import torch
import torch.nn as nn
##---------- Basic Layers ----------
def conv3x3(in_chn, out_chn, bias=True):
layer = nn.Conv2d(in_chn, out_chn, kernel_size=3, stride=1, padding=1, bias=bias)
return layer
def conv(in_channels, out_channels, kernel_size, bias=False, stride=1):
return nn.Conv2d(
in_channels, out_channels, kernel_size,
padding=(kernel_size // 2), bias=bias, stride=stride)
def bili_resize(factor):
return nn.Upsample(scale_factor=factor, mode='bilinear', align_corners=False)
##---------- Basic Blocks ----------
class UNetConvBlock(nn.Module):
def __init__(self, in_size, out_size, downsample):
super(UNetConvBlock, self).__init__()
self.downsample = downsample
self.block = SK_RDB(in_channels=in_size, growth_rate=out_size, num_layers=3)
if downsample:
self.downsample = PS_down(out_size, out_size, downscale=2)
def forward(self, x):
out = self.block(x)
if self.downsample:
out_down = self.downsample(out)
return out_down, out
else:
return out
class UNetUpBlock(nn.Module):
def __init__(self, in_size, out_size):
super(UNetUpBlock, self).__init__()
# self.up = nn.ConvTranspose2d(in_size, out_size, kernel_size=2, stride=2, bias=True)
self.up = PS_up(in_size, out_size, upscale=2)
self.conv_block = UNetConvBlock(in_size, out_size, False)
def forward(self, x, bridge):
up = self.up(x)
out = torch.cat([up, bridge], dim=1)
out = self.conv_block(out)
return out
##---------- Resizing Modules (Pixel(Un)Shuffle) ----------
class PS_down(nn.Module):
def __init__(self, in_size, out_size, downscale):
super(PS_down, self).__init__()
self.UnPS = nn.PixelUnshuffle(downscale)
self.conv1 = nn.Conv2d((downscale**2) * in_size, out_size, 1, 1, 0)
def forward(self, x):
x = self.UnPS(x) # h/2, w/2, 4*c
x = self.conv1(x)
return x
class PS_up(nn.Module):
def __init__(self, in_size, out_size, upscale):
super(PS_up, self).__init__()
self.PS = nn.PixelShuffle(upscale)
self.conv1 = nn.Conv2d(in_size//(upscale**2), out_size, 1, 1, 0)
def forward(self, x):
x = self.PS(x) # h/2, w/2, 4*c
x = self.conv1(x)
return x
##---------- Selective Kernel Feature Fusion (SKFF) ----------
class SKFF(nn.Module):
def __init__(self, in_channels, height=3, reduction=8, bias=False):
super(SKFF, self).__init__()
self.height = height
d = max(int(in_channels / reduction), 4)
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.conv_du = nn.Sequential(nn.Conv2d(in_channels, d, 1, padding=0, bias=bias), nn.PReLU())
self.fcs = nn.ModuleList([])
for i in range(self.height):
self.fcs.append(nn.Conv2d(d, in_channels, kernel_size=1, stride=1, bias=bias))
self.softmax = nn.Softmax(dim=1)
def forward(self, inp_feats):
batch_size, n_feats, H, W = inp_feats[1].shape
inp_feats = torch.cat(inp_feats, dim=1)
inp_feats = inp_feats.view(batch_size, self.height, n_feats, inp_feats.shape[2], inp_feats.shape[3])
feats_U = torch.sum(inp_feats, dim=1)
feats_S = self.avg_pool(feats_U)
feats_Z = self.conv_du(feats_S)
attention_vectors = [fc(feats_Z) for fc in self.fcs]
attention_vectors = torch.cat(attention_vectors, dim=1)
attention_vectors = attention_vectors.view(batch_size, self.height, n_feats, 1, 1)
attention_vectors = self.softmax(attention_vectors)
feats_V = torch.sum(inp_feats * attention_vectors, dim=1)
return feats_V
##---------- Dense Block ----------
class DenseLayer(nn.Module):
def __init__(self, in_channels, out_channels, I):
super(DenseLayer, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=3 // 2)
