upcunet_v3.py

import torch
from torch import nn as nn
from torch.nn import functional as F
import os, sys
import numpy as np

root_path = os.path.abspath('.')
sys.path.append(root_path)


class SEBlock(nn.Module):
    def __init__(self, in_channels, reduction=8, bias=False):
        super(SEBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels, in_channels // reduction, 1, 1, 0, bias=bias)
        self.conv2 = nn.Conv2d(in_channels // reduction, in_channels, 1, 1, 0, bias=bias)

    def forward(self, x):
        if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
            x0 = torch.mean(x.float(), dim=(2, 3), keepdim=True).half()
        else:
            x0 = torch.mean(x, dim=(2, 3), keepdim=True)
        x0 = self.conv1(x0)
        x0 = F.relu(x0, inplace=True)
        x0 = self.conv2(x0)
        x0 = torch.sigmoid(x0)
        x = torch.mul(x, x0)
        return x

    def forward_mean(self, x, x0):
        x0 = self.conv1(x0)
        x0 = F.relu(x0, inplace=True)
        x0 = self.conv2(x0)
        x0 = torch.sigmoid(x0)
        x = torch.mul(x, x0)
        return x


class UNetConv(nn.Module):
    def __init__(self, in_channels, mid_channels, out_channels, se):
        super(UNetConv, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, mid_channels, 3, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(mid_channels, out_channels, 3, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
        )
        if se:
            self.seblock = SEBlock(out_channels, reduction=8, bias=True)
        else:
            self.seblock = None

    def forward(self, x):
        z = self.conv(x)
        if self.seblock is not None:
            z = self.seblock(z)
        return z


class UNet1(nn.Module):
    def __init__(self, in_channels, out_channels, deconv):
        super(UNet1, self).__init__()
        self.conv1 = UNetConv(in_channels, 32, 64, se=False)
        self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
        self.conv2 = UNetConv(64, 128, 64, se=True)
        self.conv2_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
        self.conv3 = nn.Conv2d(64, 64, 3, 1, 0)

        if deconv:
            self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
        else:
            self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)

        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2(x2)
        x2 = self.conv2_up(x2)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)

        x1 = F.pad(x1, (-4, -4, -4, -4))
        x3 = self.conv3(x1 + x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        z = self.conv_bottom(x3)
        return z

    def forward_a(self, x):
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2.conv(x2)
        return x1, x2

    def forward_b(self, x1, x2):
        x2 = self.conv2_up(x2)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)

        x1 = F.pad(x1, (-4, -4, -4, -4))
        x3 = self.conv3(x1 + x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        z = self.conv_bottom(x3)
        return z


class UNet1x3(nn.Module):
    def __init__(self, in_channels, out_channels, deconv):
        super(UNet1x3, self).__init__()
        self.conv1 = UNetConv(in_channels, 32, 64, se=False)
        self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
        self.conv2 = UNetConv(64, 128, 64, se=True)
        self.conv2_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
        self.conv3 = nn.Conv2d(64, 64, 3, 1, 0)

        if deconv:
            self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 5, 3, 2)
        else:
            self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)

        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2(x2)
        x2 = self.conv2_up(x2)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)

        x1 = F.pad(x1, (-4, -4, -4, -4))
        x3 = self.conv3(x1 + x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        z = self.conv_bottom(x3)
        return z

    def forward_a(self, x):
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2.conv(x2)
        return x1, x2

    def forward_b(self, x1, x2):
        x2 = self.conv2_up(x2)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)

        x1 = F.pad(x1, (-4, -4, -4, -4))
        x3 = self.conv3(x1 + x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        z = self.conv_bottom(x3)
        return z


class UNet2(nn.Module):
    def __init__(self, in_channels, out_channels, deconv):
        super(UNet2, self).__init__()

        self.conv1 = UNetConv(in_channels, 32, 64, se=False)
        self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
        self.conv2 = UNetConv(64, 64, 128, se=True)
        self.conv2_down = nn.Conv2d(128, 128, 2, 2, 0)
        self.conv3 = UNetConv(128, 256, 128, se=True)
        self.conv3_up = nn.ConvTranspose2d(128, 128, 2, 2, 0)
        self.conv4 = UNetConv(128, 64, 64, se=True)
        self.conv4_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
        self.conv5 = nn.Conv2d(64, 64, 3, 1, 0)

        if deconv:
            self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
        else:
            self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)

