Models¶

This section describes the list of all models and trained checkpoints currently available in the library.

Usage¶

The following ISLR models are supported currently in the library.

Network	`encoder.type`	`decoder.type`
Recurrent Neural Networks	`pose-flattener`	`rnn`
Transformer	`pose-flattener`	`bert`
Spatio-Temporal Graph Convolution Network	`st-gcn`	`fc`
Sign Language GCN	`decoupled-gcn`	`fc`

For examples on how to use the models in configs, click here.

The following trained checkpoints are available for download across all the currently supported datasets and models.

Dataset	Metadata	BiLSTM	BERT	SL-GCN	ST-GCN
`AUTSLDataset`	autsl_metadata.zip	autsl_lstm.zip	autsl_bert.zip	autsl_slgcn.zip	autsl_stgcn.zip
`CSLDataset`	csl_metadata.zip	csl_lstm.zip	csl_bert.zip	csl_slgcn.zip	csl_stgcn.zip
`DeviSignDataset`	devisign_metadata.zip	devisign_lstm.zip	devisign_bert.zip	devisign_slgcn.zip	devisign_stgcn.zip
`GSLDataset`	gsl_metadata.zip	gsl_lstm.zip	gsl_bert.zip	gsl_slgcn.zip	gsl_stgcn.zip
`INCLUDEDataset`	include_metadata.zip	include_lstm.zip	include_bert.zip	include_slgcn.zip	include_stgcn.zip
`LSA64Dataset`	lsa64_metadata.zip	lsa64_lstm.zip	lsa64_bert.zip	lsa64_slgcn.zip	lsa64_stgcn.zip
`WLASLDataset`	wlasl_metadata.zip	wlasl_lstm.zip	wlasl_bert.zip	wlasl_slgcn.zip	wlasl_stgcn.zip

Note:

Metadata is a smaller extract from the actual dataset, which contains the actual signs to ID mappings.
- Extracted metadata path should be mentioned in the config.
The zipped checkpoints has both the config used to train as well as the trained parameters.
- For inference using the checkpoint, check the Inference section.