archs/arch_util.py

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

try:
    from .nafnet_utils.arch_util import LayerNorm2d
    from .nafnet_utils.arch_model import SimpleGate
except:
    from nafnet_utils.arch_util import LayerNorm2d
    from nafnet_utils.arch_model import SimpleGate

'''
https://github.com/wangchx67/FourLLIE.git
'''

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 make_layer(block, n_layers):
    layers = []
    for _ in range(n_layers):
        layers.append(block())
    return nn.Sequential(*layers)


class ResidualBlock_noBN(nn.Module):
    '''Residual block w/o BN
    ---Conv-ReLU-Conv-+-
     |________________|
    '''

    def __init__(self, nf=64):
        super(ResidualBlock_noBN, self).__init__()
        self.conv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
        self.conv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)

        # initialization
        initialize_weights([self.conv1, self.conv2], 0.1)

    def forward(self, x):
        identity = x
        out = F.relu(self.conv1(x), inplace=True)
        out = self.conv2(out)
        return identity + out

class SpaBlock(nn.Module):
    def __init__(self, nc):
        super(SpaBlock, self).__init__()
        self.block = nn.Sequential(
            nn.Conv2d(nc,nc,3,1,1),
            nn.LeakyReLU(0.1,inplace=True),
            nn.Conv2d(nc, nc, 3, 1, 1),
            nn.LeakyReLU(0.1, inplace=True))

    def forward(self, x):
        return x+self.block(x)

class FreBlock(nn.Module):
    def __init__(self, nc):
        super(FreBlock, self).__init__()
        self.fpre = nn.Conv2d(nc, nc, 1, 1, 0)
        self.process1 = nn.Sequential(
            nn.Conv2d(nc, nc, 1, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(nc, nc, 1, 1, 0))
        self.process2 = nn.Sequential(
            nn.Conv2d(nc, nc, 1, 1, 0),
            nn.LeakyReLU(0.1, inplace=True),
            nn.Conv2d(nc, nc, 1, 1, 0))

    def forward(self, x):
        _, _, H, W = x.shape
        x_freq = torch.fft.rfft2(self.fpre(x), norm='backward')
        mag = torch.abs(x_freq)
        pha = torch.angle(x_freq)
        mag = self.process1(mag)
        pha = self.process2(pha)
        real = mag * torch.cos(pha)
        imag = mag * torch.sin(pha)
        x_out = torch.complex(real, imag)
        x_out = torch.fft.irfft2(x_out, s=(H, W), norm='backward')

        return x_out+x

class ProcessBlock(nn.Module):
    def __init__(self, in_nc, spatial = True):
        super(ProcessBlock,self).__init__()
        self.spatial = spatial
        self.spatial_process = SpaBlock(in_nc) if spatial else nn.Identity()
        self.frequency_process = FreBlock(in_nc)
        self.cat = nn.Conv2d(2*in_nc,in_nc,1,1,0) if spatial else nn.Conv2d(in_nc,in_nc,1,1,0)

    def forward(self, x):
        xori = x
        x_freq = self.frequency_process(x)
        x_spatial = self.spatial_process(x)
        xcat = torch.cat([x_spatial,x_freq],1)
        x_out = self.cat(xcat) if self.spatial else self.cat(x_freq)

        return x_out+xori

class Attention_Light(nn.Module):
    
    def __init__(self, img_channels = 3, width = 16, spatial = False):
        super(Attention_Light, self).__init__()
        self.block = nn.Sequential(
                nn.Conv2d(in_channels = img_channels, out_channels = width//2, kernel_size = 1, padding = 0, stride = 1, groups = 1, bias = True),
                ProcessBlock(in_nc = width //2, spatial = spatial),
                nn.Conv2d(in_channels = width//2, out_channels = width, kernel_size = 1, padding = 0, stride = 1, groups = 1, bias = True),
                ProcessBlock(in_nc = width, spatial = spatial),
                nn.Conv2d(in_channels = width, out_channels = width, kernel_size = 1, padding = 0, stride = 1, groups = 1, bias = True),
                ProcessBlock(in_nc=width, spatial = spatial),
                nn.Sigmoid()
                    )
    def forward(self, input):
        return self.block(input)

