baseline_3d_resnets.py

import torch
import torch.nn as nn
import math
import torch.utils.model_zoo as model_zoo

################
#
# Modified from https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
# Adds Inflated 3D conv based on method from I3D:
# "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset"
#          https://arxiv.org/abs/1705.07750
#
# Expects BxCxTxHxW data (different from 2D version)
#
################


__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
           'resnet152']


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 conv3x3(in_planes, out_planes, stride=1):
    """ 3D convolution with padding"""
    return nn.Conv3d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)



class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = nn.BatchNorm3d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm3d(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv3d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm3d(planes)
        self.conv2 = nn.Conv3d(planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm3d(planes)
        self.conv3 = nn.Conv3d(planes, planes * self.expansion, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm3d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def __init__(self, block, layers, num_classes=150, in_channels=3, input_size=112, input_len=16):
        self.inplanes = 64
        self.in_channels = in_channels
        super(ResNet, self).__init__()
        self.conv1 = nn.Conv3d(in_channels, 64, kernel_size=7, stride=(1,2,2), padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm3d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool3d(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)
        size = int(math.ceil(input_size/32))
        length = int(math.ceil(input_len/16))
        self.avgpool = nn.AvgPool3d((length,size,size), stride=1)
        self.dropout = nn.Dropout()
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv3d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm3d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv3d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm3d(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 is BxCxTxHxW
        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)

        # average pool BxC'xT'xH'xW' to be BxC
        # mean-pooling over T, H, and W
        # may want to look into this for better fully-conv support (i.e., per-frame classification)
        # or (temporal) pool after classifying
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.dropout(x)
        x = self.fc(x)
        # return BxClasses

        return x

    def load_2d_state_dict(self, state_dict, strict=True):
        # slightly hacky trick to inflate 2d kernel to 3d by repeating over time axis
        state_dict = {k:v for k,v in state_dict.items() if 'fc' not in k}
        md = self.state_dict()
        for k,v in state_dict.items():
            if 'conv' in k or 'downsample.0' in k:
                if isinstance(v, nn.Parameter):
                    v = v.data
                if self.in_channels != 3 and k == 'conv1.weight':
                    v = torch.mean(v, dim=1).unsqueeze(1).repeat(1, self.in_channels, 1, 1)
                # CxKxHxW -> CxKxTxHxW
                D = md[k].size(2)
                v = v.unsqueeze(2).repeat(1,1,D,1,1)
                state_dict[k] = v

        md.update(state_dict)
        super(ResNet, self).load_state_dict(md, strict)


def resnet18(pretrained=False, **kwargs):
    """Constructs a ResNet-18 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        model.load_2d_state_dict(model_zoo.load_url(model_urls['resnet18']))
    return model


def resnet34(pretrained=False, **kwargs):
    """Constructs a ResNet-34 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
    if pretrained:
        model.load_2d_state_dict(model_zoo.load_url(model_urls['resnet34']))
    return model


def resnet50(pretrained=False, mode='rgb', **kwargs):
    """Constructs a ResNet-50 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    if mode == 'rgb':
        model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
    else:
        model = ResNet(Bottleneck, [3, 4, 6, 3], in_channels=2, **kwargs)
    if pretrained:
        model.load_2d_state_dict(model_zoo.load_url(model_urls['resnet50']))
    return model


def resnet101(pretrained=False, **kwargs):
    """Constructs a ResNet-101 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
    if pretrained:
        model.load_2d_state_dict(model_zoo.load_url(model_urls['resnet101']))
    return model


def resnet152(pretrained=False, **kwargs):
    """Constructs a ResNet-152 model.
    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
    if pretrained:
        model.load_2d_state_dict(model_zoo.load_url(model_urls['resnet152']))
    return model


if __name__ == '__main__':
    # test resnet 50
    import torch
    d = torch.device('cuda')
    net = resnet50(pretrained=True, mode='flow')
    net.to(d)

    vid = torch.rand((4,2,16,112,112)).to(d)

    print(net(vid).size())