CalvinBackup/SimSwapPlus
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
chenxuanhong
2022-01-10 15:03:58 +08:00
parent 573689a591
commit 3783ef0e75
57 changed files with 9520 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
components/Conditional_Discriminator_Projection_big.py
+124
View File
@@ -0,0 +1,124 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_Discriminator copy.py
# Created Date: Saturday April 18th 2020
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 29th June 2021 4:26:33 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2020 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from torch.nn import utils
class Discriminator(nn.Module):
def __init__(self, chn=32, k_size=3, n_class=3):
super().__init__()
# padding_size = int((k_size -1)/2)
slop = 0.2
enable_bias = True
# stage 1
self.block1 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size= k_size, stride = 2, padding=2,bias= enable_bias)),
nn.LeakyReLU(slop),
utils.spectral_norm(nn.Conv2d(in_channels = chn, out_channels = chn * 2 , kernel_size= k_size, stride = 2,padding=2, bias= enable_bias)), # 1/4
nn.LeakyReLU(slop)
)
self.aux_classfier1 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = chn * 2 , out_channels = chn , kernel_size= 5, bias=enable_bias)),
nn.LeakyReLU(slop),
nn.AdaptiveAvgPool2d(1),
)
self.embed1 = utils.spectral_norm(nn.Embedding(n_class, chn))
self.linear1= utils.spectral_norm(nn.Linear(chn, 1))
# stage 2
self.block2 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = chn * 2 , out_channels = chn * 4 , kernel_size= k_size, stride = 2, padding=2, bias= enable_bias)),# 1/8
nn.LeakyReLU(slop),
utils.spectral_norm(nn.Conv2d(in_channels = chn * 4, out_channels = chn * 8 , kernel_size= k_size, stride = 2,padding=2, bias= enable_bias)),# 1/16
nn.LeakyReLU(slop)
)
self.aux_classfier2 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = chn * 8 , out_channels = chn , kernel_size= 5, bias= enable_bias)),
nn.LeakyReLU(slop),
nn.AdaptiveAvgPool2d(1),
)
self.embed2 = utils.spectral_norm(nn.Embedding(n_class, chn))
self.linear2= utils.spectral_norm(nn.Linear(chn, 1))
# stage 3
self.block3 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = chn * 8 , out_channels = chn * 8 , kernel_size= k_size, stride = 2,padding=3, bias= enable_bias)),# 1/32
nn.LeakyReLU(slop),
utils.spectral_norm(nn.Conv2d(in_channels = chn * 8, out_channels = chn * 16 , kernel_size= k_size, stride = 2,padding=3, bias= enable_bias)),# 1/64
nn.LeakyReLU(slop)
)
self.aux_classfier3 = nn.Sequential(
utils.spectral_norm(nn.Conv2d(in_channels = chn * 16 , out_channels = chn, kernel_size= 5, bias= enable_bias)),
nn.LeakyReLU(slop),
nn.AdaptiveAvgPool2d(1),
)
self.embed3 = utils.spectral_norm(nn.Embedding(n_class, chn))
self.linear3= utils.spectral_norm(nn.Linear(chn, 1))
self.__weights_init__()
def __weights_init__(self):
print("Init weights")
for m in self.modules():
if isinstance(m,nn.Conv2d) or isinstance(m,nn.Linear):
nn.init.xavier_uniform_(m.weight)
try:
nn.init.zeros_(m.bias)
except:
print("No bias found!")
if isinstance(m, nn.Embedding):
nn.init.xavier_uniform_(m.weight)
def forward(self, input, condition):
h = self.block1(input)
prep1 = self.aux_classfier1(h)
prep1 = prep1.view(prep1.size()[0], -1)
y1 = self.embed1(condition)
y1 = torch.sum(y1 * prep1, dim=1, keepdim=True)
prep1 = self.linear1(prep1) + y1
h = self.block2(h)
prep2 = self.aux_classfier2(h)
prep2 = prep2.view(prep2.size()[0], -1)
y2 = self.embed2(condition)
y2 = torch.sum(y2 * prep2, dim=1, keepdim=True)
prep2 = self.linear2(prep2) + y2
h = self.block3(h)
prep3 = self.aux_classfier3(h)
prep3 = prep3.view(prep3.size()[0], -1)
y3 = self.embed3(condition)
y3 = torch.sum(y3 * prep3, dim=1, keepdim=True)
prep3 = self.linear3(prep3) + y3
out_prep = [prep1,prep2,prep3]
return out_prep
def get_outputs_len(self):
num = 0
for m in self.modules():
if isinstance(m,nn.Linear):
num+=1
return num
if __name__ == "__main__":
wocao = Discriminator().cuda()
from torchsummary import summary
summary(wocao, input_size=(3, 512, 512))
components/Conditional_Generator_Noskip.py
+114
View File
@@ -0,0 +1,114 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_Generator_tanh.py
# Created Date: Saturday April 18th 2020
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 6th July 2021 1:16:46 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2020 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
from components.Conditional_ResBlock_ModulaConv import Conditional_ResBlock
class Generator(nn.Module):
def __init__(
self,
chn=32,
k_size=3,
res_num = 5,
class_num = 3,
**kwargs):
super().__init__()
padding_size = int((k_size -1)/2)
self.resblock_list = []
self.n_class = class_num
self.encoder1 = nn.Sequential(
# nn.InstanceNorm2d(3, affine=True),
# nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size= k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
# nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size= k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn*2, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
# nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = chn*2, out_channels = chn * 4, kernel_size= k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
# nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = chn*4 , out_channels = chn * 8, kernel_size= k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
# # nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size= k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU()
)
res_size = chn * 8
for _ in range(res_num-1):
self.resblock_list += [ResBlock(res_size,k_size),]
self.resblocks = nn.Sequential(*self.resblock_list)
self.conditional_res = Conditional_ResBlock(res_size, k_size, class_num)
self.decoder1 = nn.Sequential(
DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size= k_size),
nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
DeConv(in_channels = chn * 8, out_channels = chn *4, kernel_size= k_size),
nn.InstanceNorm2d(chn *4, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
DeConv(in_channels = chn * 4, out_channels = chn * 2 , kernel_size= k_size),
nn.InstanceNorm2d(chn*2, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
DeConv(in_channels = chn *2, out_channels = chn, kernel_size= k_size),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
# nn.ReLU(),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn, out_channels =3, kernel_size= k_size, stride=1, padding=1,bias =True)
# nn.Tanh()
)
self.__weights_init__()
def __weights_init__(self):
for layer in self.encoder1:
if isinstance(layer,nn.Conv2d):
nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input, condition=None, get_feature = False):
feature = self.encoder1(input)
if get_feature:
return feature
out = self.conditional_res(feature, condition)
out = self.resblocks(out)
# n, _,h,w = out.size()
# attr = condition.view((n, self.n_class, 1, 1)).expand((n, self.n_class, h, w))
# out = torch.cat([out, attr], dim=1)
out = self.decoder1(out)
return out,feature
components/Conditional_ResBlock_ModulaConv.py
+82
View File
@@ -0,0 +1,82 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_ResBlock_v2.py
# Created Date: Tuesday June 29th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 29th June 2021 3:59:44 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