class Branch(nn.Module):
    '''
    Branch that lasts lonly the dilated convolutions
    '''
    def __init__(self, c, DW_Expand, dilation = 1, extra_depth_wise = False):
        super().__init__()
        self.dw_channel = DW_Expand * c 
        self.branch = nn.Sequential(
                       nn.Conv2d(c, c, kernel_size=3, padding=1, stride=1, groups=c, bias=True, dilation=1) if extra_depth_wise else nn.Identity(), #optional extra dw
                       nn.Conv2d(in_channels=c, out_channels=self.dw_channel, kernel_size=1, padding=0, stride=1, groups=1, bias=True, dilation = 1),
                       nn.Conv2d(in_channels=self.dw_channel, out_channels=self.dw_channel, kernel_size=3, padding=dilation, stride=1, groups=self.dw_channel,
                                            bias=True, dilation = dilation) # the dconv
        )
    def forward(self, input):
        return self.branch(input)
    
class EBlock(nn.Module):
    '''
    Change this block using Branch
    '''
    
    def __init__(self, c, DW_Expand=2, FFN_Expand=2, dilations = [1], extra_depth_wise = False):
        super().__init__()
        #we define the 2 branches
        
        self.branches = nn.ModuleList()
        for dilation in dilations:
            self.branches.append(Branch(c, DW_Expand, dilation = dilation, extra_depth_wise=extra_depth_wise))
            
        assert len(dilations) == len(self.branches)
        self.dw_channel = DW_Expand * c 
        self.sca = nn.Sequential(
                       nn.AdaptiveAvgPool2d(1),
                       nn.Conv2d(in_channels=self.dw_channel // 2, out_channels=self.dw_channel // 2, kernel_size=1, padding=0, stride=1,
                       groups=1, bias=True, dilation = 1),  
        )
        self.sg1 = SimpleGate()
        self.sg2 = SimpleGate()
        self.conv3 = nn.Conv2d(in_channels=self.dw_channel // 2, out_channels=c, kernel_size=1, padding=0, stride=1, groups=1, bias=True, dilation = 1)
        ffn_channel = FFN_Expand * c
        self.conv4 = nn.Conv2d(in_channels=c, out_channels=ffn_channel, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
        self.conv5 = nn.Conv2d(in_channels=ffn_channel // 2, out_channels=c, kernel_size=1, padding=0, stride=1, groups=1, bias=True)

        self.norm1 = LayerNorm2d(c)
        self.norm2 = LayerNorm2d(c)

        self.gamma = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)
        self.beta = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)

    def forward(self, inp):

        y = inp
        x = self.norm1(inp)
        z = 0
        for branch in self.branches:
            z += branch(x)
        
        z = self.sg1(z)
        x = self.sca(z) * z
        x = self.conv3(x)
        y = inp + self.beta * x
        #second step
        x = self.conv4(self.norm2(y)) # size [B, 2*C, H, W]
        x = self.sg2(x)  # size [B, C, H, W]
        x = self.conv5(x) # size [B, C, H, W]

        return y + x * self.gamma

#----------------------------------------------------------------------------------------------
if __name__ == '__main__':
    
    img_channel = 3
    width = 32

    enc_blks = [1, 2, 3]
    middle_blk_num = 3
    dec_blks = [3, 1, 1]
    dilations = [1, 4, 9]
    extra_depth_wise = False
    
    # net = NAFNet(img_channel=img_channel, width=width, middle_blk_num=middle_blk_num,
    #                   enc_blk_nums=enc_blks, dec_blk_nums=dec_blks)
    net  = EBlock(c = img_channel, 
                            dilations = dilations,
                            extra_depth_wise=extra_depth_wise)

    inp_shape = (3, 256, 256)

    from ptflops import get_model_complexity_info

    macs, params = get_model_complexity_info(net, inp_shape, verbose=False, print_per_layer_stat=True)


    print(macs, params)