BasicSR Versions Save

🚀

✨ Highlight

A minor version for quick bug fixed. #396 #397

🚀

:sparkles: Highlights

We reorganize the BasicSR codes. It may be incompatible with the previous v1.2.0

• Add registry mechanism
• Support pip install
• Support JIT CUDA ops
• Can be easily used as an external package to develop your own project (example project is coming soon)
• Update format: change max line length to 120
• Add degradations (data utils)

This major change may introduce bugs. If you encounter bugs, please let me know. Thanks!

🚀

:sparkles: Highlights

• Add ESRGAN and DFDNet colab
• Add FID and LPIPS metrics
• Add matlab imresize bicubic (#317)
• README add datasets download links (#318)

:bug: Bug Fixes

• PSNR and SSIM calculation on uint8 type
• Fix metrics bug in video_base_model.py (#314)

:palm_tree: Improvements

• Reorganize code structure and remove unnessary packages
• tensor2img support gray images
• Refactor DFDNet codes

🚀

:sparkles: Highlights

• Add Baidu Drive (百度网盘) download links
• Add funny emoji :relaxed:

:bug: Bug Fixes

• bgr2rgb type conversion in stylegan2 model
• Supporting training w/o validation

:palm_tree: Improvements

• download_pretrained_models.py script supports downloading all the models
• Refactor define_network functions

Hope all is well 🚀

Highlights

• Add DFDNet inference codes (ECCV20: Blind Face Restoration via Deep Multi-scale Component Dictionaries)
• Add more official StyleGAN2 pretrained models: Model Zoo
• Add New Feature section in README.

Bug Fixes

• PyTorch 1.6 uses a new serialization for torch.save. The saved model cannot be loaded by the previous PyTorch version. We updated the publish_models.py with _use_new_zipfile_serialization=False. More details.

Hope all is well 🚀

Highlights

• Add StyleGAN2 training and testing codes. Pretrained models are here.
• Fix bug: cuda prefetcher return none twice.
• Add HOWTOs for quick starts.

We will use releases to manage BasicSR :smile:

Hope all is well 🚀

This is a brand-new version of BasicSR. We have re-organized all the codes and frameworks.

Highlights

• We use Dynamic Instantiation for creating datasets, architectures, and models. So it is easier and more friendly to develop your own algorithms.
• We provide richer documents. At the same time, we also provide a Chinese version (同时也提供了中文版本的文档说明).
• We provide more pre-trained models, training examples. We also upload the training process and curves to wandb.
• Currently, it supports:
  • Training: EDSR, EDVR, ESRGAN, SRResNet, SRGAN
  • Testing: DUF, EDSR, EDVR, ESRGAN, RCAN, SRResNt, SRGAN, and TOF.
• We also mirror this codebase to Gitee码云 for easy access of Chinese users.

Sorry that this version of BasicSR is not compatible with the previous versions.

We will add more features to this codebase. And welcome contribute, and report bugs! 😆