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.
Automatic Speech Recognition — asr (the wav2letter Project)
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_trainfl_asr_testfl_asr_decode
- Tutorial/Utility Binaries
fl_asr_tutorial_inference_ctcfl_asr_tutorial_finetune_ctc
Language Modeling — lm
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_builderfl_lm_train
Image Classification — imgclass
The imgclass application is still in early, active development. It currently provides dataset abstractions for ImageNet and example training pipelines for Resnet34 and ViT (vision transformer) which can be easily extended to more complex setups.
- Binaries:
fl_img_imagenet_resnet34fl_img_imagenet_vitfl_img_eval
Object Detection — objdet
The objdet application provides tools for objet detection. It is still under active development, but currently provides dataset abstractions for COCO and an example training pipeline for DETR.
- Binaries:
fl_img_coco_detr