# -*- coding:utf-8 -*-
###################################################################
### @FilePath: \ASMegaGAN\components\Conditional_ResBlock_v2.py
### @Author: Ziang Liu
### @Date: 2021-06-28 21:30:17
### @LastEditors: Ziang Liu
### @LastEditTime: 2021-06-28 21:46:24
### @Copyright (C) 2021 SJTU. All rights reserved.
###################################################################
import torch
from torch import nn
import torch.nn.functional as F
# from ops.Conditional_BN import Conditional_BN
# from components.Adain import Adain
class Conv2DMod(nn.Module):
def __init__(self, in_channels, out_channels, kernel, demod=True, stride=1, dilation=1, eps = 1e-8, **kwargs):
super().__init__()
self.filters = out_channels
self.demod = demod
self.kernel = kernel
self.stride = stride
self.dilation = dilation
self.weight = nn.Parameter(torch.randn((out_channels, in_channels, kernel, kernel)))
self.eps = eps
padding_size = int((kernel -1)/2)
self.same_padding = nn.ReplicationPad2d(padding_size)
nn.init.kaiming_normal_(self.weight, a=0, mode='fan_in', nonlinearity='leaky_relu')
def forward(self, x, y):
b, c, h, w = x.shape
w1 = y[:, None, :, None, None]
w2 = self.weight[None, :, :, :, :]
weights = w2 * (w1 + 1)
if self.demod:
d = torch.rsqrt((weights ** 2).sum(dim=(2, 3, 4), keepdim=True) + self.eps)
weights = weights * d
x = x.reshape(1, -1, h, w)
_, _, *ws = weights.shape
weights = weights.reshape(b * self.filters, *ws)
x = self.same_padding(x)
x = F.conv2d(x, weights, groups=b)
x = x.reshape(-1, self.filters, h, w)
return x
class Conditional_ResBlock(nn.Module):
def __init__(self, in_channel, k_size = 3, n_class = 2, stride=1):
super().__init__()
self.embed1 = nn.Embedding(n_class, in_channel)
self.embed2 = nn.Embedding(n_class, in_channel)
self.conv1 = Conv2DMod(in_channels = in_channel , out_channels = in_channel, kernel= k_size, stride=stride)
self.conv2 = Conv2DMod(in_channels = in_channel , out_channels = in_channel, kernel= k_size, stride=stride)
def forward(self, input, condition):
res = input
style1 = self.embed1(condition)
h = self.conv1(res, style1)
style2 = self.embed2(condition)
h = self.conv2(h, style2)
out = h + res
return out
components/DeConv.py
+20
View File
@@ -0,0 +1,20 @@
import torch
from torch import nn
class DeConv(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size = 3, upsampl_scale = 2):
super().__init__()
self.upsampling = nn.UpsamplingNearest2d(scale_factor=upsampl_scale)
padding_size = int((kernel_size -1)/2)
# self.same_padding = nn.ReflectionPad2d(padding_size)
self.conv = nn.Conv2d(in_channels = in_channels ,padding=padding_size, out_channels = out_channels , kernel_size= kernel_size, bias= False)
self.__weights_init__()
def __weights_init__(self):
nn.init.xavier_uniform_(self.conv.weight)
def forward(self, input):
h = self.upsampling(input)
# h = self.same_padding(h)
h = self.conv(h)
return h
components/FastNST.py
+156
View File
@@ -0,0 +1,156 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_Generator_gpt_LN_encoder copy.py
# Created Date: Saturday October 9th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 26th October 2021 3:25:47 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.DeConv import DeConv
from components.network_swin import SwinTransformerBlock, PatchEmbed, PatchUnEmbed
class ImageLN(nn.Module):
def __init__(self, dim) -> None:
super().__init__()
self.layer = nn.LayerNorm(dim)
def forward(self, x):
y = self.layer(x.permute(0,2,3,1)).permute(0,3,1,2)
return y
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
class_num = kwargs["n_class"]
window_size = kwargs["window_size"]
image_size = kwargs["image_size"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
embed_dim = 96
window_size = 8
num_heads = 8
mlp_ratio = 2.
norm_layer = nn.LayerNorm
qk_scale = None
qkv_bias = True
self.patch_norm = True
self.lnnorm = norm_layer(embed_dim)
self.encoder = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
ImageLN(chn),
nn.ReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
ImageLN(chn * 2),
nn.ReLU(),
nn.Conv2d(in_channels = chn*2, out_channels = embed_dim, kernel_size=k_size, stride=2, padding=1,bias =False),
ImageLN(embed_dim),
nn.ReLU(),
)
# self.encoder2 = nn.Sequential(
# nn.Conv2d(in_channels = chn*4 , out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU()
# )
self.fea_size = (image_size//4, image_size//4)
# self.conditional_GPT = GPT_Spatial(2, res_dim, res_num, class_num)
# build blocks
self.blocks = nn.ModuleList([
SwinTransformerBlock(dim=embed_dim, input_resolution=self.fea_size,
num_heads=num_heads, window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=0.0, attn_drop=0.0,
drop_path=0.1,
norm_layer=norm_layer)
for i in range(res_num)])
self.decoder = nn.Sequential(
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn *4, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
# nn.LeakyReLU(),
DeConv(in_channels = embed_dim, out_channels = chn * 2 , kernel_size=k_size),
# nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
ImageLN(chn * 2),
nn.ReLU(),
DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
ImageLN(chn),
nn.ReLU(),
nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
self.patch_embed = PatchEmbed(
img_size=self.fea_size[0], patch_size=1, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
self.patch_unembed = PatchUnEmbed(
img_size=self.fea_size[0], patch_size=1, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
x2 = self.encoder(input)
x2 = self.patch_embed(x2)
for blk in self.blocks:
x2 = blk(x2,self.fea_size)
x2 = self.lnnorm(x2)
x2 = self.patch_unembed(x2,self.fea_size)
out = self.decoder(x2)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/FastNST_CNN.py
+129
View File
@@ -0,0 +1,129 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_Generator_gpt_LN_encoder copy.py
# Created Date: Saturday October 9th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Monday, 11th October 2021 5:22:22 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
class ImageLN(nn.Module):
def __init__(self, dim) -> None:
super().__init__()
self.layer = nn.LayerNorm(dim)
def forward(self, x):
y = self.layer(x.permute(0,2,3,1)).permute(0,3,1,2)
return y
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
class_num = kwargs["n_class"]
window_size = kwargs["window_size"]
image_size = kwargs["image_size"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
embed_dim = 96
window_size = 8
num_heads = 8
mlp_ratio = 2.
norm_layer = nn.LayerNorm
qk_scale = None
qkv_bias = True
self.patch_norm = True
self.lnnorm = norm_layer(embed_dim)
self.encoder = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn * 2),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*2, out_channels = embed_dim, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(embed_dim),
nn.LeakyReLU(),
)
# self.encoder2 = nn.Sequential(
# nn.Conv2d(in_channels = chn*4 , out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU()
# )
self.decoder = nn.Sequential(
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn *4, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
# nn.LeakyReLU(),
DeConv(in_channels = embed_dim, out_channels = chn * 2 , kernel_size=k_size),
# nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.InstanceNorm2d(chn * 2),
nn.LeakyReLU(),
DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
nn.InstanceNorm2d(chn),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
x2 = self.encoder(input)
out = self.decoder(x2)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/FastNST_CNN_Resblock.py
+110
View File
@@ -0,0 +1,110 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Conditional_Generator_gpt_LN_encoder copy.py
# Created Date: Saturday October 9th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 7:35:08 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
class ImageLN(nn.Module):
def __init__(self, dim) -> None:
super().__init__()
self.layer = nn.LayerNorm(dim)
def forward(self, x):
y = self.layer(x.permute(0,2,3,1)).permute(0,3,1,2)
return y
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
self.encoder = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*2, out_channels = chn*4, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*4 , out_channels = chn * 4, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
nn.LeakyReLU(),
)
for _ in range(res_num):
self.resblock_list += [ResBlock(chn * 4,k_size),]
self.resblocks = nn.Sequential(*self.resblock_list)
self.decoder = nn.Sequential(
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
DeConv(in_channels = chn * 4, out_channels = chn *2, kernel_size=k_size),
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
DeConv(in_channels = chn * 2, out_channels = chn * 2 , kernel_size=k_size),
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.ReLU(),
DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
nn.ReLU(),
nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
x2 = self.encoder(input)
x2 = self.resblocks(x2)
out = self.decoder(x2)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/FastNST_Liif.py
+144
View File
@@ -0,0 +1,144 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: FastNST_Liif.py
# Created Date: Thursday October 14th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 2:39:09 am
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
from components.Liif import LIIF
class ImageLN(nn.Module):
def __init__(self, dim) -> None:
super().__init__()
self.layer = nn.LayerNorm(dim)
def forward(self, x):
y = self.layer(x.permute(0,2,3,1)).permute(0,3,1,2)
return y
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
class_num = kwargs["n_class"]
window_size = kwargs["window_size"]
image_size = kwargs["image_size"]
batch_size = kwargs["batch_size"]
# mlp_in_dim = kwargs["mlp_in_dim"]
# mlp_out_dim = kwargs["mlp_out_dim"]
mlp_hidden_list = kwargs["mlp_hidden_list"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
embed_dim = 96
window_size = 8
num_heads = 8
mlp_ratio = 2.
