DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (DeepLabv2) Code

DeepLabv2의 경우 vgg16과 ResNet 두개의 버전이 있는데, 아래는 ResNet101로 구현했습니다.

#!/usr/bin/env python
# coding: utf-8

from collections import OrderedDict

import torch
import torch.nn as nn
import torch.nn.functional as F

from types import ModuleType
class Bottleneck(nn.Module):
    def __init__(self, in_ch, out_ch, stride, dilation, downsample):
        super(Bottleneck, self).__init__()
        mid_ch = out_ch // 4
        self.conv1 = nn.Conv2d(in_channels=in_ch, out_channels=mid_ch, kernel_size=1, stride=1, padding=0, dilation=dilation, bias=False)
        self.bn1 = nn.BatchNorm2d(num_features=mid_ch)
        self.conv2 = nn.Conv2d(in_channels=mid_ch, out_channels=mid_ch, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, bias=False)
        self.bn2 = nn.BatchNorm2d(num_features=mid_ch)
        self.conv3 = nn.Conv2d(in_channels=mid_ch, out_channels=out_ch, kernel_size=1, stride=1, padding=0, dilation=dilation, bias=False)
        self.bn3 = nn.BatchNorm2d(num_features=out_ch)
        self.relu = nn.ReLU(inplace=True)
        if downsample:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels=in_ch, out_channels=out_ch, kernel_size=1, stride=stride, padding=0, dilation=dilation, bias=False), 
                nn.BatchNorm2d(num_features=out_ch)
            )
            self.is_downsample = True
        else:
            self.is_downsample = False


    def forward(self, x):
        h = self.bn1(self.conv1(x))
        h = self.bn2(self.conv2(h))
        h = self.bn3(self.conv3(h))
        h = self.relu(h)
        if self.is_downsample: 
            h += self.downsample(x)
        else:
            h += x 
        return h

class ResLayer(nn.Sequential):
    def __init__(self, n_layers, in_ch, out_ch, stride, dilation, multi_grid=0):
        super(ResLayer, self).__init__()
        multi_grids = [1, 2, 2] if multi_grid else [1 for _ in range(n_layers)]
        for i in range(n_layers):
            self.add_module(
                "{}".format(i),
                Bottleneck(
                    in_ch=(in_ch if i == 0 else out_ch), 
                    out_ch=out_ch,
                    stride=(stride if i == 0 else 1),
                    dilation=dilation * multi_grids[i],
                    downsample=(True if i == 0 else False), # Downsampling is only in the first block (i=0)
                ),
            )


class IntermediateLayerGetter(nn.Sequential):
    def __init__(self, n_blocks, ch, atrous_rates, output_stride):
        super(IntermediateLayerGetter, self).__init__()
        # Stride and dilation
        if output_stride == 8:
            s = [1, 2, 1, 1]
            d = [1, 1, 2, 4]
        elif output_stride == 16:
            s = [1, 2, 2, 1]
            d = [1, 1, 1, 2]

        self.add_module("conv1", nn.Conv2d(in_channels=3, out_channels=64, kernel_size=7, stride=2, padding=3, bias=False))
        self.add_module("bn1", nn.BatchNorm2d(64))
        self.add_module("relu", nn.ReLU(inplace=True))
        self.add_module("maxpool", nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1))
        self.add_module("layer1", ResLayer(n_blocks[0], ch[0], ch[2], s[0], d[0], 0))
        self.add_module("layer2", ResLayer(n_blocks[1], ch[2], ch[3], s[1], d[1], 0))
        self.add_module("layer3", ResLayer(n_blocks[2], ch[3], ch[4], s[2], d[2], 0))
        self.add_module("layer4", ResLayer(n_blocks[3], ch[4], ch[5], s[3], d[3], 1)) # multi_grid 넣을라면 1 아니면 0


class ASPP(nn.Module):
    def __init__(self, in_channels, out_channels, n_classes=21):
        super(ASPP, self).__init__()
        # atrous 3x3, rate=6
        self.conv_3x3_r6 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=6, dilation=6)
        # atrous 3x3, rate=12
        self.conv_3x3_r12 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=12, dilation=12)
        # atrous 3x3, rate=18
        self.conv_3x3_r18 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=18, dilation=18)
        # atrous 3x3, rate=24
        self.conv_3x3_r24 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=24, dilation=24)
        self.drop_conv_3x3 = nn.Dropout2d(0.5)

        self.conv_1x1 = nn.Conv2d(out_channels, out_channels, kernel_size=1)
        self.drop_conv_1x1 = nn.Dropout2d(0.5)

        self.conv_1x1_out = nn.Conv2d(out_channels, n_classes, kernel_size=1)

    def forward(self, feature_map):
        # 1번 branch
        # shape: (batch_size, out_channels, height/output_stride, width/output_stride)
        out_3x3_r6 = self.drop_conv_3x3(F.relu(self.conv_3x3_r6(feature_map)))
        out_img_r6 = self.drop_conv_1x1(F.relu(self.conv_1x1(out_3x3_r6)))
        out_img_r6 = self.conv_1x1_out(out_img_r6)
        # 2번 branch
        # shape: (batch_size, out_channels, height/output_stride, width/output_stride)
        out_3x3_r12 = self.drop_conv_3x3(F.relu(self.conv_3x3_r12(feature_map)))
        out_img_r12 = self.drop_conv_1x1(F.relu(self.conv_1x1(out_3x3_r12)))
        out_img_r12 = self.conv_1x1_out(out_img_r12)
        # 3번 branch
        # shape: (batch_size, out_channels, height/output_stride, width/output_stride)
        out_3x3_r18 = self.drop_conv_3x3(F.relu(self.conv_3x3_r18(feature_map)))
        out_img_r18 = self.drop_conv_1x1(F.relu(self.conv_1x1(out_3x3_r18)))
        out_img_r18 = self.conv_1x1_out(out_img_r18)
        # 4번 branch
        # shape: (batch_size, out_channels, height/output_stride, width/output_stride)
        out_3x3_r24 = self.drop_conv_3x3(F.relu(self.conv_3x3_r24(feature_map)))
        out_img_r24 = self.drop_conv_1x1(F.relu(self.conv_1x1(out_3x3_r24)))
        out_img_r24 = self.conv_1x1_out(out_img_r24)
        out = sum([out_img_r6, out_img_r12, out_img_r18, out_img_r24])

        return out


class DeepLabV2(nn.Sequential):
    def __init__(self, n_classes=21):
        super(DeepLabV2, self).__init__()
        ch = [64 * 2 ** p for p in range(6)]
        self.backbone = IntermediateLayerGetter(n_blocks=[3, 4, 23, 3], ch=[64 * 2 ** p for p in range(6)], atrous_rates=[6, 12, 18, 24], output_stride=16)
        self.classifier = ASPP(2048, 256, n_classes)

    def forward(self, x): 
        h = self.backbone(x)
        h = self.classifier(h)
        h = F.interpolate(h, size=x.shape[2:], mode="bilinear", align_corners=False)
        return h