models.py

import numpy as np
from collections import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as model_zoo

class ResNetMultiImageInput(models.ResNet):
    """Constructs a resnet model with varying number of input images.
    Adapted from https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
    """
    def __init__(self, block, layers, num_classes=1000, num_input_images=1):
        super(ResNetMultiImageInput, self).__init__(block, layers)
        self.inplanes = 64
        self.conv1 = nn.Conv2d(num_input_images * 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)
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

def resnet_multiimage_input(num_layers, pretrained=False, num_input_images=1):
    """Constructs a ResNet model.
    Args:
        num_layers (int): Number of resnet layers. Must be 18 or 50
        pretrained (bool): If True, returns a model pre-trained on ImageNet
        num_input_images (int): Number of frames stacked as input
    """
    assert num_layers in [18, 50], "Can only run with 18 or 50 layer resnet"
    blocks = {18: [2, 2, 2, 2], 50: [3, 4, 6, 3]}[num_layers]
    block_type = {18: models.resnet.BasicBlock, 50: models.resnet.Bottleneck}[num_layers]
    model = ResNetMultiImageInput(block_type, blocks, num_input_images=num_input_images)
    if pretrained:
        loaded = model_zoo.load_url(models.resnet.model_urls['resnet{}'.format(num_layers)])
        loaded['conv1.weight'] = torch.cat([loaded['conv1.weight']] * num_input_images, 1) / num_input_images
        model.load_state_dict(loaded)
    return model

class ResnetEncoder(nn.Module):
    """Pytorch module for a resnet encoder """
    def __init__(self, num_layers, pretrained, num_input_images=1):
        super(ResnetEncoder, self).__init__()
        self.num_ch_enc = np.array([64, 64, 128, 256, 512])
        resnets = {18: models.resnet18,
                   34: models.resnet34,
                   50: models.resnet50,
                   101: models.resnet101,
                   152: models.resnet152}

        if num_layers not in resnets:
            raise ValueError("{} is not a valid number of resnet layers".format(num_layers))
        if num_input_images > 1:
            self.encoder = resnet_multiimage_input(num_layers, pretrained, num_input_images)
        else:
            self.encoder = resnets[num_layers](pretrained)
        if num_layers > 34:
            self.num_ch_enc[1:] *= 4

    def forward(self, input_image):
        self.features = []
        x = input_image
        x = self.encoder.conv1(x)
        x = self.encoder.bn1(x)
        self.features.append(self.encoder.relu(x))
        self.features.append(self.encoder.layer1(self.encoder.maxpool(self.features[-1])))
        self.features.append(self.encoder.layer2(self.features[-1]))
        self.features.append(self.encoder.layer3(self.features[-1]))
        self.features.append(self.encoder.layer4(self.features[-1]))
        return self.features

##################################################################
    
class PoseDecoder(nn.Module):
    def __init__(self, num_ch_enc, num_input_features=1, num_frames_to_predict_for=1, stride=1):
        super(PoseDecoder, self).__init__()
        self.num_ch_enc = num_ch_enc
        self.num_input_features = num_input_features

        if num_frames_to_predict_for is None:
            num_frames_to_predict_for = num_input_features - 1
        self.num_frames_to_predict_for = num_frames_to_predict_for

        self.convs = OrderedDict()
        self.convs[("squeeze")] = nn.Conv2d(self.num_ch_enc[-1], 256, 1)
        self.convs[("pose", 0)] = nn.Conv2d(num_input_features * 256, 256, 3, stride, 1)
        self.convs[("pose", 1)] = nn.Conv2d(256, 256, 3, stride, 1)
        self.convs[("pose", 2)] = nn.Conv2d(256, 6 * num_frames_to_predict_for, 1)
        self.relu = nn.ReLU()
        self.net = nn.ModuleList(list(self.convs.values()))

    def forward(self, input_features):
        last_features = [f[-1] for f in input_features]
        cat_features = [self.relu(self.convs["squeeze"](f)) for f in last_features]
        cat_features = torch.cat(cat_features, 1)

        out = cat_features
        for i in range(3):
            out = self.convs[("pose", i)](out)
            if i != 2:
                out = self.relu(out)
                
        out = out.mean(3).mean(2)
        pose = 0.01 * out.view(-1, 6)
        return pose


class PoseResNet(nn.Module):
    def __init__(self, num_layers = 18, pretrained = True):
        super(PoseResNet, self).__init__()
        self.encoder = ResnetEncoder(num_layers = num_layers, pretrained = pretrained, num_input_images=2)
        self.decoder = PoseDecoder(self.encoder.num_ch_enc)
    def init_weights(self):
        pass

    def forward(self, img1, img2):
        x = torch.cat([img1,img2],1)
        features = self.encoder(x)
        pose = self.decoder([features])
        return pose

