Improving model coverage

The primary goal of this repository is to provide a diverse range of models, covering many domains. This is useful both for testing coverage of Swift for TensorFlow, as well as for documenting how to implement various machine learning models.

For benchmarking purposes, it's important to us to have full coverage of the models and datasets used as part of the MLPerf training and MLPerf inference benchmark suites.

Image classification is pretty well covered, but novel architectures could still be welcome there. Other categories of models aren't as well represented, and help would be appreciated in building out those. The following list is some we've identified, but is by no means exhaustive:

BERT and variants

• BERT on SQuAD (1.1 and 2.0)
• ALBERT on SQuAD

Text generation

• Reformer

Gradient boosted trees

Image segmentation

• U-Net (2-D)
• U-Net (3-D)

Point cloud analysis

• StarNet on the Waymo Open Dataset

Speech recognition

• Deep Speech 2 on LibriSpeech
• RNN-T

Language translation

• GNMT
• Transformer on WMT English-German

Recommendation

• Neural Collaborative Filtering on MovieLens

Sentiment analysis

• Seq-CNN on IMDB
• XLNet

Reinforcement learning

• MiniGo (training)

Object detection: (currently being worked on by teams at Google) We're actively working on support for object detection models, so please hold off to avoid duplicate work. Thanks!