face rectangle

chenshijun

2019-06-05 f111cb002b9c6065fdf6bb274ce5857a9e875e8c

 code/hopenet.py

@@ -1,46 +1,12 @@
import torch
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):
    def __init__(self):
        super(Simple_CNN, self).__init__()
        self.layer1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, padding=0),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d(2))
        self.layer2 = nn.Sequential(
            nn.Conv2d(64, 128, kernel_size=3, padding=0),
            nn.BatchNorm2d(128),
            nn.ReLU(),
            nn.MaxPool2d(2))
        self.layer3 = nn.Sequential(
            nn.Conv2d(128, 256, kernel_size=3, padding=0),
            nn.BatchNorm2d(256),
            nn.ReLU(),
            nn.MaxPool2d(2))
        self.layer4 = nn.Sequential(
            nn.Conv2d(256, 512, kernel_size=3, padding=0),
            nn.BatchNorm2d(512),
            nn.ReLU(),
            nn.MaxPool2d(2))
        self.fc = nn.Linear(17*17*512, 3)

    def forward(self, x):
        out = self.layer1(x)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out
import torch.nn.functional as F

class Hopenet(nn.Module):
    # This is just Hopenet with 3 output layers for yaw, pitch and roll.
    # Hopenet with 3 output layers for yaw, pitch and roll
    # Predicts Euler angles by binning and regression with the expected value
    def __init__(self, block, layers, num_bins):
        self.inplanes = 64
        super(Hopenet, self).__init__()
@@ -58,6 +24,9 @@
        self.fc_pitch = nn.Linear(512 * block.expansion, num_bins)
        self.fc_roll = nn.Linear(512 * block.expansion, num_bins)

        # Vestigial layer from previous experiments
        self.fc_finetune = nn.Linear(512 * block.expansion + 3, 3)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
@@ -96,17 +65,17 @@

        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)
        pre_yaw = self.fc_yaw(x)
        pre_pitch = self.fc_pitch(x)
        pre_roll = self.fc_roll(x)

        return yaw, pitch, roll
        return pre_yaw, pre_pitch, pre_roll

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):
class ResNet(nn.Module):
    # ResNet for regression of 3 Euler angles.
    def __init__(self, block, layers, num_classes=1000):
        self.inplanes = 64
        super(Hopenet, self).__init__()
        super(ResNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
@@ -117,10 +86,7 @@
        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_1 = nn.Linear(512 * block.expansion, shape_bins)
        self.fc_angles = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
@@ -160,9 +126,46 @@

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc_angles(x)
        return x

class AlexNet(nn.Module):
    # AlexNet laid out as a Hopenet - classify Euler angles in bins and
    # regress the expected value.
    def __init__(self, num_bins):
        super(AlexNet, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(64, 192, kernel_size=5, padding=2),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(192, 384, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(384, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(256 * 6 * 6, 4096),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(inplace=True),
        )
        self.fc_yaw = nn.Linear(4096, num_bins)
        self.fc_pitch = nn.Linear(4096, num_bins)
        self.fc_roll = nn.Linear(4096, num_bins)

    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), 256 * 6 * 6)
        x = self.classifier(x)
        yaw = self.fc_yaw(x)
        pitch = self.fc_pitch(x)
        roll = self.fc_roll(x)
        shape_1 = self.fc_shape_1(x)

        return yaw, pitch, roll, shape_1
        return yaw, pitch, roll