        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2(x2)

        x3 = self.conv2_down(x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        x3 = self.conv3(x3)
        x3 = self.conv3_up(x3)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)

        x2 = F.pad(x2, (-4, -4, -4, -4))
        x4 = self.conv4(x2 + x3)
        x4 = self.conv4_up(x4)
        x4 = F.leaky_relu(x4, 0.1, inplace=True)

        x1 = F.pad(x1, (-16, -16, -16, -16))
        x5 = self.conv5(x1 + x4)
        x5 = F.leaky_relu(x5, 0.1, inplace=True)

        z = self.conv_bottom(x5)
        return z

    def forward_a(self, x):  # conv234结尾有se
        x1 = self.conv1(x)
        x2 = self.conv1_down(x1)
        x2 = F.leaky_relu(x2, 0.1, inplace=True)
        x2 = self.conv2.conv(x2)
        return x1, x2

    def forward_b(self, x2):  # conv234结尾有se
        x3 = self.conv2_down(x2)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)
        x3 = self.conv3.conv(x3)
        return x3

    def forward_c(self, x2, x3):  # conv234结尾有se
        x3 = self.conv3_up(x3)
        x3 = F.leaky_relu(x3, 0.1, inplace=True)

        x2 = F.pad(x2, (-4, -4, -4, -4))
        x4 = self.conv4.conv(x2 + x3)
        return x4

    def forward_d(self, x1, x4):  # conv234结尾有se
        x4 = self.conv4_up(x4)
        x4 = F.leaky_relu(x4, 0.1, inplace=True)

        x1 = F.pad(x1, (-16, -16, -16, -16))
        x5 = self.conv5(x1 + x4)
        x5 = F.leaky_relu(x5, 0.1, inplace=True)

        z = self.conv_bottom(x5)
        return z


class UpCunet2x(nn.Module):  # 完美tile，全程无损
    def __init__(self, in_channels=3, out_channels=3):
        super(UpCunet2x, self).__init__()
        self.unet1 = UNet1(in_channels, out_channels, deconv=True)
        self.unet2 = UNet2(in_channels, out_channels, deconv=False)

    def forward(self, x, tile_mode):  # 1.7G
        n, c, h0, w0 = x.shape
        if (tile_mode == 0):  # 不tile
            ph = ((h0 - 1) // 2 + 1) * 2
            pw = ((w0 - 1) // 2 + 1) * 2
            x = F.pad(x, (18, 18 + pw - w0, 18, 18 + ph - h0), 'reflect')  # 需要保证被2整除
            x = self.unet1.forward(x)
            x0 = self.unet2.forward(x)
            x1 = F.pad(x, (-20, -20, -20, -20))
            x = torch.add(x0, x1)
            if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 2, :w0 * 2]
            return x
        elif (tile_mode == 1):  # 对长边减半
            if (w0 >= h0):
                crop_size_w = ((w0 - 1) // 4 * 4 + 4) // 2  # 减半后能被2整除，所以要先被4整除
                crop_size_h = (h0 - 1) // 2 * 2 + 2  # 能被2整除
            else:
                crop_size_h = ((h0 - 1) // 4 * 4 + 4) // 2  # 减半后能被2整除，所以要先被4整除
                crop_size_w = (w0 - 1) // 2 * 2 + 2  # 能被2整除
            crop_size = (crop_size_h, crop_size_w)  # 6.6G
        elif (tile_mode == 2):  # hw都减半
            crop_size = (((h0 - 1) // 4 * 4 + 4) // 2, ((w0 - 1) // 4 * 4 + 4) // 2)  # 5.6G
        elif (tile_mode == 3):  # hw都三分之一
            crop_size = (((h0 - 1) // 6 * 6 + 6) // 3, ((w0 - 1) // 6 * 6 + 6) // 3)  # 4.2G
        elif (tile_mode == 4):  # hw都四分之一
            crop_size = (((h0 - 1) // 8 * 8 + 8) // 4, ((w0 - 1) // 8 * 8 + 8) // 4)  # 3.7G
        ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
        pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
        x = F.pad(x, (18, 18 + pw - w0, 18, 18 + ph - h0), 'reflect')
        n, c, h, w = x.shape
        se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        n_patch = 0
        tmp_dict = {}
        opt_res_dict = {}
        for i in range(0, h - 36, crop_size[0]):
            tmp_dict[i] = {}
            for j in range(0, w - 36, crop_size[1]):
                x_crop = x[:, :, i:i + crop_size[0] + 36, j:j + crop_size[1] + 36]
                n, c1, h1, w1 = x_crop.shape
                tmp0, x_crop = self.unet1.forward_a(x_crop)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                n_patch += 1
                tmp_dict[i][j] = (tmp0, x_crop)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 36, crop_size[0]):
            for j in range(0, w - 36, crop_size[1]):
                tmp0, x_crop = tmp_dict[i][j]
                x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
                opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
                tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
        se_mean1 /= n_patch
        se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        for i in range(0, h - 36, crop_size[0]):
            for j in range(0, w - 36, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
                tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
                tmp_x3 = self.unet2.forward_b(tmp_x2)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 36, crop_size[0]):
            for j in range(0, w - 36, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
                tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
                tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
        se_mean1 /= n_patch
        for i in range(0, h - 36, crop_size[0]):
            opt_res_dict[i] = {}
            for j in range(0, w - 36, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
                tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
                x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
                x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
                x_crop = torch.add(x0, x1)  # x0是unet2的最终输出
                opt_res_dict[i][j] = x_crop
        del tmp_dict
        torch.cuda.empty_cache()
        res = torch.zeros((n, c, h * 2 - 72, w * 2 - 72)).to(x.device)
        if ("Half" in x.type()):
            res = res.half()
        for i in range(0, h - 36, crop_size[0]):
            for j in range(0, w - 36, crop_size[1]):
                res[:, :, i * 2:i * 2 + h1 * 2 - 72, j * 2:j * 2 + w1 * 2 - 72] = opt_res_dict[i][j]
        del opt_res_dict
        torch.cuda.empty_cache()
        if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 2, :w0 * 2]
        return res  #