norm_layer = nn.LayerNorm
qk_scale = None
qkv_bias = True
self.patch_norm = True
self.lnnorm = norm_layer(embed_dim)
self.encoder = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn * 2),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*2, out_channels = chn*4, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*4 , out_channels = chn * 4, kernel_size=k_size, stride=2, padding=1,bias =False),
ImageLN(chn * 4),
nn.LeakyReLU(),
)
for _ in range(res_num):
self.resblock_list += [ResBlock(chn * 4,k_size),]
self.resblocks = nn.Sequential(*self.resblock_list)
# self.encoder2 = nn.Sequential(
# nn.Conv2d(in_channels = chn*4 , out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU()
# )
self.decoder = nn.Sequential(
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
DeConv(in_channels = chn * 4, out_channels = chn *2, kernel_size=k_size),
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
DeConv(in_channels = chn * 2, out_channels = chn, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.InstanceNorm2d(chn),
nn.LeakyReLU()
# DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
# nn.InstanceNorm2d(chn),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
self.upsample = LIIF(chn, 3, mlp_hidden_list)
self.upsample.gen_coord((batch_size, \
chn,image_size//2,image_size//2),(image_size,image_size))
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
x2 = self.encoder(input)
x2 = self.resblocks(x2)
out = self.decoder(x2)
out = self.upsample(out)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/FastNST_Liif_warp.py
+150
View File
@@ -0,0 +1,150 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: FastNST_Liif.py
# Created Date: Thursday October 14th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 4:33:51 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
from components.Liif_conv import LIIF
class ImageLN(nn.Module):
def __init__(self, dim) -> None:
super().__init__()
self.layer = nn.LayerNorm(dim)
def forward(self, x):
y = self.layer(x.permute(0,2,3,1)).permute(0,3,1,2)
return y
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
class_num = kwargs["n_class"]
window_size = kwargs["window_size"]
image_size = kwargs["image_size"]
batch_size = kwargs["batch_size"]
# mlp_in_dim = kwargs["mlp_in_dim"]
# mlp_out_dim = kwargs["mlp_out_dim"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
embed_dim = 96
window_size = 8
num_heads = 8
mlp_ratio = 2.
norm_layer = nn.LayerNorm
qk_scale = None
qkv_bias = True
self.patch_norm = True
self.lnnorm = norm_layer(embed_dim)
self.encoder = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*2, out_channels = chn*4, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn*4 , out_channels = chn * 4, kernel_size=k_size, stride=2, padding=1,bias =False),
nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
nn.LeakyReLU(),
)
for _ in range(res_num):
self.resblock_list += [ResBlock(chn * 4,k_size),]
self.resblocks = nn.Sequential(*self.resblock_list)
# self.encoder2 = nn.Sequential(
# nn.Conv2d(in_channels = chn*4 , out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size, stride=2, padding=1,bias =False),
# ImageLN(chn * 8),
# nn.LeakyReLU()
# )
self.decoder = nn.Sequential(
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
# DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# nn.LeakyReLU(),
DeConv(in_channels = chn * 4, out_channels = chn *2, kernel_size=k_size),
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
# DeConv(in_channels = chn * 2, out_channels = chn, kernel_size=k_size),
# # nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
# nn.InstanceNorm2d(chn, affine=True, momentum=0),
# nn.LeakyReLU()
# DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
# nn.InstanceNorm2d(chn),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
self.upsample1 = LIIF(chn*2, chn)
self.upsample1.gen_coord((batch_size, \
chn,image_size//4,image_size//4),(image_size//2,image_size//2))
self.upsample2 = LIIF(chn, chn)
self.upsample2.gen_coord((batch_size, \
chn,image_size//2,image_size//2),(image_size,image_size))
self.out_conv = nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
x2 = self.encoder(input)
x2 = self.resblocks(x2)
out = self.decoder(x2)
out = self.upsample1(out)
out = self.upsample2(out)
out = self.out_conv(out)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/FastNST_Liif_warpinvo.py
+146
View File
@@ -0,0 +1,146 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: FastNST_Liif.py
# Created Date: Thursday October 14th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 8:47:28 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
from torch import nn
from torch.nn import init
from torch.nn import functional as F
from components.ResBlock import ResBlock
from components.DeConv import DeConv
from components.Liif_invo import LIIF
class Generator(nn.Module):
def __init__(
self,
**kwargs
):
super().__init__()
chn = kwargs["g_conv_dim"]
k_size = kwargs["g_kernel_size"]
res_num = kwargs["res_num"]
class_num = kwargs["n_class"]
window_size = kwargs["window_size"]
image_size = kwargs["image_size"]
batch_size = kwargs["batch_size"]
# mlp_in_dim = kwargs["mlp_in_dim"]
# mlp_out_dim = kwargs["mlp_out_dim"]
padding_size = int((k_size -1)/2)
self.resblock_list = []
embed_dim = 96
norm_layer = nn.LayerNorm
self.img_token = nn.Sequential(
nn.Conv2d(in_channels = 3 , out_channels = chn , kernel_size=k_size, stride=1, padding=1, bias= False),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn , out_channels = chn*2, kernel_size=k_size, stride=2, padding=1,bias =False), #
nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn*2, out_channels = chn*4, kernel_size=k_size, stride=2, padding=1,bias =False),
# nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
# nn.LeakyReLU(),
# nn.Conv2d(in_channels = chn*4 , out_channels = chn * 4, kernel_size=k_size, stride=2, padding=1,bias =False),
# nn.InstanceNorm2d(chn * 4, affine=True, momentum=0),
# nn.LeakyReLU(),
)
image_size = image_size // 2
self.downsample1 = LIIF(chn * 2, chn * 4)
self.downsample1.gen_coord((batch_size, \
chn,image_size,image_size),(image_size//2,image_size//2))
image_size = image_size // 2
self.downsample2 = LIIF(chn * 4, chn * 4)
self.downsample2.gen_coord((batch_size, \
chn,image_size,image_size),(image_size//2,image_size//2))
