res
This commit is contained in:
XHChen0528
@@ -113,6 +113,8 @@ class ResnetBlock_Adain(nn.Module):
|
||||
out = x + y
|
||||
return out
|
||||
|
||||
|
||||
|
||||
class Generator(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -0,0 +1,373 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding:utf-8 -*-
|
||||
#############################################################
|
||||
# File: Generator_Invobn_config1.py
|
||||
# Created Date: Saturday February 26th 2022
|
||||
# Author: Chen Xuanhong
|
||||
# Email: chenxuanhongzju@outlook.com
|
||||
# Last Modified: Sunday, 27th February 2022 7:50:18 pm
|
||||
# Modified By: Chen Xuanhong
|
||||
# Copyright (c) 2022 Shanghai Jiao Tong University
|
||||
#############################################################
|
||||
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from components.LSTU import LSTU
|
||||
|
||||
# from components.DeConv_Invo import DeConv
|
||||
class InstanceNorm(nn.Module):
|
||||
def __init__(self, epsilon=1e-8):
|
||||
"""
|
||||
@notice: avoid in-place ops.
|
||||
https://discuss.pytorch.org/t/encounter-the-runtimeerror-one-of-the-variables-needed-for-gradient-computation-has-been-modified-by-an-inplace-operation/836/3
|
||||
"""
|
||||
super(InstanceNorm, self).__init__()
|
||||
self.epsilon = epsilon
|
||||
|
||||
def forward(self, x):
|
||||
x = x - torch.mean(x, (2, 3), True)
|
||||
tmp = torch.mul(x, x) # or x ** 2
|
||||
tmp = torch.rsqrt(torch.mean(tmp, (2, 3), True) + self.epsilon)
|
||||
return x * tmp
|
||||
|
||||
class ApplyStyle(nn.Module):
|
||||
"""
|
||||
@ref: https://github.com/lernapparat/lernapparat/blob/master/style_gan/pytorch_style_gan.ipynb
|
||||
"""
|
||||
def __init__(self, latent_size, channels):
|
||||
super(ApplyStyle, self).__init__()
|
||||
self.linear = nn.Linear(latent_size, channels * 2)
|
||||
|
||||
def forward(self, x, latent):
|
||||
style = self.linear(latent) # style => [batch_size, n_channels*2]
|
||||
shape = [-1, 2, x.size(1), 1, 1]
|
||||
style = style.view(shape) # [batch_size, 2, n_channels, ...]
|
||||
#x = x * (style[:, 0] + 1.) + style[:, 1]
|
||||
x = x * (style[:, 0] * 1 + 1.) + style[:, 1] * 1
|
||||
return x
|
||||
|
||||
class ResnetBlock_Adain(nn.Module):
|
||||
def __init__(self,
|
||||
dim,
|
||||
latent_size,
|
||||
padding_type,
|
||||
activation=nn.ReLU(True),
|
||||
res_mode="depthwise"):
|
||||
super(ResnetBlock_Adain, self).__init__()
|
||||
|
||||
p = 0
|
||||
conv1 = []
|
||||
if padding_type == 'reflect':
|
||||
conv1 += [nn.ReflectionPad2d(1)]
|
||||
elif padding_type == 'replicate':
|
||||
conv1 += [nn.ReplicationPad2d(1)]
|
||||
elif padding_type == 'zero':
|
||||
p = 1
|
||||
else:
|
||||
raise NotImplementedError('padding [%s] is not implemented' % padding_type)
|
||||
if res_mode.lower() == "conv":
|
||||
conv1 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p), InstanceNorm()]
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv1 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p,groups=dim, bias=False),
|
||||
nn.Conv2d(dim, dim, kernel_size=1),
|
||||
InstanceNorm()]
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv1 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p,groups=dim, bias=False),
|
||||
nn.Conv2d(dim, dim, kernel_size=1),
|
||||
InstanceNorm()]
|
||||
self.conv1 = nn.Sequential(*conv1)
|
||||
self.style1 = ApplyStyle(latent_size, dim)
|
||||
self.act1 = activation
|
||||
|
||||
p = 0
|
||||
conv2 = []
|
||||
if padding_type == 'reflect':
|
||||
conv2 += [nn.ReflectionPad2d(1)]
|
||||
elif padding_type == 'replicate':
|
||||
conv2 += [nn.ReplicationPad2d(1)]
|
||||
elif padding_type == 'zero':
|
||||
p = 1
|
||||
else:
|
||||
raise NotImplementedError('padding [%s] is not implemented' % padding_type)
|
||||
if res_mode.lower() == "conv":
|
||||
conv2 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p), InstanceNorm()]