class UpCunet3x(nn.Module):  # 完美tile，全程无损
    def __init__(self, in_channels=3, out_channels=3):
        super(UpCunet3x, self).__init__()
        self.unet1 = UNet1x3(in_channels, out_channels, deconv=True)
        self.unet2 = UNet2(in_channels, out_channels, deconv=False)

    def forward(self, x, tile_mode):  # 1.7G
        n, c, h0, w0 = x.shape
        if (tile_mode == 0):  # 不tile
            ph = ((h0 - 1) // 4 + 1) * 4
            pw = ((w0 - 1) // 4 + 1) * 4
            x = F.pad(x, (14, 14 + pw - w0, 14, 14 + ph - h0), 'reflect')  # 需要保证被2整除
            x = self.unet1.forward(x)
            x0 = self.unet2.forward(x)
            x1 = F.pad(x, (-20, -20, -20, -20))
            x = torch.add(x0, x1)
            if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 3, :w0 * 3]
            return x
        elif (tile_mode == 1):  # 对长边减半
            if (w0 >= h0):
                crop_size_w = ((w0 - 1) // 8 * 8 + 8) // 2  # 减半后能被4整除，所以要先被8整除
                crop_size_h = (h0 - 1) // 4 * 4 + 4  # 能被4整除
            else:
                crop_size_h = ((h0 - 1) // 8 * 8 + 8) // 2  # 减半后能被4整除，所以要先被8整除
                crop_size_w = (w0 - 1) // 4 * 4 + 4  # 能被4整除
            crop_size = (crop_size_h, crop_size_w)  # 6.6G
        elif (tile_mode == 2):  # hw都减半
            crop_size = (((h0 - 1) // 8 * 8 + 8) // 2, ((w0 - 1) // 8 * 8 + 8) // 2)  # 5.6G
        elif (tile_mode == 3):  # hw都三分之一
            crop_size = (((h0 - 1) // 12 * 12 + 12) // 3, ((w0 - 1) // 12 * 12 + 12) // 3)  # 4.2G
        elif (tile_mode == 4):  # hw都四分之一
            crop_size = (((h0 - 1) // 16 * 16 + 16) // 4, ((w0 - 1) // 16 * 16 + 16) // 4)  # 3.7G
        ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
        pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
        x = F.pad(x, (14, 14 + pw - w0, 14, 14 + ph - h0), 'reflect')
        n, c, h, w = x.shape
        se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        n_patch = 0
        tmp_dict = {}
        opt_res_dict = {}
        for i in range(0, h - 28, crop_size[0]):
            tmp_dict[i] = {}
            for j in range(0, w - 28, crop_size[1]):
                x_crop = x[:, :, i:i + crop_size[0] + 28, j:j + crop_size[1] + 28]
                n, c1, h1, w1 = x_crop.shape
                tmp0, x_crop = self.unet1.forward_a(x_crop)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                n_patch += 1
                tmp_dict[i][j] = (tmp0, x_crop)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 28, crop_size[0]):
            for j in range(0, w - 28, crop_size[1]):
                tmp0, x_crop = tmp_dict[i][j]
                x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
                opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
                tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
        se_mean1 /= n_patch
        se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        for i in range(0, h - 28, crop_size[0]):
            for j in range(0, w - 28, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
                tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
                tmp_x3 = self.unet2.forward_b(tmp_x2)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 28, crop_size[0]):
            for j in range(0, w - 28, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
                tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
                tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
        se_mean1 /= n_patch
        for i in range(0, h - 28, crop_size[0]):
            opt_res_dict[i] = {}
            for j in range(0, w - 28, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
                tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
                x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
                x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
                x_crop = torch.add(x0, x1)  # x0是unet2的最终输出
                opt_res_dict[i][j] = x_crop  #
        del tmp_dict
        torch.cuda.empty_cache()
        res = torch.zeros((n, c, h * 3 - 84, w * 3 - 84)).to(x.device)
        if ("Half" in x.type()):
            res = res.half()
        for i in range(0, h - 28, crop_size[0]):
            for j in range(0, w - 28, crop_size[1]):
                res[:, :, i * 3:i * 3 + h1 * 3 - 84, j * 3:j * 3 + w1 * 3 - 84] = opt_res_dict[i][j]
        del opt_res_dict
        torch.cuda.empty_cache()
        if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 3, :w0 * 3]
        return res