self.relu = nn.ReLU(inplace=True)
self.sk = SKFF(out_channels, height=2, reduction=8, bias=False)
def forward(self, x):
x1 = self.relu(self.conv(x))
# output = torch.cat([x, x1], 1) # -> RDB
output = self.sk((x, x1))
return output
##---------- Selective Kernel Residual Dense Block (SK-RDB) ----------
class SK_RDB(nn.Module):
def __init__(self, in_channels, growth_rate, num_layers):
super(SK_RDB, self).__init__()
self.identity = nn.Conv2d(in_channels, growth_rate, 1, 1, 0)
self.layers = nn.Sequential(
*[DenseLayer(in_channels, in_channels, I=i) for i in range(num_layers)]
)
self.lff = nn.Conv2d(in_channels, growth_rate, kernel_size=1)
def forward(self, x):
res = self.identity(x)
x = self.layers(x)
x = self.lff(x)
return res + x
##---------- testNet ----------
class SRMNet(nn.Module):
def __init__(self, in_chn=3, wf=96, depth=4):
super(SRMNet, self).__init__()
self.depth = depth
self.down_path = nn.ModuleList()
self.bili_down = bili_resize(0.5)
self.conv_01 = nn.Conv2d(in_chn, wf, 3, 1, 1)
# encoder of UNet-64
prev_channels = 0
for i in range(depth): # 0,1,2,3
downsample = True if (i + 1) < depth else False
self.down_path.append(UNetConvBlock(prev_channels + wf, (2 ** i) * wf, downsample))
prev_channels = (2 ** i) * wf
# decoder of UNet-64
self.up_path = nn.ModuleList()
self.skip_conv = nn.ModuleList()
self.conv_up = nn.ModuleList()
self.bottom_conv = nn.Conv2d(prev_channels, wf, 3, 1, 1)
self.bottom_up = bili_resize(2 ** (depth-1))
for i in reversed(range(depth - 1)):
self.up_path.append(UNetUpBlock(prev_channels, (2 ** i) * wf))
self.skip_conv.append(nn.Conv2d((2 ** i) * wf, (2 ** i) * wf, 3, 1, 1))
self.conv_up.append(nn.Sequential(*[bili_resize(2 ** i), nn.Conv2d((2 ** i) * wf, wf, 3, 1, 1)]))
# *[nn.Conv2d((2 ** i) * wf, wf, 3, 1, 1), bili_resize(2 ** i)])
prev_channels = (2 ** i) * wf
self.final_ff = SKFF(in_channels=wf, height=depth)
self.last = conv3x3(prev_channels, in_chn, bias=True)
def forward(self, x):
img = x
scale_img = img
##### shallow conv #####
x1 = self.conv_01(img)
encs = []
######## UNet-64 ########
# Down-path (Encoder)
for i, down in enumerate(self.down_path):
if i == 0: # top layer
x1, x1_up = down(x1)
encs.append(x1_up)
elif (i + 1) < self.depth: # middle layer
scale_img = self.bili_down(scale_img)
left_bar = self.conv_01(scale_img)
x1 = torch.cat([x1, left_bar], dim=1)
x1, x1_up = down(x1)
encs.append(x1_up)
else: # lowest layer
scale_img = self.bili_down(scale_img)
left_bar = self.conv_01(scale_img)
x1 = torch.cat([x1, left_bar], dim=1)
x1 = down(x1)
# Up-path (Decoder)
ms_result = [self.bottom_up(self.bottom_conv(x1))]
for i, up in enumerate(self.up_path):
x1 = up(x1, self.skip_conv[i](encs[-i - 1]))
ms_result.append(self.conv_up[i](x1))
# Multi-scale selective feature fusion
msff_result = self.final_ff(ms_result)
##### Reconstruct #####
out_1 = self.last(msff_result) + img
return out_1
if __name__ == "__main__":
from thop import profile
input = torch.ones(1, 3, 256, 256, dtype=torch.float, requires_grad=False)
model = SRMNet(in_chn=3, wf=96, depth=4)
out = model(input)
flops, params = profile(model, inputs=(input,))
# RDBlayer = SK_RDB(in_channels=64, growth_rate=64, num_layers=3)
# print(RDBlayer)
# out = RDBlayer(input)
# flops, params = profile(RDBlayer, inputs=(input,))
print('input shape:', input.shape)
print('parameters:', params/1e6)
print('flops', flops/1e9)
print('output shape', out.shape)