README.md 4.99 KB
Newer Older
mashun1's avatar
diffbir  
mashun1 committed
1
# DiffBIR
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
2

mashun1's avatar
diffbir  
mashun1 committed
3
## 论文
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
4

mashun1's avatar
diffbir  
mashun1 committed
5
**DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior**
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
6

mashun1's avatar
diffbir  
mashun1 committed
7
* https://arxiv.org/abs/2308.15070
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
8

mashun1's avatar
diffbir  
mashun1 committed
9
## 模型结构
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
10

mashun1's avatar
diffbir  
mashun1 committed
11
第一阶段模型使用8个Swin Transformer blocks(RSTB),每个RSTB中包含6个Swin Transformer Layers(STL),其中head数为6,window size为8.
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
12

mashun1's avatar
diffbir  
mashun1 committed
13
第二阶段模型基于Stable Diffusioin 2.1-base,创建了一个与Unet中encoder block与middle block相同的网络。
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
14

mashun1's avatar
diffbir  
mashun1 committed
15
![Alt text](images/image.png)
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
16

mashun1's avatar
diffbir  
mashun1 committed
17
## 算法原理
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
18

mashun1's avatar
diffbir  
mashun1 committed
19
用途:该算法为两阶段算法,可以提升图像的分辨率。
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
20

mashun1's avatar
diffbir  
mashun1 committed
21
第一阶段使用复原模块,从具有未知和复杂降质的低质量(LQ)图像中恢复清晰图像;
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
22

mashun1's avatar
diffbir  
mashun1 committed
23
第二阶段使用生成模块来重新生成丢失的信息。
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
24

mashun1's avatar
diffbir  
mashun1 committed
25
## 环境配置
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
26
27


mashun1's avatar
diffbir  
mashun1 committed
28
### Docker(方法一)
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
29

mashun1's avatar
diffbir  
mashun1 committed
30
31
32
    docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:1.13.1-centos7.6-dtk-23.04.1-py39-latest
    docker run --shm-size 10g --network=host --name=diffbir --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v 项目地址(绝对路径):/home/ -it <your IMAGE ID> bash
    pip install -r requirements.txt
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
33
34


mashun1's avatar
diffbir  
mashun1 committed
35
### Docker(方法二)
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
36

mashun1's avatar
diffbir  
mashun1 committed
37
38
39
40
    # 需要在对应的目录下
    docker build -t <IMAGE_NAME>:<TAG> .
    # <your IMAGE ID>用以上拉取的docker的镜像ID替换
    docker run -it --shm-size 10g --network=host --name=diffbir --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined <your IMAGE ID> bash
41

mashun1's avatar
diffbir  
mashun1 committed
42
43
44
### Anaconda (方法三)
1、关于本项目DCU显卡所需的特殊深度学习库可从光合开发者社区下载安装:
https://developer.hpccube.com/tool/
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
45

mashun1's avatar
diffbir  
mashun1 committed
46
47
48
49
50
51
52
    DTK驱动:dtk23.04.1
    python:python3.9
    torch:1.13.1
    torchvision:0.14.1
    torchaudio:0.13.1
    deepspeed:0.9.2
    apex:0.1
53

mashun1's avatar
diffbir  
mashun1 committed
54
Tips:以上dtk驱动、python、torch等DCU相关工具版本需要严格一一对应
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
55

mashun1's avatar
diffbir  
mashun1 committed
56
2、其它非特殊库参照requirements.txt安装
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
57

mashun1's avatar
diffbir  
mashun1 committed
58
    pip install -r requirements.txt
59

0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
60

mashun1's avatar
diffbir  
mashun1 committed
61
## 数据集
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
62

mashun1's avatar
diffbir  
mashun1 committed
63
下载地址(训练+测试集):https://www.image-net.org/ (imagenet1k)
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
64

mashun1's avatar
diffbir  
mashun1 committed
65
66
67
68
    datasets
        |- train
            |- n01440764
                |- xxx.JPEG
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
69

ziyannchen's avatar
ziyannchen committed
70

mashun1's avatar
diffbir  
mashun1 committed
71
## 训练
72

mashun1's avatar
diffbir  
mashun1 committed
73
### 阶段一
74

mashun1's avatar
diffbir  
mashun1 committed
75
76
77
78
79
80
1、数据准备:该操作用于生成训练以及验证数据路径列表

    python scripts/make_file_list.py \  
    --img_folder [hq_dir_path] \         # 包含图片的文件夹
    --val_size [validation_set_size] \   # 验证集大小
    --save_folder [save_dir_path] \      # 路径列表保存文件夹
ziyannchen's avatar
ziyannchen committed
81
    --follow_links
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
82