class UpCunet4x(nn.Module):  # 完美tile，全程无损
    def __init__(self, in_channels=3, out_channels=3):
        super(UpCunet4x, self).__init__()
        self.unet1 = UNet1(in_channels, 64, deconv=True)
        self.unet2 = UNet2(64, 64, deconv=False)
        self.ps = nn.PixelShuffle(2)
        self.conv_final = nn.Conv2d(64, 12, 3, 1, padding=0, bias=True)

    def forward(self, x, tile_mode):
        n, c, h0, w0 = x.shape
        x00 = x
        if (tile_mode == 0):  # 不tile
            ph = ((h0 - 1) // 2 + 1) * 2
            pw = ((w0 - 1) // 2 + 1) * 2
            x = F.pad(x, (19, 19 + pw - w0, 19, 19 + ph - h0), 'reflect')  # 需要保证被2整除
            x = self.unet1.forward(x)
            x0 = self.unet2.forward(x)
            x1 = F.pad(x, (-20, -20, -20, -20))
            x = torch.add(x0, x1)
            x = self.conv_final(x)
            x = F.pad(x, (-1, -1, -1, -1))
            x = self.ps(x)
            if (w0 != pw or h0 != ph): x = x[:, :, :h0 * 4, :w0 * 4]
            x += F.interpolate(x00, scale_factor=4, mode='nearest')
            return x
        elif (tile_mode == 1):  # 对长边减半
            if (w0 >= h0):
                crop_size_w = ((w0 - 1) // 4 * 4 + 4) // 2  # 减半后能被2整除，所以要先被4整除
                crop_size_h = (h0 - 1) // 2 * 2 + 2  # 能被2整除
            else:
                crop_size_h = ((h0 - 1) // 4 * 4 + 4) // 2  # 减半后能被2整除，所以要先被4整除
                crop_size_w = (w0 - 1) // 2 * 2 + 2  # 能被2整除
            crop_size = (crop_size_h, crop_size_w)  # 6.6G
        elif (tile_mode == 2):  # hw都减半
            crop_size = (((h0 - 1) // 4 * 4 + 4) // 2, ((w0 - 1) // 4 * 4 + 4) // 2)  # 5.6G
        elif (tile_mode == 3):  # hw都三分之一
            crop_size = (((h0 - 1) // 6 * 6 + 6) // 3, ((w0 - 1) // 6 * 6 + 6) // 3)  # 4.1G
        elif (tile_mode == 4):  # hw都四分之一
            crop_size = (((h0 - 1) // 8 * 8 + 8) // 4, ((w0 - 1) // 8 * 8 + 8) // 4)  # 3.7G
        ph = ((h0 - 1) // crop_size[0] + 1) * crop_size[0]
        pw = ((w0 - 1) // crop_size[1] + 1) * crop_size[1]
        x = F.pad(x, (19, 19 + pw - w0, 19, 19 + ph - h0), 'reflect')
        n, c, h, w = x.shape
        se_mean0 = torch.zeros((n, 64, 1, 1)).to(x.device)
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        n_patch = 0
        tmp_dict = {}
        opt_res_dict = {}
        for i in range(0, h - 38, crop_size[0]):
            tmp_dict[i] = {}
            for j in range(0, w - 38, crop_size[1]):
                x_crop = x[:, :, i:i + crop_size[0] + 38, j:j + crop_size[1] + 38]
                n, c1, h1, w1 = x_crop.shape
                tmp0, x_crop = self.unet1.forward_a(x_crop)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(x_crop.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(x_crop, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                n_patch += 1
                tmp_dict[i][j] = (tmp0, x_crop)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 38, crop_size[0]):
            for j in range(0, w - 38, crop_size[1]):
                tmp0, x_crop = tmp_dict[i][j]
                x_crop = self.unet1.conv2.seblock.forward_mean(x_crop, se_mean0)
                opt_unet1 = self.unet1.forward_b(tmp0, x_crop)
