Developer's Guide
Code restructuring
- model architecture: The core code is located in model.py and can modify the network layer structure
- loss function: Adjust the weights of perceptual loss and adversarial loss in train.py
Customizing the training process
- Data preparation: Collection of paired data (with watermarked original map + non-watermarked true value map)
- Parameter Configuration: Modify the hyperparameters in config.py
- priming training::
python train.py --dataset 数据集路径 --batch_size 8
Practical advice
- Get better GPU resources with Google Colab Pro!
- Some of the underlying network weights can be frozen for small sample training
- Recommended use of TensorBoard to monitor the training process
Extended Directions
May try:
- Integration of Stable Diffusion's repair capabilities
- Development of GUI interfaces to lower the barrier to use
- Adapting the PyTorch framework to improve development efficiency
This answer comes from the articleWatermark Removal: open source image watermark removal tool, picture watermark recovery original imageThe
May not be reproduced without permission:AI productivity tools " How to do secondary development or model training based on Watermark Removal?
