参考图像彩色化网络修改流程（自用版）

一. SGA-Eliminating Gradient Conflict in Reference-based Line-Art Colorization（2022ECCV）

1. 修改config.yml
修改前

python 复制代码

  EPOCH: 40
  BATCH_SIZE: 16
  NUM_WORKER : 4
  TRAIN_DIR : 'anime' # 'anime' or 'afhq_cat' or 'afhq_dog' or afhq_wild

修改后

python 复制代码

  EPOCH: 400
  BATCH_SIZE: 8
  NUM_WORKER : 0
  TRAIN_DIR : 'nighttime' # 'anime' or 'afhq_cat' or 'afhq_dog' or afhq_wild

添加

python 复制代码

USE_TENSORBOARD : 'True'

2. 修改data_loader.py
添加

python 复制代码

 elif config['TRAINING_CONFIG']['TRAIN_DIR'] == 'nighttime':
     self.img_dir = r'F:\RefDataset\KAIST\train\refB'
     self.skt_dir = r'F:\RefDataset\KAIST\train\nightA'
     self.data_list = glob.glob(os.path.join(self.img_dir, '*.jpg'))

3. 修改model.py

python 复制代码

self.gcn3 = Gconv(in_features=channel, out_features=channel)
self.gcn4 = Gconv(in_features=channel, out_features=channel)

修改为

python 复制代码

self.gcn3 = Gconv(in_ch=channel, out_ch=channel)
self.gcn4 = Gconv(in_ch=channel, out_ch=channel)

二. SCFT-Reference-Based Sketch Image Colorization using Augmented-Self Reference and Dense Semantic Correspondence（2020CVPR）

和SGA修改一致
1. 修改config.yml
修改前

python 复制代码

  EPOCH: 40
  BATCH_SIZE: 16
  NUM_WORKER : 4
  TRAIN_DIR : 'anime' # 'anime' or 'afhq_cat' or 'afhq_dog' or afhq_wild

修改后

python 复制代码

  EPOCH: 400
  BATCH_SIZE: 8
  NUM_WORKER : 0
  TRAIN_DIR : 'nighttime' # 'anime' or 'afhq_cat' or 'afhq_dog' or afhq_wild

添加

python 复制代码

USE_TENSORBOARD : 'True'

2. 修改data_loader.py
添加

python 复制代码

 elif config['TRAINING_CONFIG']['TRAIN_DIR'] == 'nighttime':
     self.img_dir = r'F:\RefDataset\KAIST\train\refB'
     self.skt_dir = r'F:\RefDataset\KAIST\train\nightA'
     self.data_list = glob.glob(os.path.join(self.img_dir, '*.jpg'))