Skip to content

Pytorch Improvement

Javier Duarte requested to merge github/fork/Duchstf/pytorch_dev into master

Created by: Duchstf

As discussed in #255 (closed), this PR aims to improve Pytorch/Brevitas support for hls4ml. Many thanks to Pedro for initializing some works on this.

The Pytorch converter is now modeled after the Keras converter (i.e function to process a layer is accessed using a decorator).

However, since Pytorch's default data format is channel_first, synthesizing Pytorch CNN models with streaming is not possible yet. Hopefully this will be addressed in a later PR.

Current status:

  1. Conv1D, Conv2D layers.
  2. Batchnorm 1D & 2D layer.

Merge request reports