Adapter-Incremental-Continual-Learning-AST

This repo provides the official PyTorch implementation of the paper "Adapter Incremental Continual Learning of Efficient Audio Spectrogram Transformers" which is presented at INTERSPEECH 2023. Run AICL_train.py to train the Audio Spectrogram Transformer (AST) in a Task Incremental Continual Learning setup.

AICL Architecture

Frequency-Time factorized Attention (FTA)

To visualize the masked attention (which enables FTA) refer FTA.ipynb