Training for hopenet works.

natanielruiz

2017-07-07 6664c6d52fad58e396861946a3bed7d5afc4d44d

Training for hopenet works.

 code/datasets.py

@@ -84,10 +84,51 @@

        # We get the pose in radians
        pose = utils.get_ypr_from_mat(os.path.join(self.data_dir, self.y_train[index] + self.annot_ext))
        # And convert to positive degrees.
        pose = pose * 180 / np.pi + 90
        # And convert to degrees.
        pitch = pose[0] * 180 / np.pi
        yaw = pose[1] * 180 / np.pi
        roll = pose[2] * 180 / np.pi
        # Bin values
        bins = np.array(range(-99, 102, 3))
        labels = torch.LongTensor(np.digitize([yaw, pitch, roll], bins) - 1)

        label = torch.FloatTensor(pose)
        if self.transform is not None:
            img = self.transform(img)

        return img, labels, self.X_train[index]

    def __len__(self):
        # 122,450
        return self.length

class AFLW2000_binned(Dataset):
    def __init__(self, data_dir, filename_path, transform, img_ext='.jpg', annot_ext='.mat'):
        self.data_dir = data_dir
        self.transform = transform
        self.img_ext = img_ext
        self.annot_ext = annot_ext

        filename_list = get_list_from_filenames(filename_path)

        self.X_train = filename_list
        self.y_train = filename_list
        self.length = len(filename_list)

    def __getitem__(self, index):
        img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
        img = img.convert('RGB')

        # We get the pose in radians
        pose = utils.get_ypr_from_mat(os.path.join(self.data_dir, self.y_train[index] + self.annot_ext))
        # And convert to degrees.
        pitch, yaw, roll = pose * 180 / np.pi
        # Bin values
        bins = np.array(range(-99, 102, 3))
        binned_pitch = torch.DoubleTensor(np.digitize(pitch, bins) - 1)
        binned_yaw = torch.DoubleTensor(np.digitize(yaw, bins) - 1)
        binned_roll = torch.DoubleTensor(np.digitize(roll, bins) - 1)

        label = binned_yaw, binned_pitch, binned_roll

        if self.transform is not None:
            img = self.transform(img)
@@ -95,7 +136,7 @@
        return img, label, self.X_train[index]

    def __len__(self):
        # 122,450
        # 2,000
        return self.length

def get_list_from_filenames(file_path):

 code/hopenet.py

@@ -2,6 +2,7 @@
import torch.nn as nn
import torchvision.datasets as dsets
from torch.autograd import Variable
import math

# CNN Model (2 conv layer)
class Simple_CNN(nn.Module):
@@ -37,3 +38,66 @@
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out

class Hopenet(nn.Module):
    # This is just Hopenet with 3 output layers for yaw, pitch and roll.
    def __init__(self, block, layers, num_bins):
        self.inplanes = 64
        super(Hopenet, 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)

        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)

        return yaw, pitch, roll

 code/test_resnet_bins.py

@@ -13,7 +13,7 @@
import os
import argparse

from datasets import AFLW2000
import datasets
import hopenet
import utils

@@ -55,9 +55,10 @@

    print 'Loading data.'

    transformations = transforms.Compose([transforms.Scale(224),transforms.RandomCrop(224), transforms.ToTensor()])
    transformations = transforms.Compose([transforms.Scale(224),
    transforms.RandomCrop(224), transforms.ToTensor()])

    pose_dataset = AFLW2000(args.data_dir, args.filename_list,
    pose_dataset = datasets.AFLW2000_binned(args.data_dir, args.filename_list,
                                transformations)
    test_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
                                               batch_size=batch_size,
@@ -69,7 +70,9 @@

    # Test the Model
    model.eval()  # Change model to 'eval' mode (BN uses moving mean/var).
    error = .0
    yaw_correct = 0
    pitch_correct = 0
    roll_correct = 0
    total = 0
    for i, (images, labels, name) in enumerate(test_loader):
        images = Variable(images).cuda(gpu)
@@ -78,13 +81,14 @@
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        # TODO: There are more efficient ways.
        yaw_correct += (outputs[:][0] == labels[:][0])
        pitch_correct += (outputs[:][])
        for idx in xrange(len(outputs)):
            # if abs(outputs[idx].data[1] - labels[idx].data[1]) * 180 / np.pi > 30:
            print name
            print abs(outputs[idx].data - labels[idx].data) * 180 / np.pi, 180 * outputs[idx].data / np.pi, labels[idx].data * 180 / np.pi
            # error += utils.mse_loss(outputs[idx], labels[idx])
            error += abs(outputs[idx].data - labels[idx].data) * 180 / np.pi
            yaw_correct += (outputs[idx].data[0] == labels[idx].data[0])
            pitch_correct += (outputs[idx].data[1] == labels[idx].data[1])
            roll_correct += (outputs[idx].data[2] == labels[idx].data[2])


