='iter_ref', help='Number of iterative refinement passes.',
          default=1, type=int)
    parser.add_argument('--dataset', dest='dataset', help='Dataset type.', default='Pose_300W_LP', type=str)
    args = parser.parse_args()
    return args

@@ -111,7 +114,7 @@
    # ResNet101 with 3 outputs
    # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66)
    # ResNet50
    model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66)
    model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66, args.iter_ref)
    # ResNet18
    # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66)
    load_filtered_state_dict(model, model_zoo.load_url(model_urls['resnet50']))
@@ -122,8 +125,21 @@
    transforms.RandomCrop(224), transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])

    pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list,
                                transformations)
    if args.dataset == 'Pose_300W_LP':
        pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'AFLW2000':
        pose_dataset = datasets.AFLW2000(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'BIWI':
        pose_dataset = datasets.BIWI(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'AFLW':
        pose_dataset = datasets.AFLW(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'AFLW_aug':
        pose_dataset = datasets.AFLW_aug(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'AFW':
        pose_dataset = datasets.AFW(args.data_dir, args.filename_list, transformations)
    else:
        print 'Error: not a valid dataset name'
        sys.exit()
    train_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
                                               batch_size=batch_size,
                                               shuffle=True,
@@ -133,6 +149,7 @@
    softmax = nn.Softmax()
    criterion = nn.CrossEntropyLoss().cuda()
    reg_criterion = nn.MSELoss().cuda()
    smooth_l1_loss = nn.SmoothL1Loss().cuda()
    # Regression loss coefficient
    alpha = args.alpha

@@ -141,18 +158,23 @@

    optimizer = torch.optim.Adam([{'params': get_ignored_params(model), 'lr': 0},
                                  {'params': get_non_ignored_params(model), 'lr': args.lr},
                                  {'params': get_fc_params(model), 'lr': args.lr * 2}],
                                  {'params': get_fc_params(model), 'lr': args.lr * 5}],
                                   lr = args.lr)

    print 'Ready to train network.'

    print 'First phase of training.'
    for epoch in range(num_epochs):
        for i, (images, labels, name) in enumerate(train_loader):
        for i, (images, labels, cont_labels, name) in enumerate(train_loader):
            images = Variable(images.cuda(gpu))
            label_yaw = Variable(labels[:,0].cuda(gpu))
            label_pitch = Variable(labels[:,1].cuda(gpu))
            label_roll = Variable(labels[:,2].cuda(gpu))

            label_angles = Variable(cont_labels[:,:3].cuda(gpu))
            label_yaw_cont = Variable(cont_labels[:,0].cuda(gpu))
            label_pitch_cont = Variable(cont_labels[:,1].cuda(gpu))
            label_roll_cont = Variable(cont_labels[:,2].cuda(gpu))

            optimizer.zero_grad()
            model.zero_grad()
@@ -169,22 +191,20 @@
            pitch_predicted = softmax(pre_pitch)
            roll_predicted = softmax(pre_roll)

            yaw_predicted = torch.sum(yaw_predicted * idx_tensor, 1)
            pitch_predicted = torch.sum(pitch_predicted * idx_tensor, 1)
            roll_predicted = torch.sum(roll_predicted * idx_tensor, 1)
            yaw_predicted = torch.sum(yaw_predicted * idx_tensor, 1) * 3 - 99
            pitch_predicted = torch.sum(pitch_predicted * idx_tensor, 1) * 3 - 99
            roll_predicted = torch.sum(roll_predicted * idx_tensor, 1) * 3 - 99

            loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw.float())
            loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch.float())
            loss_reg_roll = reg_criterion(roll_predicted, label_roll.float())
            loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw_cont)
            loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch_cont)
            loss_reg_roll = reg_criterion(roll_predicted, label_roll_cont)