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv2 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p,groups=dim, bias=False),
|
||||
nn.Conv2d(dim, dim, kernel_size=1),
|
||||
InstanceNorm()]
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv2 += [nn.Conv2d(dim, dim, kernel_size=3, padding=p,groups=dim, bias=False),
|
||||
nn.Conv2d(dim, dim, kernel_size=1),
|
||||
InstanceNorm()]
|
||||
self.conv2 = nn.Sequential(*conv2)
|
||||
self.style2 = ApplyStyle(latent_size, dim)
|
||||
|
||||
|
||||
def forward(self, x, dlatents_in_slice):
|
||||
y = self.conv1(x)
|
||||
y = self.style1(y, dlatents_in_slice)
|
||||
y = self.act1(y)
|
||||
y = self.conv2(y)
|
||||
y = self.style2(y, dlatents_in_slice)
|
||||
out = x + y
|
||||
return out
|
||||
|
||||
class ResUpSampleBlock(nn.Module):
|
||||
def __init__(self,
|
||||
in_dim,
|
||||
out_dim,
|
||||
latent_size,
|
||||
activation=nn.LeakyReLU(0.2),
|
||||
res_mode="depthwise"):
|
||||
super(ResnetBlock_Adain, self).__init__()
|
||||
conv1 = []
|
||||
self.in1 = InstanceNorm()
|
||||
self.in2 = InstanceNorm()
|
||||
if res_mode.lower() == "conv":
|
||||
|
||||
conv1 += [activation,
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv1 += [activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=in_dim, bias=False),
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv1 += [activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=in_dim, bias=False),
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=1)]
|
||||
|
||||
self.conv1 = nn.Sequential(*conv1)
|
||||
self.style1 = ApplyStyle(latent_size, in_dim)
|
||||
|
||||
conv2 = []
|
||||
if res_mode.lower() == "conv":
|
||||
conv2 += [activation,
|
||||
nn.Conv2d(out_dim, out_dim, kernel_size=3, padding=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv2 += [activation,
|
||||
nn.Conv2d(out_dim, out_dim, kernel_size=3, padding=1,groups=out_dim, bias=False),
|
||||
nn.Conv2d(out_dim, out_dim, kernel_size=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv2 += [activation,
|
||||
nn.Conv2d(out_dim, out_dim, kernel_size=3, padding=1,groups=out_dim, bias=False),
|
||||
nn.Conv2d(out_dim, out_dim, kernel_size=1)]
|
||||
self.conv2 = nn.Sequential(*conv2)
|
||||
self.style2 = ApplyStyle(latent_size, out_dim)
|
||||
|
||||
self.reshape1_1 = nn.Conv2d(in_dim, out_dim, kernel_size=1)
|
||||
self.resampling = nn.UpsamplingBilinear2d(scale_factor=2)
|
||||
|
||||
|
||||
def forward(self, x, dlatents_in_slice):
|
||||
y = self.in1(x)
|
||||
y = self.style1(y, dlatents_in_slice)
|
||||
y = self.conv1(y)
|
||||
y = self.resampling(y)
|
||||
y = self.in2(y)
|
||||
y = self.style2(y, dlatents_in_slice)
|
||||
y = self.conv2(y)
|
||||
res = self.reshape1_1(x)
|
||||
res = self.resampling(res)
|
||||
out = res + y
|
||||
return out
|
||||
|
||||
|
||||
class ResDownSampleBlock(nn.Module):
|
||||
def __init__(self,
|
||||
in_dim,
|
||||
out_dim,
|
||||
activation=nn.LeakyReLU(0.2),
|
||||
res_mode="depthwise"):
|
||||
super(ResnetBlock_Adain, self).__init__()
|
||||
conv1 = []
|
||||
if res_mode.lower() == "conv":
|
||||
|
||||
conv1 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv1 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=in_dim, bias=False),
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv1 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=in_dim, bias=False),
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=1)]
|
||||
|
||||
self.conv1 = nn.Sequential(*conv1)
|
||||
|
||||
conv2 = []
|
||||
if res_mode.lower() == "conv":
|
||||
conv2 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise":
|
||||
conv2 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=out_dim, bias=False),
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=1)]
|
||||
|
||||
elif res_mode.lower() == "depthwise_eca":
|
||||
conv2 += [
|
||||
nn.BatchNorm2d(in_dim),
|
||||
activation,
|
||||
nn.Conv2d(in_dim, in_dim, kernel_size=3, padding=1,groups=out_dim, bias=False),