for _ in range(res_num):
self.resblock_list += [ResBlock(chn * 4,k_size),]
self.resblocks = nn.Sequential(*self.resblock_list)
# self.decoder = nn.Sequential(
# # DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# # nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# # nn.LeakyReLU(),
# # DeConv(in_channels = chn * 8, out_channels = chn * 8, kernel_size=k_size),
# # nn.InstanceNorm2d(chn * 8, affine=True, momentum=0),
# # nn.LeakyReLU(),
# DeConv(in_channels = chn * 4, out_channels = chn *2, kernel_size=k_size),
# nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
# nn.LeakyReLU(),
# # DeConv(in_channels = chn * 2, out_channels = chn, kernel_size=k_size),
# # # nn.InstanceNorm2d(chn * 2, affine=True, momentum=0),
# # nn.InstanceNorm2d(chn, affine=True, momentum=0),
# # nn.LeakyReLU()
# # DeConv(in_channels = chn *2, out_channels = chn, kernel_size=k_size),
# # nn.InstanceNorm2d(chn),
# # nn.LeakyReLU(),
# # nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
# )
image_size = image_size // 2
self.upsample1 = LIIF(chn*4, chn * 4)
self.upsample1.gen_coord((batch_size, \
chn,image_size,image_size),(image_size*2,image_size*2))
image_size = image_size * 2
self.upsample2 = LIIF(chn*4, chn * 2)
self.upsample2.gen_coord((batch_size, \
chn,image_size,image_size),(image_size*2,image_size*2))
# image_size = image_size * 2
# self.upsample2 = LIIF(chn, chn)
# self.upsample2.gen_coord((batch_size, \
# chn,image_size,image_size),(image_size*2,image_size*2))
self.decoder = nn.Sequential(
DeConv(in_channels = chn * 2, out_channels = chn, kernel_size=k_size),
nn.InstanceNorm2d(chn, affine=True, momentum=0),
nn.LeakyReLU(),
nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
)
# self.out_conv = nn.Conv2d(in_channels = chn, out_channels =3, kernel_size=k_size, stride=1, padding=1,bias =True)
# self.__weights_init__()
# def __weights_init__(self):
# for layer in self.encoder:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
# for layer in self.encoder2:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
def forward(self, input):
out = self.img_token(input)
out = self.downsample1(out)
out = self.downsample2(out)
out = self.resblocks(out)
out = self.upsample1(out)
out = self.upsample2(out)
out = self.decoder(out)
return out
if __name__ == '__main__':
upscale = 4
window_size = 8
height = 1024
width = 1024
model = Generator()
print(model)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/Involution.py
+303
View File
@@ -0,0 +1,303 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Involution.py
# Created Date: Tuesday July 20th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 20th July 2021 10:35:52 am
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
from torch.autograd import Function
import torch
from torch.nn.modules.utils import _pair
import torch.nn.functional as F
import torch.nn as nn
from mmcv.cnn import ConvModule
from collections import namedtuple
import cupy
from string import Template
Stream = namedtuple('Stream', ['ptr'])
def Dtype(t):
if isinstance(t, torch.cuda.FloatTensor):
return 'float'
elif isinstance(t, torch.cuda.DoubleTensor):
return 'double'
@cupy._util.memoize(for_each_device=True)
def load_kernel(kernel_name, code, **kwargs):
code = Template(code).substitute(**kwargs)
kernel_code = cupy.cuda.compile_with_cache(code)
return kernel_code.get_function(kernel_name)
CUDA_NUM_THREADS = 1024
kernel_loop = '''
#define CUDA_KERNEL_LOOP(i, n) \
for (int i = blockIdx.x * blockDim.x + threadIdx.x; \
i < (n); \
i += blockDim.x * gridDim.x)
'''
def GET_BLOCKS(N):
return (N + CUDA_NUM_THREADS - 1) // CUDA_NUM_THREADS
_involution_kernel = kernel_loop + '''
extern "C"
__global__ void involution_forward_kernel(
const ${Dtype}* bottom_data, const ${Dtype}* weight_data, ${Dtype}* top_data) {
CUDA_KERNEL_LOOP(index, ${nthreads}) {
const int n = index / ${channels} / ${top_height} / ${top_width};
const int c = (index / ${top_height} / ${top_width}) % ${channels};
const int h = (index / ${top_width}) % ${top_height};
const int w = index % ${top_width};
const int g = c / (${channels} / ${groups});
${Dtype} value = 0;
#pragma unroll
for (int kh = 0; kh < ${kernel_h}; ++kh) {
#pragma unroll
for (int kw = 0; kw < ${kernel_w}; ++kw) {
const int h_in = -${pad_h} + h * ${stride_h} + kh * ${dilation_h};
const int w_in = -${pad_w} + w * ${stride_w} + kw * ${dilation_w};
if ((h_in >= 0) && (h_in < ${bottom_height})
&& (w_in >= 0) && (w_in < ${bottom_width})) {
const int offset = ((n * ${channels} + c) * ${bottom_height} + h_in)
* ${bottom_width} + w_in;
const int offset_weight = ((((n * ${groups} + g) * ${kernel_h} + kh) * ${kernel_w} + kw) * ${top_height} + h)
* ${top_width} + w;
value += weight_data[offset_weight] * bottom_data[offset];
}
}
}
top_data[index] = value;
}
}
'''
_involution_kernel_backward_grad_input = kernel_loop + '''
extern "C"
__global__ void involution_backward_grad_input_kernel(
const ${Dtype}* const top_diff, const ${Dtype}* const weight_data, ${Dtype}* const bottom_diff) {
CUDA_KERNEL_LOOP(index, ${nthreads}) {
const int n = index / ${channels} / ${bottom_height} / ${bottom_width};
const int c = (index / ${bottom_height} / ${bottom_width}) % ${channels};
const int h = (index / ${bottom_width}) % ${bottom_height};
const int w = index % ${bottom_width};
const int g = c / (${channels} / ${groups});
${Dtype} value = 0;
#pragma unroll
for (int kh = 0; kh < ${kernel_h}; ++kh) {
#pragma unroll
for (int kw = 0; kw < ${kernel_w}; ++kw) {
const int h_out_s = h + ${pad_h} - kh * ${dilation_h};
const int w_out_s = w + ${pad_w} - kw * ${dilation_w};
if (((h_out_s % ${stride_h}) == 0) && ((w_out_s % ${stride_w}) == 0)) {
const int h_out = h_out_s / ${stride_h};
const int w_out = w_out_s / ${stride_w};
if ((h_out >= 0) && (h_out < ${top_height})
&& (w_out >= 0) && (w_out < ${top_width})) {
const int offset = ((n * ${channels} + c) * ${top_height} + h_out)
* ${top_width} + w_out;
const int offset_weight = ((((n * ${groups} + g) * ${kernel_h} + kh) * ${kernel_w} + kw) * ${top_height} + h_out)
* ${top_width} + w_out;
value += weight_data[offset_weight] * top_diff[offset];
}
}
}
}
bottom_diff[index] = value;
}
}
'''
_involution_kernel_backward_grad_weight = kernel_loop + '''
extern "C"
__global__ void involution_backward_grad_weight_kernel(
const ${Dtype}* const top_diff, const ${Dtype}* const bottom_data, ${Dtype}* const buffer_data) {
