model_resnet.py

from typing import Optional, Callable, Type, Union, List

import torch
import torch.nn as nn
from torch.functional import F
from torch import Tensor


class NLBlockND(nn.Module):
    def __init__(self, in_channels, inter_channels=None, mode='embedded',
                 dimension=3, bn_layer=True):
        """Implementation of Non-Local Block with 4 different pairwise functions but doesn't include subsampling trick
        args:
            in_channels: original channel size (1024 in the paper)
            inter_channels: channel size inside the block if not specifed reduced to half (512 in the paper)
            mode: supports Gaussian, Embedded Gaussian, Dot Product, and Concatenation
            dimension: can be 1 (temporal), 2 (spatial), 3 (spatiotemporal)
            bn_layer: whether to add batch norm
        """
        super(NLBlockND, self).__init__()

        assert dimension in [1, 2, 3]

        if mode not in ['gaussian', 'embedded', 'dot', 'concatenate']:
            raise ValueError('`mode` must be one of `gaussian`, `embedded`, `dot` or `concatenate`')

        self.mode = mode
        self.dimension = dimension

        self.in_channels = in_channels
        self.inter_channels = inter_channels

        # the channel size is reduced to half inside the block
        if self.inter_channels is None:
            self.inter_channels = in_channels // 2
            if self.inter_channels == 0:
                self.inter_channels = 1

        # assign appropriate convolutional, max pool, and batch norm layers for different dimensions
        if dimension == 3:
            conv_nd = nn.Conv3d
            max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
            bn = nn.BatchNorm3d
        elif dimension == 2:
            conv_nd = nn.Conv2d
            max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
            bn = nn.BatchNorm2d
        else:
            conv_nd = nn.Conv1d
            max_pool_layer = nn.MaxPool1d(kernel_size=(2))
            bn = nn.BatchNorm1d

        # function g in the paper which goes through conv. with kernel size 1
        self.g = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)

        # add BatchNorm layer after the last conv layer
        if bn_layer:
            self.W_z = nn.Sequential(
                    conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1),
                    bn(self.in_channels)
                )
            # from section 4.1 of the paper, initializing params of BN ensures that the initial state of non-local block is identity mapping
            nn.init.constant_(self.W_z[1].weight, 0)
            nn.init.constant_(self.W_z[1].bias, 0)
        else:
            self.W_z = conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1)

            # from section 3.3 of the paper by initializing Wz to 0, this block can be inserted to any existing architecture
            nn.init.constant_(self.W_z.weight, 0)
            nn.init.constant_(self.W_z.bias, 0)

        # define theta and phi for all operations except gaussian
        if self.mode == "embedded" or self.mode == "dot" or self.mode == "concatenate":
            self.theta = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)
            self.phi = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)

        if self.mode == "concatenate":
            self.W_f = nn.Sequential(
                    nn.Conv2d(in_channels=self.inter_channels * 2, out_channels=1, kernel_size=1),
                    nn.ReLU()
                )

    def forward(self, x):
        """
        args
            x: (N, C, T, H, W) for dimension=3; (N, C, H, W) for dimension 2; (N, C, T) for dimension 1
        """

        batch_size = x.size(0)

        # (N, C, THW)
        # this reshaping and permutation is from the spacetime_nonlocal function in the original Caffe2 implementation
        g_x = self.g(x).view(batch_size, self.inter_channels, -1)
        g_x = g_x.permute(0, 2, 1)

        if self.mode == "gaussian":
            theta_x = x.view(batch_size, self.in_channels, -1)
            phi_x = x.view(batch_size, self.in_channels, -1)
            theta_x = theta_x.permute(0, 2, 1)
            f = torch.matmul(theta_x, phi_x)

        elif self.mode == "embedded" or self.mode == "dot":
            theta_x = self.theta(x).view(batch_size, self.inter_channels, -1)
            phi_x = self.phi(x).view(batch_size, self.inter_channels, -1)
            theta_x = theta_x.permute(0, 2, 1)
            f = torch.matmul(theta_x, phi_x)

        elif self.mode == "concatenate":
            theta_x = self.theta(x).view(batch_size, self.inter_channels, -1, 1)
            phi_x = self.phi(x).view(batch_size, self.inter_channels, 1, -1)

            h = theta_x.size(2)
            w = phi_x.size(3)
            theta_x = theta_x.repeat(1, 1, 1, w)
            phi_x = phi_x.repeat(1, 1, h, 1)

            concat = torch.cat([theta_x, phi_x], dim=1)
            f = self.W_f(concat)
            f = f.view(f.size(0), f.size(2), f.size(3))

        if self.mode == "gaussian" or self.mode == "embedded":
            f_div_C = F.softmax(f, dim=-1)
        elif self.mode == "dot" or self.mode == "concatenate":
            N = f.size(-1) # number of position in x
            f_div_C = f / N

        y = torch.matmul(f_div_C, g_x)

        # contiguous here just allocates contiguous chunk of memory
        y = y.permute(0, 2, 1).contiguous()
        y = y.view(batch_size, self.inter_channels, *x.size()[2:])