|
||||
nn.Conv2d(in_dim, out_dim, kernel_size=1)]
|
||||
self.conv2 = nn.Sequential(*conv2)
|
||||
|
||||
self.reshape1_1 = nn.Conv2d(in_dim, out_dim, kernel_size=1)
|
||||
self.resampling = nn.AvgPool2d(3,2,1)
|
||||
|
||||
|
||||
def forward(self, x):
|
||||
y = self.conv1(y)
|
||||
y = self.resampling(y)
|
||||
y = self.conv2(y)
|
||||
res = self.reshape1_1(x)
|
||||
res = self.resampling(res)
|
||||
out = res + y
|
||||
return out
|
||||
|
||||
|
||||
class Generator(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
**kwargs
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
id_dim = kwargs["id_dim"]
|
||||
k_size = kwargs["g_kernel_size"]
|
||||
res_num = kwargs["res_num"]
|
||||
in_channel = kwargs["in_channel"]
|
||||
up_mode = kwargs["up_mode"]
|
||||
|
||||
aggregator = kwargs["aggregator"]
|
||||
res_mode = kwargs["res_mode"]
|
||||
|
||||
padding_size= int((k_size -1)/2)
|
||||
padding_type= 'reflect'
|
||||
|
||||
activation = nn.LeakyReLU(0.2)
|
||||
from components.DeConv_Depthwise import DeConv
|
||||
|
||||
# self.first_layer = nn.Sequential(nn.ReflectionPad2d(3), nn.Conv2d(3, 64, kernel_size=7, padding=0, bias=False),
|
||||
# nn.BatchNorm2d(64), activation)
|
||||
self.first_layer = nn.Sequential(nn.ReflectionPad2d(1),
|
||||
nn.Conv2d(3, in_channel, kernel_size=3, stride=2, padding=0, bias=False),
|
||||
nn.BatchNorm2d(in_channel),
|
||||
activation) # 256
|
||||
# self.first_layer = nn.Sequential(nn.Conv2d(3, 64, kernel_size=3, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(64), activation)
|
||||
### downsample
|
||||
self.down1 = ResDownSampleBlock(in_channel, in_channel*2,res_mode=res_mode)
|
||||
# nn.Sequential(
|
||||
# nn.Conv2d(in_channel, in_channel*2, stride=2, kernel_size=3, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*2),
|
||||
# activation) # 128
|
||||
|
||||
self.down2 = ResDownSampleBlock(in_channel*2, in_channel*4,res_mode=res_mode)
|
||||
# nn.Sequential(
|
||||
# nn.Conv2d(in_channel*2, in_channel*4, stride=2, kernel_size=3, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*4),
|
||||
# activation) # 64
|
||||
|
||||
# self.lstu = LSTU(in_channel*4,in_channel*4,in_channel*8,4)
|
||||
|
||||
self.down3 = ResDownSampleBlock(in_channel*4, in_channel*8,res_mode=res_mode)
|
||||
# nn.Sequential(
|
||||
# nn.Conv2d(in_channel*4, in_channel*8, stride=2, kernel_size=3, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*8),
|
||||
# activation) # 32
|
||||
|
||||
# self.down4 = nn.Sequential(
|
||||
# nn.Conv2d(in_channel*8, in_channel*8, stride=2, kernel_size=3, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*8),
|
||||
# activation)
|
||||
|
||||
|
||||
|
||||
### resnet blocks
|
||||
# BN = []
|
||||
# for i in range(res_num):
|
||||
# BN += [
|
||||
# ResnetBlock_Adain(in_channel*8, latent_size=id_dim,
|
||||
# padding_type=padding_type, activation=activation, res_mode=res_mode)]
|
||||
# self.BottleNeck = nn.Sequential(*BN)
|
||||
|
||||
self.up4 = ResDownSampleBlock(in_channel*8,in_channel*8,id_dim,res_mode=res_mode) # 64
|
||||
# nn.Sequential(
|
||||
# nn.Upsample(scale_factor=2, mode='bilinear'),
|
||||
# nn.Conv2d(in_channel*8, in_channel*8, kernel_size=3, stride=1, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*8),
|
||||
# activation
|
||||
# )
|
||||
|
||||
self.up3 = ResDownSampleBlock(in_channel*8,in_channel*4,id_dim,res_mode=res_mode) # 128
|
||||
# nn.Sequential(
|
||||
# nn.Upsample(scale_factor=2, mode='bilinear'),
|
||||
# nn.Conv2d(in_channel*8, in_channel*4, kernel_size=3, stride=1, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*4),
|
||||
# activation
|
||||
# )
|
||||
|
||||
self.up2 = ResDownSampleBlock(in_channel*4,in_channel*2,id_dim,res_mode=res_mode) # 256
|
||||
# nn.Sequential(
|
||||
# nn.Upsample(scale_factor=2, mode='bilinear'),
|
||||