                tmp_x1, tmp_x2 = self.unet2.forward_a(opt_unet1)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x2.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x2, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2)
        se_mean1 /= n_patch
        se_mean0 = torch.zeros((n, 128, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean0 = se_mean0.half()
        for i in range(0, h - 38, crop_size[0]):
            for j in range(0, w - 38, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2 = tmp_dict[i][j]
                tmp_x2 = self.unet2.conv2.seblock.forward_mean(tmp_x2, se_mean1)
                tmp_x3 = self.unet2.forward_b(tmp_x2)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x3.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x3, dim=(2, 3), keepdim=True)
                se_mean0 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x2, tmp_x3)
        se_mean0 /= n_patch
        se_mean1 = torch.zeros((n, 64, 1, 1)).to(x.device)  # 64#128#128#64
        if ("Half" in x.type()):
            se_mean1 = se_mean1.half()
        for i in range(0, h - 38, crop_size[0]):
            for j in range(0, w - 38, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x2, tmp_x3 = tmp_dict[i][j]
                tmp_x3 = self.unet2.conv3.seblock.forward_mean(tmp_x3, se_mean0)
                tmp_x4 = self.unet2.forward_c(tmp_x2, tmp_x3)
                if ("Half" in x.type()):  # torch.HalfTensor/torch.cuda.HalfTensor
                    tmp_se_mean = torch.mean(tmp_x4.float(), dim=(2, 3), keepdim=True).half()
                else:
                    tmp_se_mean = torch.mean(tmp_x4, dim=(2, 3), keepdim=True)
                se_mean1 += tmp_se_mean
                tmp_dict[i][j] = (opt_unet1, tmp_x1, tmp_x4)
        se_mean1 /= n_patch
        for i in range(0, h - 38, crop_size[0]):
            opt_res_dict[i] = {}
            for j in range(0, w - 38, crop_size[1]):
                opt_unet1, tmp_x1, tmp_x4 = tmp_dict[i][j]
                tmp_x4 = self.unet2.conv4.seblock.forward_mean(tmp_x4, se_mean1)
                x0 = self.unet2.forward_d(tmp_x1, tmp_x4)
                x1 = F.pad(opt_unet1, (-20, -20, -20, -20))
                x_crop = torch.add(x0, x1)  # x0是unet2的最终输出
                x_crop = self.conv_final(x_crop)
                x_crop = F.pad(x_crop, (-1, -1, -1, -1))
                x_crop = self.ps(x_crop)
                opt_res_dict[i][j] = x_crop
        del tmp_dict
        torch.cuda.empty_cache()
        res = torch.zeros((n, c, h * 4 - 152, w * 4 - 152)).to(x.device)
        if ("Half" in x.type()):
            res = res.half()
        for i in range(0, h - 38, crop_size[0]):
            for j in range(0, w - 38, crop_size[1]):
                # print(opt_res_dict[i][j].shape,res[:, :, i * 4:i * 4 + h1 * 4 - 144, j * 4:j * 4 + w1 * 4 - 144].shape)
                res[:, :, i * 4:i * 4 + h1 * 4 - 152, j * 4:j * 4 + w1 * 4 - 152] = opt_res_dict[i][j]
        del opt_res_dict
        torch.cuda.empty_cache()
        if (w0 != pw or h0 != ph): res = res[:, :, :h0 * 4, :w0 * 4]
        res += F.interpolate(x00, scale_factor=4, mode='nearest')
        return res  #


