next

natanielruiz

2017-09-21 9a02f63f4d5692399a95cb889e8f7629a165c28e

next

 code/datasets.py

@@ -38,7 +38,9 @@
        x_max = max(pt2d[0,:])
        y_max = max(pt2d[1,:])

        k = 0.35
        # k = 0.35 was being used beforehand
        # k = 0.2 to 0.40
        k = np.random.random_sample() * 0.2 + 0.2
        x_min -= 0.6 * k * abs(x_max - x_min)
        y_min -= 2 * k * abs(y_max - y_min)
        x_max += 0.6 * k * abs(x_max - x_min)
@@ -59,15 +61,10 @@
            roll = -roll
            img = img.transpose(Image.FLIP_LEFT_RIGHT)

        # Rotate?
        # rnd = np.random.random_sample()
        # if rnd < 0.5:
        #     if roll >= 0:
        #         img = img.rotate(30)
        #         roll -= 30
        #     else:
        #         img = img.rotate(-30)
        #         roll += 30
        # Blur?
        rnd = np.random.random_sample()
        if rnd < 0.05:
            img = img.filter(ImageFilter.BLUR)

        # Bin values
        bins = np.array(range(-99, 102, 3))
@@ -82,7 +79,7 @@
        if self.transform is not None:
            img = self.transform(img)

        return img, labels, cont_labels, elf.X_train[index]
        return img, labels, cont_labels, self.X_train[index]

    def __len__(self):
        # 122,450

 code/hopenet.py

@@ -125,88 +125,3 @@
            angles.append(self.fc_finetune(torch.cat((preangles, x), 1)))

        return pre_yaw, pre_pitch, pre_roll, angles

class Hopenet_shape(nn.Module):
    # This is just Hopenet with 3 output layers for yaw, pitch and roll.
    def __init__(self, block, layers, num_bins, shape_bins):
        self.inplanes = 64
        super(Hopenet_shape, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
        self.avgpool = nn.AvgPool2d(7)
        self.fc_yaw = nn.Linear(512 * block.expansion, num_bins)
        self.fc_pitch = nn.Linear(512 * block.expansion, num_bins)
        self.fc_roll = nn.Linear(512 * block.expansion, num_bins)
        self.fc_shape_0 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_1 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_2 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_3 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_4 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_5 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_6 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_7 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_8 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_shape_9 = nn.Linear(512 * block.expansion, shape_bins)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        yaw = self.fc_yaw(x)
        pitch = self.fc_pitch(x)
        roll = self.fc_roll(x)

        shape = []
        shape.append(self.fc_shape_0(x))
        shape.append(self.fc_shape_1(x))
        shape.append(self.fc_shape_2(x))
        shape.append(self.fc_shape_3(x))
        shape.append(self.fc_shape_4(x))
        shape.append(self.fc_shape_5(x))
        shape.append(self.fc_shape_6(x))
        shape.append(self.fc_shape_7(x))
        shape.append(self.fc_shape_8(x))
        shape.append(self.fc_shape_9(x))

        return yaw, pitch, roll, shape

 code/test_preangles.py

@@ -71,6 +71,8 @@
        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 == 'Pose_300W_LP':
        pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list, transformations)
    elif args.dataset == 'AFW':
        pose_dataset = datasets.AFW(args.data_dir, args.filename_list, transformations)
    else:

 code/train_preangles.py

@@ -18,6 +18,8 @@
import hopenet
import torch.utils.model_zoo as model_zoo

import time

model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
@@ -32,8 +34,6 @@
    parser.add_argument('--gpu', dest='gpu_id', help='GPU device id to use [0]',
            default=0, type=int)
    parser.add_argument('--num_epochs', dest='num_epochs', help='Maximum number of training epochs.',
          default=5, type=int)
    parser.add_argument('--num_epochs_ft', dest='num_epochs_ft', help='Maximum number of finetuning epochs.',
          default=5, type=int)
    parser.add_argument('--batch_size', dest='batch_size', help='Batch size.',
          default=16, type=int)
@@ -103,7 +103,6 @@

    cudnn.enabled = True
    num_epochs = args.num_epochs
    num_epochs_ft = args.num_epochs_ft
    batch_size = args.batch_size
    gpu = args.gpu_id

@@ -123,7 +122,6 @@
    transformations = transforms.Compose([transforms.Scale(240),
    transforms.RandomCrop(224), transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])


    if args.dataset == 'Pose_300W_LP':
        pose_dataset = datasets.Pose_300W_LP(args.data_dir, args.filename_list, transformations)
@@ -146,9 +144,9 @@
                                               num_workers=2)

    model.cuda(gpu)
    softmax = nn.Softmax()
    criterion = nn.CrossEntropyLoss().cuda()
    reg_criterion = nn.MSELoss().cuda()
    softmax = nn.Softmax().cuda(gpu)
    criterion = nn.CrossEntropyLoss().cuda(gpu)
    reg_criterion = nn.MSELoss().cuda(gpu)
    # Regression loss coefficient
    alpha = args.alpha

@@ -161,25 +159,26 @@
                                   lr = args.lr)

    print 'Ready to train network.'

    print 'First phase of training.'
    for epoch in range(num_epochs):
        start = time.time()
        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))
            print i
            print 'start: ', time.time() - start
            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))
            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()

            pre_yaw, pre_pitch, pre_roll, angles = model(images)

            # Cross entropy loss
            loss_yaw = criterion(pre_yaw, label_yaw)
            loss_pitch = criterion(pre_pitch, label_pitch)
@@ -198,7 +197,6 @@
            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
@@ -209,6 +207,8 @@
            torch.autograd.backward(loss_seq, grad_seq)
            optimizer.step()

            print 'end: ', time.time() - start

            if (i+1) % 100 == 0:
                print ('Epoch [%d/%d], Iter [%d/%d] Losses: Yaw %.4f, Pitch %.4f, Roll %.4f'
                       %(epoch+1, num_epochs, i+1, len(pose_dataset)//batch_size, loss_yaw.data[0], loss_pitch.data[0], loss_roll.data[0]))