##################################################################

class ConvBlock(nn.Module):
    """Layer to perform a convolution followed by ELU"""
    def __init__(self, in_channels, out_channels):
        super(ConvBlock, self).__init__()
        self.conv = Conv3x3(in_channels, out_channels)
        self.nonlin = nn.ELU(inplace=True)
    def forward(self, x):
        out = self.conv(x)
        out = self.nonlin(out)
        return out

class Conv3x3(nn.Module):
    """Layer to pad and convolve input"""
    def __init__(self, in_channels, out_channels, use_refl=True):
        super(Conv3x3, self).__init__()
        if use_refl:
            self.pad = nn.ReflectionPad2d(1)
        else:
            self.pad = nn.ZeroPad2d(1)
        self.conv = nn.Conv2d(int(in_channels), int(out_channels), 3)
    def forward(self, x):
        out = self.pad(x)
        out = self.conv(out)
        return out

def upsample(x):
    """Upsample input tensor by a factor of 2"""
    return F.interpolate(x, scale_factor=2, mode="nearest")

class DepthDecoder(nn.Module):
    def __init__(self, num_ch_enc, scales=range(4), num_output_channels=1, use_skips=True):
        super(DepthDecoder, self).__init__()
        self.alpha = 10
        self.beta = 0.01

        self.num_output_channels = num_output_channels
        self.use_skips = use_skips
        self.upsample_mode = 'nearest'
        self.scales = scales
        self.num_ch_enc = num_ch_enc
        self.num_ch_dec = np.array([16, 32, 64, 128, 256])

        # Decoder
        self.convs = OrderedDict()
        for i in range(4, -1, -1):
            # upconv_0
            num_ch_in = self.num_ch_enc[-1] if i == 4 else self.num_ch_dec[i + 1]
            num_ch_out = self.num_ch_dec[i]
            self.convs[("upconv", i, 0)] = ConvBlock(num_ch_in, num_ch_out)

            # upconv_1
            num_ch_in = self.num_ch_dec[i]
            if self.use_skips and i > 0:
                num_ch_in += self.num_ch_enc[i - 1]
            num_ch_out = self.num_ch_dec[i]
            self.convs[("upconv", i, 1)] = ConvBlock(num_ch_in, num_ch_out)
        for s in self.scales:
            self.convs[("dispconv", s)] = Conv3x3(self.num_ch_dec[s], self.num_output_channels)
        self.decoder = nn.ModuleList(list(self.convs.values()))
        self.sigmoid = nn.Sigmoid()

    def forward(self, input_features):
        self.outputs = []
        x = input_features[-1]
        
        for i in range(4, -1, -1):
            x = self.convs[("upconv", i, 0)](x)
            x = [upsample(x)]
            if self.use_skips and i > 0:
                x += [input_features[i - 1]]
            x = torch.cat(x, 1)
            x = self.convs[("upconv", i, 1)](x)
            if i in self.scales:
                self.outputs.append(self.alpha * self.sigmoid(self.convs[("dispconv", i)](x)) + self.beta)
        self.outputs = self.outputs[::-1]
        return self.outputs


class DispResNet(nn.Module):
    def __init__(self, num_layers = 18, pretrained = True):
        super(DispResNet, self).__init__()
        self.encoder = ResnetEncoder(num_layers = num_layers, pretrained = pretrained, num_input_images=1)
        self.decoder = DepthDecoder(self.encoder.num_ch_enc)
    def init_weights(self):
        pass
    def forward(self, x):
        features = self.encoder(x)
        outputs = self.decoder(features)        
        if self.training:
            return outputs
        else:
            return outputs[0]