CUDA_KERNEL_LOOP(index, ${nthreads}) {
const int h = (index / ${top_width}) % ${top_height};
const int w = index % ${top_width};
const int kh = (index / ${kernel_w} / ${top_height} / ${top_width})
% ${kernel_h};
const int kw = (index / ${top_height} / ${top_width}) % ${kernel_w};
const int h_in = -${pad_h} + h * ${stride_h} + kh * ${dilation_h};
const int w_in = -${pad_w} + w * ${stride_w} + kw * ${dilation_w};
if ((h_in >= 0) && (h_in < ${bottom_height})
&& (w_in >= 0) && (w_in < ${bottom_width})) {
const int g = (index / ${kernel_h} / ${kernel_w} / ${top_height} / ${top_width}) % ${groups};
const int n = (index / ${groups} / ${kernel_h} / ${kernel_w} / ${top_height} / ${top_width}) % ${num};
${Dtype} value = 0;
#pragma unroll
for (int c = g * (${channels} / ${groups}); c < (g + 1) * (${channels} / ${groups}); ++c) {
const int top_offset = ((n * ${channels} + c) * ${top_height} + h)
* ${top_width} + w;
const int bottom_offset = ((n * ${channels} + c) * ${bottom_height} + h_in)
* ${bottom_width} + w_in;
value += top_diff[top_offset] * bottom_data[bottom_offset];
}
buffer_data[index] = value;
} else {
buffer_data[index] = 0;
}
}
}
'''
class _involution(Function):
@staticmethod
def forward(ctx, input, weight, stride, padding, dilation):
assert input.dim() == 4 and input.is_cuda
assert weight.dim() == 6 and weight.is_cuda
batch_size, channels, height, width = input.size()
kernel_h, kernel_w = weight.size()[2:4]
output_h = int((height + 2 * padding[0] - (dilation[0] * (kernel_h - 1) + 1)) / stride[0] + 1)
output_w = int((width + 2 * padding[1] - (dilation[1] * (kernel_w - 1) + 1)) / stride[1] + 1)
output = input.new(batch_size, channels, output_h, output_w)
n = output.numel()
with torch.cuda.device_of(input):
f = load_kernel('involution_forward_kernel', _involution_kernel, Dtype=Dtype(input), nthreads=n,
num=batch_size, channels=channels, groups=weight.size()[1],
bottom_height=height, bottom_width=width,
top_height=output_h, top_width=output_w,
kernel_h=kernel_h, kernel_w=kernel_w,
stride_h=stride[0], stride_w=stride[1],
dilation_h=dilation[0], dilation_w=dilation[1],
pad_h=padding[0], pad_w=padding[1])
f(block=(CUDA_NUM_THREADS,1,1),
grid=(GET_BLOCKS(n),1,1),
args=[input.data_ptr(), weight.data_ptr(), output.data_ptr()],
stream=Stream(ptr=torch.cuda.current_stream().cuda_stream))
ctx.save_for_backward(input, weight)
ctx.stride, ctx.padding, ctx.dilation = stride, padding, dilation
return output
@staticmethod
def backward(ctx, grad_output):
assert grad_output.is_cuda and grad_output.is_contiguous()
input, weight = ctx.saved_tensors
stride, padding, dilation = ctx.stride, ctx.padding, ctx.dilation
batch_size, channels, height, width = input.size()
kernel_h, kernel_w = weight.size()[2:4]
output_h, output_w = grad_output.size()[2:]
grad_input, grad_weight = None, None
opt = dict(Dtype=Dtype(grad_output),
num=batch_size, channels=channels, groups=weight.size()[1],
bottom_height=height, bottom_width=width,
top_height=output_h, top_width=output_w,
kernel_h=kernel_h, kernel_w=kernel_w,
stride_h=stride[0], stride_w=stride[1],
dilation_h=dilation[0], dilation_w=dilation[1],
pad_h=padding[0], pad_w=padding[1])
with torch.cuda.device_of(input):
if ctx.needs_input_grad[0]:
grad_input = input.new(input.size())
n = grad_input.numel()
opt['nthreads'] = n
f = load_kernel('involution_backward_grad_input_kernel',
_involution_kernel_backward_grad_input, **opt)
f(block=(CUDA_NUM_THREADS,1,1),
grid=(GET_BLOCKS(n),1,1),
args=[grad_output.data_ptr(), weight.data_ptr(), grad_input.data_ptr()],
stream=Stream(ptr=torch.cuda.current_stream().cuda_stream))
if ctx.needs_input_grad[1]:
grad_weight = weight.new(weight.size())
n = grad_weight.numel()
opt['nthreads'] = n
f = load_kernel('involution_backward_grad_weight_kernel',
_involution_kernel_backward_grad_weight, **opt)
f(block=(CUDA_NUM_THREADS,1,1),
grid=(GET_BLOCKS(n),1,1),
args=[grad_output.data_ptr(), input.data_ptr(), grad_weight.data_ptr()],
stream=Stream(ptr=torch.cuda.current_stream().cuda_stream))
return grad_input, grad_weight, None, None, None
def _involution_cuda(input, weight, bias=None, stride=1, padding=0, dilation=1):
""" involution kernel
"""
assert input.size(0) == weight.size(0)
assert input.size(-2)//stride == weight.size(-2)
assert input.size(-1)//stride == weight.size(-1)
if input.is_cuda:
out = _involution.apply(input, weight, _pair(stride), _pair(padding), _pair(dilation))
if bias is not None:
out += bias.view(1,-1,1,1)
else:
raise NotImplementedError
return out
class involution(nn.Module):
def __init__(self,
channels,
kernel_size,
stride):
super(involution, self).__init__()
self.kernel_size = kernel_size
self.stride = stride
self.channels = channels
reduction_ratio = 4
self.group_channels = 8
self.groups = self.channels // self.group_channels
self.seblock = nn.Sequential(
nn.Conv2d(in_channels = channels, out_channels = channels // reduction_ratio, kernel_size= 1),
nn.InstanceNorm2d(channels // reduction_ratio, affine=True, momentum=0),
nn.ReLU(),
nn.Conv2d(in_channels = channels // reduction_ratio, out_channels = kernel_size**2 * self.groups, kernel_size= 1)
)
# self.conv1 = ConvModule(
# in_channels=channels,
# out_channels=channels // reduction_ratio,
# kernel_size=1,
# conv_cfg=None,
# norm_cfg=dict(type='BN'),
# act_cfg=dict(type='ReLU'))
# self.conv2 = ConvModule(
# in_channels=channels // reduction_ratio,
# out_channels=kernel_size**2 * self.groups,
# kernel_size=1,
# stride=1,
# conv_cfg=None,
# norm_cfg=None,
# act_cfg=None)
if stride > 1:
self.avgpool = nn.AvgPool2d(stride, stride)
def forward(self, x):
# weight = self.conv2(self.conv1(x if self.stride == 1 else self.avgpool(x)))
weight = self.seblock(x)
b, c, h, w = weight.shape
weight = weight.view(b, self.groups, self.kernel_size, self.kernel_size, h, w)
out = _involution_cuda(x, weight, stride=self.stride, padding=(self.kernel_size-1)//2)
return out
components/Liif.py
+146
View File
@@ -0,0 +1,146 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Liif.py
# Created Date: Monday October 18th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 10:27:09 am
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
def make_coord(shape, ranges=None, flatten=True):
""" Make coordinates at grid centers.