    print('Test MSE error of the model on the ' + str(total) +
    ' test images: %.4f' % (error / total))
    print('Test accuracies of the model on the ' + str(total) +
    ' test images. Yaw: %.4f %%, Pitch: %.4f %%, Roll: %.4f %%' % (yaw_correct / total,
    pitch_correct / total, roll_correct / total))

 code/train_resnet_bins.py

@@ -13,8 +13,17 @@
import os
import argparse

from datasets import Pose_300W_LP
import datasets
import hopenet
import torch.utils.model_zoo as model_zoo

model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}

def parse_args():
    """Parse input arguments."""
@@ -36,6 +45,41 @@

    return args

def get_ignored_params(model):
    # Generator function that yields ignored params.
    b = []
    b.append(model.conv1)
    b.append(model.bn1)
    b.append(model.layer1)
    b.append(model.layer2)
    b.append(model.layer3)
    b.append(model.layer4)
    for i in range(len(b)):
        for j in b[i].modules():
            for k in j.parameters():
                yield k

def get_non_ignored_params(model):
    # Generator function that yields params that will be optimized.
    b = []
    b.append(model.fc_yaw)
    b.append(model.fc_pitch)
    b.append(model.fc_roll)
    for i in range(len(b)):
        for j in b[i].modules():
            for k in j.parameters():
                    yield k

def load_filtered_state_dict(model, snapshot):
    # By user apaszke from discuss.pytorch.org
    model_dict = model.state_dict()
    # 1. filter out unnecessary keys
    snapshot = {k: v for k, v in snapshot.items() if k in model_dict}
    # 2. overwrite entries in the existing state dict
    model_dict.update(snapshot)
    # 3. load the new state dict
    model.load_state_dict(model_dict)

if __name__ == '__main__':
    args = parse_args()

@@ -47,21 +91,16 @@
    if not os.path.exists('output/snapshots'):
        os.makedirs('output/snapshots')

    model = torchvision.models.resnet18(pretrained=True)
    for param in model.parameters():
        param.requires_grad = False
    # Parameters of newly constructed modules have requires_grad=True by default
    num_ftrs = model.fc.in_features
    model.fc_pitch = nn.Linear(num_ftrs, 3)
    model.fc_yaw = nn.Linear(num_ftrs, 3)
    model.fc_roll = nn.Linear(num_ftrs, )

    # ResNet18 with 3 outputs.
    model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66)
    load_filtered_state_dict(model, model_zoo.load_url(model_urls['resnet18']))
    
    print 'Loading data.'

    transformations = transforms.Compose([transforms.Scale(230),transforms.RandomCrop(224),
    transformations = transforms.Compose([transforms.Scale(224),transforms.RandomCrop(224),
                                          transforms.ToTensor()])

    pose_dataset = Pose_300W_LP(args.data_dir, args.filename_list,
    pose_dataset = datasets.Pose_300W_LP_binned(args.data_dir, args.filename_list,
                                transformations)
    train_loader = torch.utils.data.DataLoader(dataset=pose_dataset,
                                               batch_size=batch_size,
@@ -69,31 +108,40 @@
                                               num_workers=2)

    model.cuda(gpu)
    criterion = nn.MSELoss(size_average = True)
    optimizer = torch.optim.Adam(model.fc.parameters(), lr = args.lr)
    criterion = nn.CrossEntropyLoss()
    optimizer = torch.optim.Adam([{'params': get_ignored_params(model), 'lr': .0},
                                  {'params': get_non_ignored_params(model), 'lr': args.lr}],
                                  lr = args.lr)

    print 'Ready to train network.'

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

            optimizer.zero_grad()
            outputs = model(images)
            loss = criterion(outputs, labels)
            loss.backward()
            yaw, pitch, roll = model(images)
            loss_yaw = criterion(yaw, label_yaw)
            loss_pitch = criterion(pitch, label_pitch)
            loss_roll = criterion(roll, label_roll)

            loss_seq = [loss_yaw, loss_pitch, loss_roll]
            grad_seq = [torch.Tensor(1).cuda(gpu) for _ in range(len(loss_seq))]
            torch.autograd.backward(loss_seq, grad_seq)
            optimizer.step()

            if (i+1) % 100 == 0:
                print ('Epoch [%d/%d], Iter [%d/%d] Loss: %.4f'
                       %(epoch+1, num_epochs, i+1, len(pose_dataset)//batch_size, loss.data[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]))

        # Save models at even numbered epochs.
        if epoch % 1 == 0 and epoch < num_epochs - 1:
            print 'Taking snapshot...'
            torch.save(model.state_dict(),
            'output/snapshots/resnet18_epoch_' + str(epoch+1) + '.pkl')
            'output/snapshots/resnet18_binned_epoch_' + str(epoch+1) + '.pkl')

    # Save the final Trained Model
    torch.save(model.state_dict(), 'output/snapshots/resnet18_epoch_' + str(epoch+1) + '.pkl')
    torch.save(model.state_dict(), 'output/snapshots/resnet18_binned_epoch_' + str(epoch+1) + '.pkl')