pokouqiege/models/line_detect/heads/decoder.py

import torch
from torch import nn

from models.base.high_reso_resnet import Bottleneck, resnet101fpn
from typing import Any, Callable, List, Optional, Type, Union

class TestModel(nn.Sequential):
    def __init__(self,block: Type[Union[Bottleneck]]):
        super().__init__()
        self.encoder=resnet101fpn(out_channels=256)
        self.decoder=FPNDecoder(block)

    def forward(self, x):
        res=self.encoder(x)
        for k in res.keys():
            print(f'k:{k}')
        # print(f'res:{res}')

        out=self.decoder(res)
        return out


class FPNDecoder(nn.Sequential):
    def __init__(self,block: Type[Union[Bottleneck]], in_channels=256):
        super().__init__()
        self._norm_layer = nn.BatchNorm2d
        self.inplanes = 64
        self.dilation = 1
        self.groups = 1
        self.base_width = 64

        self.decoder0=self._make_decoder_layer(block,256,256)

        self.upconv4 = self._make_upsample_layer(256 , 256)
        self.upconv3 = self._make_upsample_layer(512, 256)
        self.upconv2 = self._make_upsample_layer(512, 256)
        self.upconv1 = self._make_upsample_layer(512, 256)
        self.upconv0 = self._make_upsample_layer(512, 256)

        self.final_up = self._make_upsample_layer(512, 128)

        self.decoder4 = self._make_decoder_layer(block, 512, 256)
        self.decoder3 = self._make_decoder_layer(block, 512, 256)
        self.decoder2 = self._make_decoder_layer(block, 512, 256)
        self.decoder1 = self._make_decoder_layer(block, 512, 256)

        self.final_conv = nn.Conv2d(128, 1, kernel_size=1)



    def forward(self, fpn_res):
        # ------------------
        # Encoder
        # ------------------
        e0 = fpn_res['0']  # [B, 64, H/4, W/4]
        print(f'e0:{e0.shape}')
        e1 = fpn_res['1']  # [B, 256, H/4, W/4]
        print(f'e1:{e1.shape}')
        e2 = fpn_res['2']  # [B, 512, H/8, W/8]
        print(f'e2:{e2.shape}')
        e3 = fpn_res['3']  # [B, 1024, H/16, W/16]
        print(f'e3:{e3.shape}')
        e4 = fpn_res['4']  # [B, 2048, H/32, W/32]
        print(f'e4:{e4.shape}')

        # ------------------
        # Decoder
        # ------------------

        d4 = self.upconv4(e4)  # [B, 1024, H/16, W/16]
        print(f'd4 = self.upconv4(e4):{d4.shape}')
        d4 = torch.cat([d4, e3], dim=1)  # [B, 2048, H/16, W/16]
        print(f' d4 = torch.cat([d4, e3], dim=1):{d4.shape}')
        d4 = self.decoder4(d4)  # [B, 2048, H/16, W/16]
        print(f'd4 = self.decoder4(d4):{d4.shape}')

        d3 = self.upconv3(d4)  # [B, 512, H/8, W/8]
        print(f'd3 = self.upconv3(d4):{d3.shape}')
        d3 = torch.cat([d3, e2], dim=1)  # [B, 1024, H/8, W/8]
        print(f'd3 = torch.cat([d3, e2], dim=1):{d3.shape}')
        d3 = self.decoder3(d3)  # [B, 1024, H/8, W/8]
        print(f'd3 = self.decoder3(d3):{d3.shape}')

        d2 = self.upconv2(d3)  # [B, 256, H/4, W/4]
        print(f'd2 = self.upconv2(d3):{d2.shape}')
        d2 = torch.cat([d2, e1], dim=1)  # [B, 512, H/4, W/4]
        print(f'd2 = torch.cat([d2, e1], dim=1):{d2.shape}')
        d2 = self.decoder2(d2)  # [B, 512, H/4, W/4]
        print(f'd2 = self.decoder2(d2):{d2.shape}')

        d1 = self.upconv1(d2)  # [B, 64, H/2, W/2]
        print(f'd1 = self.upconv1(d2):{d1.shape}')
        d1 = torch.cat([d1, e0], dim=1)  # [B, 128, H/2, W/2]
        print(f'd1 = torch.cat([d1, e0], dim=1):{d1.shape}')
        d1 = self.decoder1(d1)  # [B, 128, H/2, W/2]
        print(f'd1 =self.decoder1(d1):{d1.shape}')

        # ------------------
        # Output Head
        # ------------------
        d0=self.final_up(d1)
        out = self.final_conv(d0)  # [B, num_classes, H/2, W/2]
        print(f'out:{out.shape}')

        return out





    def _make_upsample_layer(self, in_channels: int, out_channels: int) -> nn.Module:
        """
        使用转置卷积进行上采样
        """
        return nn.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2)

    def _make_decoder_layer(self, block: Type[Union[Bottleneck]], in_channels: int,
                            out_channels: int, blocks: int = 1) -> nn.Sequential:
        """
        """
        # assert in_channels == out_channels, "in_channels must equal out_channels"
        layers = []
        for _ in range(blocks):
            layers.append(
                block(
                    in_channels,
                    in_channels // block.expansion,
                    groups=self.groups,
                    base_width=self.base_width,
                    dilation=self.dilation,
                    norm_layer=self._norm_layer,
                )
            )
        return nn.Sequential(*layers)

if __name__ == '__main__':
    model=TestModel(Bottleneck)
    x=torch.randn(3,3,512,512)
    out=model(x)