        W_y = self.W_z(y)
        # residual connection
        z = W_y + x

        return z


def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
    """3x3 convolution with padding"""
    return nn.Conv2d(
        in_planes,
        out_planes,
        kernel_size=3,
        stride=stride,
        padding=dilation,
        groups=groups,
        bias=False,
        dilation=dilation,
    )


def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


class BasicBlock(nn.Module):
    expansion: int = 1

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64:
            raise ValueError("BasicBlock only supports groups=1 and base_width=64")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(inplanes, planes, stride, dilation=dilation)
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes, dilation=dilation)
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        identity = x

        out = self.conv1(x)    # 3x3 convolution
        out = self.bn1(out)    # Batch normalization
        out = self.relu(out)   # Relu
        out = self.conv2(out)  # 3x3 convolution
        out = self.bn2(out)    # Batch normalization

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

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

        return out


class DecBlock(nn.Module):
    def __init__(self, in_ch, out_ch):
        super().__init__()
        self.conv1 = nn.Conv2d(in_ch, out_ch, 3, padding=1)
        self.relu  = nn.ReLU()
        self.conv2 = nn.Conv2d(out_ch, out_ch, 3, padding=1)
    
    def forward(self, x):
        ret = self.conv2(self.relu(self.conv1(x)))
        return ret



class ResNet(nn.Module):
    def __init__(
        self,
        block: Type[BasicBlock],
        layers: List[int],
        zero_init_residual: bool = False,
        groups: int = 1,
        width_per_group: int = 64,
        dilation: Optional[List[int]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None,
    ) -> None:
        super().__init__()
        #_log_api_usage_once(self)
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if dilation is None:
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            dilation = [1, 1, 1]
        if len(dilation) != 3:
            raise ValueError(
                "dilation should be None "
                f"or a 3-element tuple, got {dilation}"
            )
        self.groups = groups
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = norm_layer(self.inplanes)
        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, dilation=dilation[0])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilation=dilation[1], nonLocalBlock = True)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilation=dilation[2], nonLocalBlock = True)

        #self.deconv1 = nn.ConvTranspose2d(512, 512, 2, stride=2, padding=0, output_padding=0)
        #self.dec1 = DecBlock(512 + 256, 256)
        #self.deconv2 = nn.ConvTranspose2d(256, 256, 2, stride=2, padding=0, output_padding=0)
        #self.dec2 = DecBlock(256 + 128, 128)
        #self.deconv3 = nn.ConvTranspose2d(128, 128, 2, stride=2, padding=0, output_padding=0)
        #self.dec3 = DecBlock(128 + 64, 92)

        self.deconv1 = nn.ConvTranspose2d(512, 512, 2, stride=2, padding=0, output_padding=0)
        self.deconv2 = nn.ConvTranspose2d(512 + 256, 256, 2, stride=2, padding=0, output_padding=0)
        self.deconv3 = nn.ConvTranspose2d(256 + 128, 128, 2, stride=2, padding=0, output_padding=0)

        self.deconv4 = nn.Conv2d(128 + 64, 94, kernel_size=1)



        for m in self.modules():
            if isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]

    def _make_layer(
        self,
        block: Type[BasicBlock],
        planes: int,
        blocks: int,
        stride: int = 1,
        dilation: int = 1,
        nonLocalBlock = False
    ) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(
            block(
                self.inplanes, planes, stride, downsample, self.groups, self.base_width, dilation, norm_layer
            )
        )
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            if nonLocalBlock and i == blocks - 1:
                layers.append(NLBlockND(self.inplanes, self.inplanes, dimension=2))
            layers.append(
                block(
                    self.inplanes,
                    planes,
                    groups=self.groups,
                    base_width=self.base_width,
                    dilation=dilation,
                    norm_layer=norm_layer,
                )
            )

        return nn.Sequential(*layers)

    def _forward_impl(self, x: Tensor) -> Tensor:
        # See note [TorchScript super()]
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x1 = self.layer1(x)
        x2 = self.layer2(x1)
        x3 = self.layer3(x2)
        x4 = self.layer4(x3)

        x = x4

        x = self.relu(self.deconv1(x))
        x = torch.cat([x, x3.detach()], dim = 1)
        #x = self.relu(self.dec1(x))

        x = self.relu(self.deconv2(x))
        x = torch.cat([x, x2.detach()], dim = 1)
        #x = self.relu(self.dec2(x))

        x = self.relu(self.deconv3(x))
        x = torch.cat([x[:,:,:-1], x1.detach()], dim = 1)
        #x = self.dec3(x)

        x = self.deconv4(x)

        #x = torch.sigmoid(x)
        x = torch.permute(x, (0, 2, 3, 1))
        x = torch.softmax(x, dim=3)
        x = torch.permute(x, (0, 3, 1, 2))

        x = torch.clamp(x, min=1e-4, max=1-1e-4)

        return x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)


def makeModel():
    model = ResNet(BasicBlock, [2, 2, 2, 2], dilation=[1, 2, 2])
    return model