"""
coord_seqs = []
for i, n in enumerate(shape):
print("i: %d, n: %d"%(i,n))
if ranges is None:
v0, v1 = -1, 1
else:
v0, v1 = ranges[i]
r = (v1 - v0) / (2 * n)
seq = v0 + r + (2 * r) * torch.arange(n).float()
coord_seqs.append(seq)
ret = torch.stack(torch.meshgrid(*coord_seqs), dim=-1)
if flatten:
ret = ret.view(-1, ret.shape[-1])
return ret
class MLP(nn.Module):
def __init__(self, in_dim, out_dim, hidden_list):
super().__init__()
layers = []
lastv = in_dim
for hidden in hidden_list:
layers.append(nn.Linear(lastv, hidden))
layers.append(nn.ReLU())
lastv = hidden
layers.append(nn.Linear(lastv, out_dim))
self.layers = nn.Sequential(*layers)
def forward(self, x):
shape = x.shape[:-1]
x = self.layers(x.view(-1, x.shape[-1]))
return x.view(*shape, -1)
class LIIF(nn.Module):
def __init__(self, mlp_in_dim, mlp_out_dim, mlp_hidden_list):
super().__init__()
imnet_in_dim = mlp_in_dim
imnet_in_dim *= 9
imnet_in_dim += 2 # attach coord
imnet_in_dim += 2
self.imnet = MLP(imnet_in_dim, mlp_out_dim, mlp_hidden_list).cuda()
def gen_coord(self, in_shape, output_size):
self.vx_lst = [-1, 1]
self.vy_lst = [-1, 1]
eps_shift = 1e-6
self.image_size=output_size
# field radius (global: [-1, 1])
rx = 2 / in_shape[-2] / 2
ry = 2 / in_shape[-1] / 2
coord = make_coord(output_size,flatten=False) \
.expand(in_shape[0],output_size[0],output_size[1],2) \
.view(in_shape[0],output_size[0]*output_size[1],2)
cell = torch.ones_like(coord)
cell[:, :, 0] *= 2 / coord.shape[-2]
cell[:, :, 1] *= 2 / coord.shape[-1]
feat_coord = make_coord(in_shape[-2:], flatten=False) \
.permute(2, 0, 1) \
.unsqueeze(0).expand(in_shape[0], 2, *in_shape[-2:])
areas = []
self.rel_coord = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
self.rel_cell = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
self.coord_ = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
for vx in self.vx_lst:
for vy in self.vy_lst:
self.coord_[(vx+1)//2,(vy+1)//2,:, :, :] = coord.clone()
self.coord_[(vx+1)//2,(vy+1)//2,:, :, 0] += vx * rx + eps_shift
self.coord_[(vx+1)//2,(vy+1)//2,:, :, 1] += vy * ry + eps_shift
self.coord_.clamp_(-1 + 1e-6, 1 - 1e-6)
q_coord = F.grid_sample(
feat_coord, self.coord_[(vx+1)//2,(vy+1)//2,:, :, :].flip(-1).unsqueeze(1),
mode='nearest', align_corners=False)[:, :, 0, :] \
.permute(0, 2, 1)
self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, :] = coord - q_coord
self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, :] = cell.clone()
self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
area = torch.abs(self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] * self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1])
areas.append(area + 1e-9)
tot_area = torch.stack(areas).sum(dim=0)
t = areas[0]; areas[0] = areas[3]; areas[3] = t
t = areas[1]; areas[1] = areas[2]; areas[2] = t
self.area_weights = []
for item in areas:
self.area_weights.append((item / tot_area).unsqueeze(-1).cuda())
self.rel_coord = self.rel_coord.cuda()
self.rel_cell = self.rel_cell.cuda()
self.coord_ = self.coord_.cuda()
def forward(self, feat):
# B K*K*Cin H W
feat = F.unfold(feat, 3, padding=1).view(
feat.shape[0], feat.shape[1] * 9, feat.shape[2], feat.shape[3])
preds = []
for vx in [0,1]:
for vy in [0,1]:
q_feat = F.grid_sample(
feat, self.coord_[vx,vy,:,:,:].flip(-1).unsqueeze(1),
mode='nearest', align_corners=False)[:, :, 0, :] \
.permute(0, 2, 1)
inp = torch.cat([q_feat, self.rel_coord[vx,vy,:,:,:], self.rel_cell[vx,vy,:,:,:]], dim=-1)
bs, q = self.coord_[0,0,:,:,:].shape[:2]
pred = self.imnet(inp.view(bs * q, -1)).view(bs, q, -1)
# print("pred shape: ",pred.shape)
preds.append(pred)
ret = 0
for pred, area in zip(preds, self.area_weights):
ret = ret + pred * area
return ret.permute(0, 2, 1).view(-1,3,self.image_size[0],self.image_size[1])
components/Liif_conv.py
+156
View File
@@ -0,0 +1,156 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Liif.py
# Created Date: Monday October 18th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 4:26:26 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
def make_coord(shape, ranges=None, flatten=True):
""" Make coordinates at grid centers.
"""
coord_seqs = []
for i, n in enumerate(shape):
print("i: %d, n: %d"%(i,n))
if ranges is None:
v0, v1 = -1, 1
else:
v0, v1 = ranges[i]
r = (v1 - v0) / (2 * n)
seq = v0 + r + (2 * r) * torch.arange(n).float()
coord_seqs.append(seq)
ret = torch.stack(torch.meshgrid(*coord_seqs), dim=-1)
if flatten:
ret = ret.view(-1, ret.shape[-1])
return ret
class MLP(nn.Module):
def __init__(self, in_dim, out_dim, hidden_list):
super().__init__()
layers = []
lastv = in_dim
for hidden in hidden_list:
layers.append(nn.Linear(lastv, hidden))
layers.append(nn.ReLU())
lastv = hidden
layers.append(nn.Linear(lastv, out_dim))
self.layers = nn.Sequential(*layers)
def forward(self, x):
shape = x.shape[:-1]
x = self.layers(x.view(-1, x.shape[-1]))
return x.view(*shape, -1)
class LIIF(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
imnet_in_dim = in_dim
# imnet_in_dim += 2 # attach coord
# imnet_in_dim += 2
self.imnet = nn.Sequential( \
nn.Conv2d(in_channels = imnet_in_dim, out_channels = out_dim, kernel_size= 3,padding=1),
nn.InstanceNorm2d(out_dim, affine=True, momentum=0),
nn.LeakyReLU(),
# nn.Conv2d(in_channels = out_dim, out_channels = out_dim, kernel_size= 3,padding=1),
# nn.InstanceNorm2d(out_dim),
# nn.LeakyReLU(),
)
def gen_coord(self, in_shape, output_size):
self.vx_lst = [-1, 1]
self.vy_lst = [-1, 1]
eps_shift = 1e-6
self.image_size=output_size
# field radius (global: [-1, 1])
rx = 2 / in_shape[-2] / 2
ry = 2 / in_shape[-1] / 2
self.coord = make_coord(output_size,flatten=False) \
.expand(in_shape[0],output_size[0],output_size[1],2)
# cell = torch.ones_like(coord)
# cell[:, :, 0] *= 2 / coord.shape[-2]
# cell[:, :, 1] *= 2 / coord.shape[-1]
# feat_coord = make_coord(in_shape[-2:], flatten=False) \
# .permute(2, 0, 1) \
# .unsqueeze(0).expand(in_shape[0], 2, *in_shape[-2:])
# areas = []
# self.rel_coord = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# self.rel_cell = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# self.coord_ = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# for vx in self.vx_lst:
# for vy in self.vy_lst:
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, :] = coord.clone()
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, 0] += vx * rx + eps_shift
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, 1] += vy * ry + eps_shift
# self.coord_.clamp_(-1 + 1e-6, 1 - 1e-6)
# q_coord = F.grid_sample(
# feat_coord, self.coord_[(vx+1)//2,(vy+1)//2,:, :, :].flip(-1).unsqueeze(1),
# mode='nearest', align_corners=False)[:, :, 0, :] \
# .permute(0, 2, 1)
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, :] = coord - q_coord
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, :] = cell.clone()
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
# area = torch.abs(self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] * self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1])
# areas.append(area + 1e-9)
# tot_area = torch.stack(areas).sum(dim=0)
# t = areas[0]; areas[0] = areas[3]; areas[3] = t
# t = areas[1]; areas[1] = areas[2]; areas[2] = t
# self.area_weights = []
# for item in areas:
# self.area_weights.append((item / tot_area).unsqueeze(-1).cuda())
# self.rel_coord = self.rel_coord.cuda()
# self.rel_cell = self.rel_cell.cuda()
# self.coord_ = self.coord_.cuda()
self.coord = self.coord.cuda()
def forward(self, feat):
# B K*K*Cin H W
# feat = F.unfold(feat, 3, padding=1).view(
# feat.shape[0], feat.shape[1] * 9, feat.shape[2], feat.shape[3])
# preds = []
# for vx in [0,1]:
# for vy in [0,1]:
# print("feat shape: ", feat.shape)
# print("coor shape: ", self.coord.shape)
q_feat = F.grid_sample(
feat, self.coord,
mode='bilinear', align_corners=False)
out = self.imnet(q_feat)
# inp = torch.cat([q_feat, self.rel_coord[vx,vy,:,:,:], self.rel_cell[vx,vy,:,:,:]], dim=-1)
# bs, q = self.coord_[0,0,:,:,:].shape[:2]
# pred = self.imnet(inp.view(bs * q, -1)).view(bs, q, -1)
# # print("pred shape: ",pred.shape)
# preds.append(pred)
# ret = 0
# for pred, area in zip(preds, self.area_weights):
# ret = ret + pred * area
# print("warp output shape: ",out.shape)
return out
components/Liif_invo.py
+164
View File
@@ -0,0 +1,164 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: Liif.py
# Created Date: Monday October 18th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 8:25:18 pm
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
from components.Involution import involution
def make_coord(shape, ranges=None, flatten=True):
""" Make coordinates at grid centers.
