代码地址：https://github.com/sczhou/CodeFormer/releases/tag/v0.1.0

论文的一些简略介绍：
https://qq742971636.blog.csdn.net/article/details/134562550

BasicSR介绍

CodeFormer整个项目都沿袭BasicSR，了解一下BasicSR很有必要：

https://mp.csdn.net/mp_blog/creation/success/135674803

环境

# git clone this repository
git clone https://github.com/sczhou/CodeFormer
cd CodeFormer

# create new anaconda env
conda create -n codeformer python=3.8 -y
conda activate codeformer

conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia

# install python dependencies
pip3 install -r requirements.txt
python basicsr/setup.py develop

conda install -c conda-forge dlib (only for face detection or cropping with dlib)

数据

找一些高清人脸数据1024*1024。

人脸数据需要对齐，对齐方式为： https://qq742971636.blog.csdn.net/article/details/135521146

阶段 I - VQGAN

训练VQGAN:

python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 basicsr/train.py -opt options/VQGAN_512_ds32_nearest_stage1.yml --launcher pytorch

CUDA_VISIBLE_DEVICES=0,2,3 python -m torch.distributed.launch --nproc_per_node=3 --master_port=4321 basicsr/train.py -opt options/VQGAN_512_ds32_nearest_stage1.yml --launcher pytorch # 指定三张显卡训练，对应VQGAN_512_ds32_nearest_stage1.yaml也是需要修改的

训练完VQGAN后，可以通过下面代码预先获得训练数据集的密码本序列，从而加速后面阶段的训练过程:

python scripts/generate_latent_gt.py

如果你不需要训练自己的VQGAN，可以在Release v0.1.0文档中找到预训练的VQGAN (vqgan_code1024.pth)和对应的密码本序列 (latent_gt_code1024.pth): https://github.com/sczhou/CodeFormer/releases/tag/v0.1.0

打开日志查看训练过程：

tensorboard --logdir="/ssd/xiedong/CodeFormer/tb_logger/20240116_182107_VQGAN-512-ds32-nearest-stage1" --bind_all

VQGAN本身就是一个图生图的网络，在中间使用transformer将特征图转为embedding. 而 CodeFormer就是要利用这每张图的embedding来进行面部修复。

下面代码里用vqgan_code1024.pth获取训练数据的密码本，vqgan_code1024.pth的encoder输出的是2563232的特征图，由embedding给到1*1024，最终所有图保存为一个pytorch文件。

import argparse
import glob
import numpy as np
import os
import cv2
import torch
from torchvision.transforms.functional import normalize
from tqdm import tqdm

from basicsr.utils import imwrite, img2tensor, tensor2img

from basicsr.utils.registry import ARCH_REGISTRY

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('-i', '--test_path', type=str, default='/ssd/xiedong/FFHQ/faces_hq_sr')
    parser.add_argument('-o', '--save_root', type=str, default='/ssd/xiedong/FFHQ/lt_output')
    parser.add_argument('--codebook_size', type=int, default=1024)
    parser.add_argument('--ckpt_path', type=str, default='/ssd/xiedong/CodeFormer/weights/vqgan/vqgan_code1024.pth')
    args = parser.parse_args()

    if args.save_root.endswith('/'):  # solve when path ends with /
        args.save_root = args.save_root[:-1]
    dir_name = os.path.abspath(args.save_root)
    os.makedirs(dir_name, exist_ok=True)

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    test_path = args.test_path
    save_root = args.save_root
    ckpt_path = args.ckpt_path
    codebook_size = args.codebook_size

    vqgan = ARCH_REGISTRY.get('VQAutoEncoder')(512, 64, [1, 2, 2, 4, 4, 8], 'nearest',
                                               codebook_size=codebook_size).to(device)
    checkpoint = torch.load(ckpt_path)['params_ema']

    vqgan.load_state_dict(checkpoint)
    vqgan.eval()

    sum_latent = np.zeros((codebook_size)).astype('float64')
    size_latent = 32
    latent = {}
    latent['orig'] = {}
    latent['hflip'] = {}
    for i in ['orig', 'hflip']:
        # for i in ['hflip']:
        for img_path in tqdm(sorted(glob.glob(os.path.join(test_path, '*.[jp][pn]g')))):
            img_name = os.path.basename(img_path)
            img = cv2.imread(img_path)
            if i == 'hflip':
                cv2.flip(img, 1, img)
            img = img2tensor(img / 255., bgr2rgb=True, float32=True)
            normalize(img, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
            img = img.unsqueeze(0).to(device)
            with torch.no_grad():
                # output = net(img)[0]
                # x, feat_dict = vqgan.encoder(img, True)
                x = vqgan.encoder(img)
                x, _, log = vqgan.quantize(x)
            # del output
            torch.cuda.empty_cache()

            min_encoding_indices = log['min_encoding_indices']
            min_encoding_indices = min_encoding_indices.view(size_latent, size_latent)
            latent[i][img_name[:-4]] = min_encoding_indices.cpu().numpy()
            print(img_name, latent[i][img_name[:-4]].shape)

    latent_save_path = os.path.join(save_root, f'latent_gt_code{codebook_size}.pth')
    torch.save(latent, latent_save_path)
    print(f'\nLatent GT code are saved in {save_root}')

阶段 II - CodeFormer (w=0)

w=0 是需要模型完全追求抽象美学，w=1 是需要模型完全追求与原图相似。

在第一个阶段，得到了每张图对应的embedding。

训练密码本训练预测模块:

python -m torch.distributed.launch --nproc_per_node=8 --master_port=4322 basicsr/train.py -opt options/CodeFormer_stage2.yml --launcher pytorch

预训练CodeFormer第二阶段模型 (codeformer_stage2.pth)可以在Releases v0.1.0文档里下载: https://github.com/sczhou/CodeFormer/releases/tag/v0.1.0

阶段 III - CodeFormer (w=1)

训练可调模块:

python -m torch.distributed.launch --nproc_per_node=8 --master_port=4323 basicsr/train.py -opt options/CodeFormer_stage3.yml --launcher pytorch

预训练CodeFormer模型 (codeformer.pth)可以在Releases v0.1.0文档里下载: https://github.com/sczhou/CodeFormer/releases/tag/v0.1.0