# nn.Conv2d(in_channel*4, in_channel*2, kernel_size=3, stride=1, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel*2),
|
||||
# activation
|
||||
# )
|
||||
|
||||
self.up1 = ResDownSampleBlock(in_channel*2,in_channel,id_dim,res_mode=res_mode) # 512
|
||||
# nn.Sequential(
|
||||
# nn.Upsample(scale_factor=2, mode='bilinear'),
|
||||
# nn.Conv2d(in_channel*2, in_channel, kernel_size=3, stride=1, padding=1, bias=False),
|
||||
# nn.BatchNorm2d(in_channel),
|
||||
# activation
|
||||
# )
|
||||
# self.last_layer = nn.Sequential(nn.Conv2d(64, 3, kernel_size=3, padding=1))
|
||||
self.last_layer = nn.Sequential(nn.ReflectionPad2d(1),
|
||||
nn.Conv2d(in_channel, 3, kernel_size=3, padding=0))
|
||||
# self.last_layer = nn.Sequential(nn.ReflectionPad2d(3),
|
||||
# nn.Conv2d(64, 3, kernel_size=7, padding=0))
|
||||
|
||||
|
||||
# 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, img, id):
|
||||
res = self.first_layer(img)
|
||||
res = self.down1(res)
|
||||
res = self.down2(res)
|
||||
res = self.down3(res)
|
||||
res = self.up4(res,id)
|
||||
res = self.up3(res,id)
|
||||
res = self.up2(res,id) # + skip
|
||||
res = self.up1(res,id)
|
||||
res = self.last_layer(res)
|
||||
|
||||
return res
|
||||
+3
-3
@@ -31,7 +31,7 @@ def getParameters():
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
# general settings
|
||||
parser.add_argument('-v', '--version', type=str, default='cycle_lstu1',
|
||||
parser.add_argument('-v', '--version', type=str, default='cycle_res1',
|
||||
help="version name for train, test, finetune")
|
||||
parser.add_argument('-t', '--tag', type=str, default='cycle',
|
||||
help="tag for current experiment")
|
||||
@@ -46,9 +46,9 @@ def getParameters():
|
||||
|
||||
# training
|
||||
parser.add_argument('--experiment_description', type=str,
|
||||
default="cycle配合LSTU")
|
||||
default="cycle配合残差decoder,ID注入放在decoder中")
|
||||
|
||||
parser.add_argument('--train_yaml', type=str, default="train_cycleloss.yaml")
|
||||
parser.add_argument('--train_yaml', type=str, default="train_cycleloss_res.yaml")
|
||||
|
||||
# system logger
|
||||
parser.add_argument('--logger', type=str,
|
||||
|
||||
@@ -0,0 +1,70 @@
|
||||
# Related scripts
|
||||
train_script_name: multi_gpu_cycle
|
||||
|
||||
# models' scripts
|
||||
model_configs:
|
||||
g_model:
|
||||
script: Generator_Res_config
|
||||
class_name: Generator
|
||||
module_params:
|
||||
id_dim: 512
|
||||
g_kernel_size: 3
|
||||
in_channel: 64
|
||||
res_num: 9
|
||||
up_mode: bilinear
|
||||
aggregator: "conv"
|
||||
res_mode: "conv"
|
||||
|
||||
d_model:
|
||||
script: projected_discriminator
|
||||
class_name: ProjectedDiscriminator
|
||||
module_params:
|
||||
diffaug: False
|
||||
interp224: False
|
||||
backbone_kwargs: {}
|
||||
|
||||
# arcface_ckpt: arcface_torch/checkpoints/glint360k_cosface_r100_fp16_backbone.pth
|
||||
arcface_ckpt: arcface_ckpt/arcface_checkpoint.tar
|
||||
|
||||
# Training information
|
||||
batch_size: 20
|
||||
|
||||
# Dataset
|
||||
dataloader: VGGFace2HQ_multigpu
|
||||
dataset_name: vggface2_hq
|
||||
dataset_params:
|
||||
random_seed: 1234
|
||||
dataloader_workers: 6
|
||||
|
||||
eval_dataloader: DIV2K_hdf5
|
||||
eval_dataset_name: DF2K_H5_Eval
|
||||
eval_batch_size: 2
|
||||
|
||||
# Dataset
|
||||
|
||||
# Optimizer
|
||||
optim_type: Adam
|
||||
g_optim_config:
|
||||
lr: 0.0006
|
||||
betas: [ 0, 0.99]
|
||||
eps: !!float 1e-8
|
||||
|
||||
d_optim_config:
|
||||
lr: 0.0006
|
||||
betas: [ 0, 0.99]
|
||||
eps: !!float 1e-8
|
||||
|
||||
id_weight: 25.0
|
||||
reconstruct_weight: 10.0
|
||||
rec_feature_match_weight: 10.0
|
||||
cycle_feature_match_weight: 10.0
|
||||
cycle_weight: 10.0
|
||||
|
||||
# Log
|
||||
log_step: 400
|
||||
model_save_step: 10000
|
||||
total_step: 1000000
|
||||
sample_step: 1000
|
||||
checkpoint_names:
|
||||
generator_name: Generator
|
||||
discriminator_name: Discriminator
|
||||
Reference in New Issue
Block a user