mashun1's avatar
diffbir  
mashun1 committed
83
84
85
2、修改配置文件

    修改 `configs/dataset`中相应的yaml配置文件
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
86

mashun1's avatar
diffbir  
mashun1 committed
87
    修改 `configs/train_swinir.yaml`配置文件
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
88

mashun1's avatar
diffbir  
mashun1 committed
89
90
91
3、训练

    python train.py --config [training_config_path]
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
92

mashun1's avatar
diffbir  
mashun1 committed
93
注意:该阶段训练得到的模型将用于第二阶段的训练。
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
94

mashun1's avatar
diffbir  
mashun1 committed
95
### 阶段二
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
96

mashun1's avatar
diffbir  
mashun1 committed
97
1、模型准备(Stable Diffusion v2.1): https://huggingface.co/stabilityai/stable-diffusion-2-1-base/resolve/main/v2-1_512-ema-pruned.ckpt
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
98
99


mashun1's avatar
diffbir  
mashun1 committed
100
101
102
103
104
105
106
107
108
109
110
111
112
2、初始化模型参数

    python scripts/make_stage2_init_weight.py \
    --cldm_config configs/model/cldm.yaml \
    --sd_weight [sd_v2.1_ckpt_path] \
    --swinir_weight [swinir_ckpt_path] \  # 第一阶段训练得到的模型
    --output [init_weight_output_path]    # 初始化模型保存地址

3、修改配置文件

    修改`configs/train_cldm.yaml`配置文件

4、训练
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
113
114
115

    python train.py --config [training_config_path]

mashun1's avatar
diffbir  
mashun1 committed
116
## 推理
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
117

mashun1's avatar
diffbir  
mashun1 committed
118
### general Image
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
119

mashun1's avatar
diffbir  
mashun1 committed
120
121
模型下载地址:
* https://huggingface.co/lxq007/DiffBIR/resolve/main/general_full_v1.ckpt
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
122

mashun1's avatar
diffbir  
mashun1 committed
123
* https://huggingface.co/lxq007/DiffBIR/resolve/main/general_swinir_v1.ckpt
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
124

mashun1's avatar
diffbir  
mashun1 committed
125
126
127
128
129
130
131
132
133
134
        python inference.py \
        --input inputs/demo/general \
        --config configs/model/cldm.yaml \
        --ckpt weights/general_full_v1.ckpt \  
        --reload_swinir --swinir_ckpt weights/general_swinir_v1.ckpt \  
        --steps 50 \
        --sr_scale 4 \
        --color_fix_type wavelet \
        --output results/demo/general \
        --device cuda [--tiled --tile_size 512 --tile_stride 256]
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
135

mashun1's avatar
diffbir  
mashun1 committed
136
注意:方括号中的参数为可选项,模型也可以替换为在训练阶段得到的
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
137

mashun1's avatar
diffbir  
mashun1 committed
138
### Face Image
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
139

mashun1's avatar
diffbir  
mashun1 committed
140
模型下载地址:
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
141

mashun1's avatar
diffbir  
mashun1 committed
142
* https://huggingface.co/lxq007/DiffBIR/resolve/main/face_full_v1.ckpt
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
143

mashun1's avatar
diffbir  
mashun1 committed
144
for aligned face inputs
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
145

mashun1's avatar
diffbir  
mashun1 committed
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
    python inference_face.py \
    --input inputs/demo/face/aligned \
    --sr_scale 1 \
    --output results/demo/face/aligned \
    --has_aligned \
    --device cuda

for unaligned face inputs

    python inference_face.py \
    --input inputs/demo/face/whole_img \
    --sr_scale 2 \
    --output results/demo/face/whole_img \
    --bg_upsampler DiffBIR \
    --device cuda

## result

恢复后的图像

![Alt text](images/samples_step-004900_e-000008_b-001203.png)

低质量图像

![Alt text](images/lq_step-004900_e-000008_b-001203.png)

### 精度



## 应用场景
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
177

mashun1's avatar
diffbir  
mashun1 committed
178
### 算法类别
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
179

mashun1's avatar
diffbir  
mashun1 committed
180
`图像超分`
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
181

mashun1's avatar
diffbir  
mashun1 committed
182
### 热点应用行业
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
183

mashun1's avatar
diffbir  
mashun1 committed
184
`媒体,科研,教育`
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
185

mashun1's avatar
diffbir  
mashun1 committed
186
## 源码仓库及问题反馈
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
187
188
189



mashun1's avatar
diffbir  
mashun1 committed
190
## 参考
0x3f3f3f3fun's avatar
0x3f3f3f3fun committed
191

mashun1's avatar
diffbir  
mashun1 committed
192
193
* https://github.com/XPixelGroup/DiffBIR
* https://github.com/XPixelGroup/DiffBIR/issues/55