"""
coord_seqs = []
for i, n in enumerate(shape):
print("i: %d, n: %d"%(i,n))
if ranges is None:
v0, v1 = -1, 1
else:
v0, v1 = ranges[i]
r = (v1 - v0) / (2 * n)
seq = v0 + r + (2 * r) * torch.arange(n).float()
coord_seqs.append(seq)
ret = torch.stack(torch.meshgrid(*coord_seqs), dim=-1)
if flatten:
ret = ret.view(-1, ret.shape[-1])
return ret
class MLP(nn.Module):
def __init__(self, in_dim, out_dim, hidden_list):
super().__init__()
layers = []
lastv = in_dim
for hidden in hidden_list:
layers.append(nn.Linear(lastv, hidden))
layers.append(nn.ReLU())
lastv = hidden
layers.append(nn.Linear(lastv, out_dim))
self.layers = nn.Sequential(*layers)
def forward(self, x):
shape = x.shape[:-1]
x = self.layers(x.view(-1, x.shape[-1]))
return x.view(*shape, -1)
class LIIF(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
imnet_in_dim = in_dim
# imnet_in_dim += 2 # attach coord
# imnet_in_dim += 2
self.conv1x1 = nn.Conv2d(in_channels = imnet_in_dim, out_channels = out_dim, kernel_size= 1)
# self.same_padding = nn.ReflectionPad2d(padding_size)
# self.conv = involution(out_dim,5,1)
self.imnet = nn.Sequential( \
# nn.Conv2d(in_channels = imnet_in_dim, out_channels = out_dim, kernel_size= 3,padding=1),
involution(out_dim,5,1),
nn.InstanceNorm2d(out_dim, affine=True, momentum=0),
nn.LeakyReLU(),
# nn.Conv2d(in_channels = out_dim, out_channels = out_dim, kernel_size= 3,padding=1),
# nn.InstanceNorm2d(out_dim),
# nn.LeakyReLU(),
)
def gen_coord(self, in_shape, output_size):
self.vx_lst = [-1, 1]
self.vy_lst = [-1, 1]
eps_shift = 1e-6
self.image_size=output_size
# field radius (global: [-1, 1])
rx = 2 / in_shape[-2] / 2
ry = 2 / in_shape[-1] / 2
self.coord = make_coord(output_size,flatten=False) \
.expand(in_shape[0],output_size[0],output_size[1],2)
# cell = torch.ones_like(coord)
# cell[:, :, 0] *= 2 / coord.shape[-2]
# cell[:, :, 1] *= 2 / coord.shape[-1]
# feat_coord = make_coord(in_shape[-2:], flatten=False) \
# .permute(2, 0, 1) \
# .unsqueeze(0).expand(in_shape[0], 2, *in_shape[-2:])
# areas = []
# self.rel_coord = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# self.rel_cell = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# self.coord_ = torch.zeros((2,2,in_shape[0],output_size[0]*output_size[1],2))
# for vx in self.vx_lst:
# for vy in self.vy_lst:
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, :] = coord.clone()
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, 0] += vx * rx + eps_shift
# self.coord_[(vx+1)//2,(vy+1)//2,:, :, 1] += vy * ry + eps_shift
# self.coord_.clamp_(-1 + 1e-6, 1 - 1e-6)
# q_coord = F.grid_sample(
# feat_coord, self.coord_[(vx+1)//2,(vy+1)//2,:, :, :].flip(-1).unsqueeze(1),
# mode='nearest', align_corners=False)[:, :, 0, :] \
# .permute(0, 2, 1)
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, :] = coord - q_coord
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
# self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, :] = cell.clone()
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 0] *= in_shape[-2]
# self.rel_cell[(vx+1)//2,(vy+1)//2,:, :, 1] *= in_shape[-1]
# area = torch.abs(self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 0] * self.rel_coord[(vx+1)//2,(vy+1)//2,:, :, 1])
# areas.append(area + 1e-9)
# tot_area = torch.stack(areas).sum(dim=0)
# t = areas[0]; areas[0] = areas[3]; areas[3] = t
# t = areas[1]; areas[1] = areas[2]; areas[2] = t
# self.area_weights = []
# for item in areas:
# self.area_weights.append((item / tot_area).unsqueeze(-1).cuda())
# self.rel_coord = self.rel_coord.cuda()
# self.rel_cell = self.rel_cell.cuda()
# self.coord_ = self.coord_.cuda()
self.coord = self.coord.cuda()
def forward(self, feat):
# B K*K*Cin H W
# feat = F.unfold(feat, 3, padding=1).view(
# feat.shape[0], feat.shape[1] * 9, feat.shape[2], feat.shape[3])
# preds = []
# for vx in [0,1]:
# for vy in [0,1]:
# print("feat shape: ", feat.shape)
# print("coor shape: ", self.coord.shape)
q_feat = self.conv1x1(feat)
q_feat = F.grid_sample(
q_feat, self.coord,
mode='bilinear', align_corners=False)
out = self.imnet(q_feat)
# inp = torch.cat([q_feat, self.rel_coord[vx,vy,:,:,:], self.rel_cell[vx,vy,:,:,:]], dim=-1)
# bs, q = self.coord_[0,0,:,:,:].shape[:2]
# pred = self.imnet(inp.view(bs * q, -1)).view(bs, q, -1)
# # print("pred shape: ",pred.shape)
# preds.append(pred)
# ret = 0
# for pred, area in zip(preds, self.area_weights):
# ret = ret + pred * area
# print("warp output shape: ",out.shape)
return out
components/ResBlock.py
+38
View File
@@ -0,0 +1,38 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: ResBlock.py
# Created Date: Monday July 5th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Monday, 5th July 2021 12:18:18 am
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
from torch import nn
class ResBlock(nn.Module):
def __init__(self, in_channel, k_size = 3, stride=1):
super().__init__()
padding_size = int((k_size -1)/2)
self.block = nn.Sequential(
nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = in_channel , out_channels = in_channel , kernel_size= k_size, stride=stride, bias= False),
nn.InstanceNorm2d(in_channel, affine=True, momentum=0),
nn.ReflectionPad2d(padding_size),
nn.Conv2d(in_channels = in_channel , out_channels = in_channel , kernel_size= k_size, stride=stride, bias= False),
nn.InstanceNorm2d(in_channel, affine=True, momentum=0)
)
self.__weights_init__()
def __weights_init__(self):
for m in self.modules():
if isinstance(m,nn.Conv2d):
nn.init.xavier_uniform_(m.weight)
def forward(self, input):
res = input
h = self.block(input)
out = h + res
return out
components/Transform.py
+14
View File
@@ -0,0 +1,14 @@
import torch
from torch import nn
class Transform_block(nn.Module):
def __init__(self, k_size = 10):
super().__init__()
padding_size = int((k_size -1)/2)
# self.padding = nn.ReplicationPad2d(padding_size)
self.pool = nn.AvgPool2d(k_size, stride=1,padding=padding_size)
def forward(self, input_image):
# h = self.padding(input)
out = self.pool(input_image)
return out
components/network_swin.py
+854
View File
@@ -0,0 +1,854 @@
# -----------------------------------------------------------------------------------
# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
# Originally Written by Ze Liu, Modified by Jingyun Liang.
# -----------------------------------------------------------------------------------
import math
import torch
import torch.nn as nn
import torch.utils.checkpoint as checkpoint
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
class Mlp(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
def window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
return windows
def window_reverse(windows, window_size, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class WindowAttention(nn.Module):
r""" Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter(
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
nn.init
trunc_normal_(self.relative_position_bias_table, std=.02)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
"""
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
B_, N, C = x.shape
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
q = q * self.scale
attn = (q @ k.transpose(-2, -1))
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(-1, self.num_heads, N, N)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
def extra_repr(self) -> str:
return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class SwinTransformerBlock(nn.Module):
r""" Swin Transformer Block.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
num_heads (int): Number of attention heads.
