Flashlight application libraries are domain-specific libraries build on top of the flashlight core and flashlight lib. They provide lightweight, unopinionated pipelines and tools that are easily modifiable for training or inference across tasks. Below are supported applications; new applications are under active development.

The asr application provides tools for audio processing/augmentation, acoustic model training, beam search decoding, and preprocessing/preparing audio data for use. Full documentation for usage and binaries can be found here.

• Binaries:
  • fl_asr_train
  • fl_asr_test
  • fl_asr_decode
• Tutorial/Utility Binaries
  • fl_asr_tutorial_inference_ctc
  • fl_asr_tutorial_finetune_ctc

The lm application provides tools for text preprocessing and language model training for both auto-regressive and BERT-style models. Full documentation for usage and binaries can be found here.

• Binaries:
  • fl_lm_dictionary_builder
  • fl_lm_train

The imgclass application is still in early, active development. It currently provides dataset abstractions for ImageNet and an example training pipeline for Resnet34 which can be easily extended to more complex setups.

• Binaries:
  • fl_img_imagenet_resnet34