            # print yaw_predicted, label_yaw.float(), loss_reg_yaw
            # Total loss
            loss_yaw += alpha * loss_reg_yaw
            loss_pitch += alpha * loss_reg_pitch
            loss_roll += alpha * loss_reg_roll

            loss_seq = [loss_yaw, loss_pitch, loss_roll]
            # loss_seq = [loss_reg_yaw, loss_reg_pitch, loss_reg_roll]
            grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
            torch.autograd.backward(loss_seq, grad_seq)
            optimizer.step()
@@ -204,12 +224,16 @@

    print 'Second phase of training (finetuning layer).'
    for epoch in range(num_epochs_ft):
        for i, (images, labels, name) in enumerate(train_loader):
        for i, (images, labels, cont_labels, name) in enumerate(train_loader):
            images = Variable(images.cuda(gpu))
            label_yaw = Variable(labels[:,0].cuda(gpu))
            label_pitch = Variable(labels[:,1].cuda(gpu))
            label_roll = Variable(labels[:,2].cuda(gpu))
            label_angles = Variable(labels[:,:3].cuda(gpu))

            label_angles = Variable(cont_labels[:,:3].cuda(gpu))
            label_yaw_cont = Variable(cont_labels[:,0].cuda(gpu))
            label_pitch_cont = Variable(cont_labels[:,1].cuda(gpu))
            label_roll_cont = Variable(cont_labels[:,2].cuda(gpu))

            optimizer.zero_grad()
            model.zero_grad()
@@ -226,13 +250,13 @@
            pitch_predicted = softmax(pre_pitch)
            roll_predicted = softmax(pre_roll)

            yaw_predicted = torch.sum(yaw_predicted.data * idx_tensor, 1)
            pitch_predicted = torch.sum(pitch_predicted.data * idx_tensor, 1)
            roll_predicted = torch.sum(roll_predicted.data * idx_tensor, 1)
            yaw_predicted = torch.sum(yaw_predicted * idx_tensor, 1) * 3 - 99
            pitch_predicted = torch.sum(pitch_predicted * idx_tensor, 1) * 3 - 99
            roll_predicted = torch.sum(roll_predicted * idx_tensor, 1) * 3 - 99

            loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw.float())
            loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch.float())
            loss_reg_roll = reg_criterion(roll_predicted, label_roll.float())
            loss_reg_yaw = reg_criterion(yaw_predicted, label_yaw_cont)
            loss_reg_pitch = reg_criterion(pitch_predicted, label_pitch_cont)
            loss_reg_roll = reg_criterion(roll_predicted, label_roll_cont)

            # Total loss
            loss_yaw += alpha * loss_reg_yaw
@@ -240,9 +264,17 @@
            loss_roll += alpha * loss_reg_roll

            # Finetuning loss
            loss_angles = reg_criterion(angles[0], label_angles.float())
            loss_seq = [loss_yaw, loss_pitch, loss_roll]
            for idx in xrange(1,len(angles)):
                label_angles_residuals = label_angles - angles[0] * 3 - 99
                for idy in xrange(1,idx):
                    label_angles_residuals += angles[idy] * 3 - 99
                label_angles_residuals = label_angles_residuals.detach()
                # Reconvert to other unit
                label_angles_residuals = label_angles_residuals / 3.0 + 33
                loss_angles = smooth_l1_loss(angles[idx], label_angles_residuals)
                loss_seq.append(loss_angles)

            loss_seq = [loss_yaw, loss_pitch, loss_roll, loss_angles]
            grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
            torch.autograd.backward(loss_seq, grad_seq)
            optimizer.step()
@@ -255,11 +287,7 @@
                #     'output/snapshots/' + args.output_string + '_iter_'+ str(i+1) + '.pkl')

        # Save models at numbered epochs.
        if epoch % 1 == 0 and epoch < num_epochs_ft - 1:
        if epoch % 1 == 0 and epoch < num_epochs_ft:
            print 'Taking snapshot...'
            torch.save(model.state_dict(),
            'output/snapshots/' + args.output_string + '_epoch_'+ str(num_epochs+epoch+1) + '.pkl')


    # Save the final Trained Model
    torch.save(model.state_dict(), 'output/snapshots/' + args.output_string + '_epoch_' + str(num_epochs+epoch+1) + '.pkl')