window_size (int): Window size.
shift_size (int): Shift size for SW-MSA.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float, optional): Stochastic depth rate. Default: 0.0
act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
self.norm1 = norm_layer(dim)
self.attn = WindowAttention(
dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = nn.Dropout(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
if self.shift_size > 0:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
h_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
w_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
cnt = 0
for h in h_slices:
for w in w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
return attn_mask
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
# assert L == H * W, "input feature has wrong size"
shortcut = x
x = self.norm1(x)
x = x.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
# partition windows
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
else:
attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
else:
x = shifted_x
x = x.view(B, H * W, C)
x = shortcut + self.drop_path(x)
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
def flops(self):
flops = 0
H, W = self.input_resolution
# norm1
flops += self.dim * H * W
# W-MSA/SW-MSA
nW = H * W / self.window_size / self.window_size
flops += nW * self.attn.flops(self.window_size * self.window_size)
# mlp
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# norm2
flops += self.dim * H * W
return flops
class PatchMerging(nn.Module):
r""" Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self,
input_resolution,
dim, norm_layer = nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
x = self.norm(x)
x = self.reduction(x)
return x
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
class BasicLayer(nn.Module):
""" A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local window size.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList([
SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
num_heads=num_heads, window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
norm_layer=norm_layer)
for i in range(depth)])
# patch merging layer
if downsample is not None:
self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
else:
self.downsample = None
def forward(self, x, x_size):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
if self.downsample is not None:
x = self.downsample(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
def flops(self):
flops = 0
for blk in self.blocks:
flops += blk.flops()
if self.downsample is not None:
flops += self.downsample.flops()
return flops
class RSTB(nn.Module):
"""Residual Swin Transformer Block (RSTB).
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local window size.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
img_size: Input image size.
patch_size: Patch size.
resi_connection: The convolutional block before residual connection.
"""
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
img_size=224, patch_size=4, resi_connection='1conv'):
super(RSTB, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = BasicLayer(dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint)
if resi_connection == '1conv':
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1))
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
def forward(self, x, x_size):
return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
def flops(self):
flops = 0
flops += self.residual_group.flops()
H, W = self.input_resolution
flops += H * W * self.dim * self.dim * 9
flops += self.patch_embed.flops()
flops += self.patch_unembed.flops()
return flops
class PatchEmbed(nn.Module):
r""" Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
flops = 0
H, W = self.img_size
if self.norm is not None:
flops += H * W * self.embed_dim
return flops
class PatchUnEmbed(nn.Module):
r""" Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
class Upsample(nn.Sequential):
"""Upsample module.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat):
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
super(Upsample, self).__init__(*m)
class UpsampleOneStep(nn.Sequential):
"""UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
Used in lightweight SR to save parameters.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
self.num_feat = num_feat
self.input_resolution = input_resolution
m = []
m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
super(UpsampleOneStep, self).__init__(*m)
def flops(self):
H, W = self.input_resolution
flops = H * W * self.num_feat * 3 * 9
return flops
class SwinIR(nn.Module):
r""" SwinIR
A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
Args:
img_size (int | tuple(int)): Input image size. Default 64
patch_size (int | tuple(int)): Patch size. Default: 1
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
window_size (int): Window size. Default: 7
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
patch_norm (bool): If True, add normalization after patch embedding. Default: True
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range: Image range. 1. or 255.
upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(self, img_size=64, patch_size=1, in_chans=3,
embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
**kwargs):
super(SwinIR, self).__init__()
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.img_range = img_range
if in_chans == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
self.upscale = upscale
self.upsampler = upsampler
#####################################################################################################
################################### 1, shallow feature extraction ###################################
self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
#####################################################################################################
################################### 2, deep feature extraction ######################################
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
trunc_normal_(self.absolute_pos_embed, std=.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
# build Residual Swin Transformer blocks (RSTB)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(dim=embed_dim,
input_resolution=(patches_resolution[0],
patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection
)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
# build the last conv layer in deep feature extraction
if resi_connection == '1conv':
self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
#####################################################################################################
################################ 3, high quality image reconstruction ################################
if self.upsampler == 'pixelshuffle':
# for classical SR
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.upsample = Upsample(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
(patches_resolution[0], patches_resolution[1]))
elif self.upsampler == 'nearest+conv':
# for real-world SR (less artifacts)
assert self.upscale == 4, 'only support x4 now.'
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'absolute_pos_embed'}
@torch.jit.ignore
def no_weight_decay_keywords(self):
return {'relative_position_bias_table'}
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == 'pixelshuffle':
# for classical SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
elif self.upsampler == 'nearest+conv':
# for real-world SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
x = x + self.conv_last(res)
x = x / self.img_range + self.mean
return x
def flops(self):
flops = 0
H, W = self.patches_resolution
flops += H * W * 3 * self.embed_dim * 9
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops()
flops += H * W * 3 * self.embed_dim * self.embed_dim
flops += self.upsample.flops()
return flops
if __name__ == '__main__':
upscale = 4
window_size = 8
height = (1024 // upscale // window_size + 1) * window_size
width = (720 // upscale // window_size + 1) * window_size
model = SwinIR(upscale=2, img_size=(height, width),
window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect')
print(model)
print(height, width, model.flops() / 1e9)
x = torch.randn((1, 3, height, width))
x = model(x)
print(x.shape)
components/warp_invo.py
+45
View File
@@ -0,0 +1,45 @@
#!/usr/bin/env python3
# -*- coding:utf-8 -*-
#############################################################
# File: warp_invo.py
# Created Date: Tuesday October 19th 2021
# Author: Chen Xuanhong
# Email: chenxuanhongzju@outlook.com
# Last Modified: Tuesday, 19th October 2021 11:27:13 am
# Modified By: Chen Xuanhong
# Copyright (c) 2021 Shanghai Jiao Tong University
#############################################################
from torch import nn
from components.Involution import involution
class DeConv(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size = 3, upsampl_scale = 2, padding="reflect"):
super().__init__()
self.upsampling = nn.UpsamplingNearest2d(scale_factor=upsampl_scale)
padding_size = int((kernel_size -1)/2)
self.conv1x1 = nn.Conv2d(in_channels = in_channels, out_channels = out_channels, kernel_size= 1)
# self.same_padding = nn.ReflectionPad2d(padding_size)
if padding.lower() == "reflect":
self.conv = involution(out_channels,5,1)
# self.conv = nn.Sequential(
# nn.ReflectionPad2d(padding_size),
# nn.Conv2d(in_channels = in_channels, out_channels = out_channels, kernel_size= kernel_size, bias= False))
# for layer in self.conv:
# if isinstance(layer,nn.Conv2d):
# nn.init.xavier_uniform_(layer.weight)
elif padding.lower() == "zero":
self.conv = involution(out_channels,5,1)
# nn.init.xavier_uniform_(self.conv.weight)
# self.__weights_init__()
# def __weights_init__(self):
# nn.init.xavier_uniform_(self.conv.weight)
def forward(self, input):
h = self.conv1x1(input)
h = self.upsampling(h)
h = self.conv(h)
return h