可逆模块的实现思路和示例

可逆模块的实现思路和示例

论文：Invertible Image Signal Processing

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import torch.nn.init as init

from .modules import InvertibleConv1x1


def initialize_weights(net_l, scale=1):
    if not isinstance(net_l, list):
        net_l = [net_l]
    for net in net_l:
        for m in net.modules():
            if isinstance(m, nn.Conv2d):
                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
                m.weight.data *= scale  # for residual block
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
                m.weight.data *= scale
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                init.constant_(m.weight, 1)
                init.constant_(m.bias.data, 0.0)


def initialize_weights_xavier(net_l, scale=1):
    if not isinstance(net_l, list):
        net_l = [net_l]
    for net in net_l:
        for m in net.modules():
            if isinstance(m, nn.Conv2d):
                init.xavier_normal_(m.weight)
                m.weight.data *= scale  # for residual block
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                init.xavier_normal_(m.weight)
                m.weight.data *= scale
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                init.constant_(m.weight, 1)
                init.constant_(m.bias.data, 0.0)


class DenseBlock(nn.Module):
    def __init__(self, channel_in, channel_out, init='xavier', gc=32, bias=True):
        super(DenseBlock, self).__init__()
        self.conv1 = nn.Conv2d(channel_in, gc, 3, 1, 1, bias=bias)
        self.conv2 = nn.Conv2d(channel_in + gc, gc, 3, 1, 1, bias=bias)
        self.conv3 = nn.Conv2d(channel_in + 2 * gc, gc, 3, 1, 1, bias=bias)
        self.conv4 = nn.Conv2d(channel_in + 3 * gc, gc, 3, 1, 1, bias=bias)
        self.conv5 = nn.Conv2d(channel_in + 4 * gc, channel_out, 3, 1, 1, bias=bias)
        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)

        if init == 'xavier':
            initialize_weights_xavier([self.conv1, self.conv2, self.conv3, self.conv4], 0.1)
        else:
            initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4], 0.1)
        initialize_weights(self.conv5, 0)
    
    def forward(self, x):
        x1 = self.lrelu(self.conv1(x))
        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))

        return x5

def subnet(net_structure, init='xavier'):
    def constructor(channel_in, channel_out):
        if net_structure == 'DBNet':
            if init == 'xavier':
                return DenseBlock(channel_in, channel_out, init)
            else:
                return DenseBlock(channel_in, channel_out)
            # return UNetBlock(channel_in, channel_out)
        else:
            return None

    return constructor


class InvBlock(nn.Module):
    def __init__(self, subnet_constructor, channel_num, channel_split_num, clamp=0.8):
        super(InvBlock, self).__init__()
        # channel_num: 3
        # channel_split_num: 1

        self.split_len1 = channel_split_num # 1
        self.split_len2 = channel_num - channel_split_num # 2 

        self.clamp = clamp

        self.F = subnet_constructor(self.split_len2, self.split_len1)
        self.G = subnet_constructor(self.split_len1, self.split_len2)
        self.H = subnet_constructor(self.split_len1, self.split_len2)

        in_channels = 3        
        self.invconv = InvertibleConv1x1(in_channels, LU_decomposed=True)
        self.flow_permutation = lambda z, logdet, rev: self.invconv(z, logdet, rev)
        
    def forward(self, x, rev=False):
        if not rev:            
            # invert1x1conv 
            x, logdet = self.flow_permutation(x, logdet=0, rev=False) 
            
            # split to 1 channel and 2 channel. 
            x1, x2 = (x.narrow(1, 0, self.split_len1), x.narrow(1, self.split_len1, self.split_len2)) 

            y1 = x1 + self.F(x2) # 1 channel 
            self.s = self.clamp * (torch.sigmoid(self.H(y1)) * 2 - 1)
            y2 = x2.mul(torch.exp(self.s)) + self.G(y1) # 2 channel 
            out = torch.cat((y1, y2), 1)
        else:
            # split. 
            x1, x2 = (x.narrow(1, 0, self.split_len1), x.narrow(1, self.split_len1, self.split_len2)) 
            self.s = self.clamp * (torch.sigmoid(self.H(x1)) * 2 - 1)
            y2 = (x2 - self.G(x1)).div(torch.exp(self.s)) 
            y1 = x1 - self.F(y2) 

            x = torch.cat((y1, y2), 1)            

            # inv permutation 
            out, logdet = self.flow_permutation(x, logdet=0, rev=True)

        return out

class InvISPNet(nn.Module):
    def __init__(self, channel_in=3, channel_out=3, subnet_constructor=subnet('DBNet'), block_num=8):
        super(InvISPNet, self).__init__()
        operations = []

        current_channel = channel_in
        channel_num = channel_in
        channel_split_num = 1

        for j in range(block_num): 
            b = InvBlock(subnet_constructor, channel_num, channel_split_num) # one block is one flow step. 
            operations.append(b)
        
        self.operations = nn.ModuleList(operations)

        self.initialize()

    def initialize(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                init.xavier_normal_(m.weight)
                m.weight.data *= 1.  # for residual block
                if m.bias is not None:
                    m.bias.data.zero_() 
            elif isinstance(m, nn.Linear):
                init.xavier_normal_(m.weight)
                m.weight.data *= 1.
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                init.constant_(m.weight, 1)
                init.constant_(m.bias.data, 0.0)
    
    def forward(self, x, rev=False):
        out = x # x: [N,3,H,W] 
        
        if not rev: 
            for op in self.operations:
                out = op.forward(out, rev)
        else:
            for op in reversed(self.operations):
                out = op.forward(out, rev)
        
        return out