Version 0.4.0 is the first stable release as we have reached 100% completion state based on what we have planned out to do. The details of the changes are in the CHANGELOG. Please take a look.

Thank you and happy hacking!

Added

• Implemented reconstruction loss.
• Saving reconstructed image as file.
• Improve training speed by using PyTorch DataParallel to wrap our model.
  • PyTorch will parallelized the model and data over multiple GPUs.
• Supports training:
  • on CPU (tested with macOS Sierra)
  • on one GPU (tested with 1 Tesla K80 GPU)
  • on multiple GPU (tested with 8 GPUs)
  • with or without CUDA (tested with CUDA version 8.0.61)
  • cuDNN 5 (tested with cuDNN 5.1.3)

Changed

• More intuitive variable naming.

Fixed

• Resolve Pylint warnings and reformat code.
• Missing square in equation 4 for margin (class) loss.

Initial release. API is stable. The code runs. It's safe to use for development but not ready for general production usage.