class RealWaifuUpScaler(object):
    def __init__(self, scale, weight_path, half, device):
        weight = torch.load(weight_path, map_location="cpu")
        self.model = eval("UpCunet%sx" % scale)()
        if (half == True):
            self.model = self.model.half().to(device)
        else:
            self.model = self.model.to(device)
        self.model.load_state_dict(weight, strict=True)
        self.model.eval()
        self.half = half
        self.device = device

    def np2tensor(self, np_frame):
        if (self.half == False):
            return torch.from_numpy(np.transpose(np_frame, (2, 0, 1))).unsqueeze(0).to(self.device).float() / 255
        else:
            return torch.from_numpy(np.transpose(np_frame, (2, 0, 1))).unsqueeze(0).to(self.device).half() / 255

    def tensor2np(self, tensor):
        if (self.half == False):
            return (
                np.transpose((tensor.data.squeeze() * 255.0).round().clamp_(0, 255).byte().cpu().numpy(), (1, 2, 0)))
        else:
            return (np.transpose((tensor.data.squeeze().float() * 255.0).round().clamp_(0, 255).byte().cpu().numpy(),
                                 (1, 2, 0)))

    def __call__(self, frame, tile_mode):
        with torch.no_grad():
            tensor = self.np2tensor(frame)
            result = self.tensor2np(self.model(tensor, tile_mode))
        return result


if __name__ == "__main__":
    ###########inference_img
    import time, cv2, sys
    from time import time as ttime

    for weight_path, scale in [("weights_v3/up2x-latest-denoise3x.pth", 2), ("weights_v3/up3x-latest-denoise3x.pth", 3),
                               ("weights_v3/up4x-latest-denoise3x.pth", 4)]:
        for tile_mode in [0, 1, 2, 3, 4]:
            upscaler2x = RealWaifuUpScaler(scale, weight_path, half=True, device="cuda:0")
            input_dir = "%s/input_dir1" % root_path
            output_dir = "%s/opt-dir-all-test" % root_path
            os.makedirs(output_dir, exist_ok=True)
            for name in os.listdir(input_dir):
                print(name)
                tmp = name.split(".")
                inp_path = os.path.join(input_dir, name)
                suffix = tmp[-1]
                prefix = ".".join(tmp[:-1])
                tmp_path = os.path.join(root_path, "tmp", "%s.%s" % (int(time.time() * 1000000), suffix))
                print(inp_path, tmp_path)
                # 支持中文路径
                # os.link(inp_path, tmp_path)#win用硬链接
                os.symlink(inp_path, tmp_path)  # linux用软链接
                frame = cv2.imread(tmp_path)[:, :, [2, 1, 0]]
                t0 = ttime()
                result = upscaler2x(frame, tile_mode=tile_mode)[:, :, ::-1]
                t1 = ttime()
                print(prefix, "done", t1 - t0)
                tmp_opt_path = os.path.join(root_path, "tmp", "%s.%s" % (int(time.time() * 1000000), suffix))
                cv2.imwrite(tmp_opt_path, result)
                n = 0
                while (1):
                    if (n == 0):
                        suffix = "_%sx_tile%s.png" % (scale, tile_mode)
                    else:
                        suffix = "_%sx_tile%s_%s.png" % (scale, tile_mode, n)  #
                    if (os.path.exists(os.path.join(output_dir, prefix + suffix)) == False):
                        break
                    else:
                        n += 1
                final_opt_path = os.path.join(output_dir, prefix + suffix)
                os.rename(tmp_opt_path, final_opt_path)
                os.remove(tmp_path)