Efficient Training of Audio Transformers with Patchout

11 Oct 2021  ·  Khaled Koutini, Jan Schlüter, Hamid Eghbal-zadeh, Gerhard Widmer ·

The great success of transformer-based models in natural language processing (NLP) has led to various attempts at adapting these architectures to other domains such as vision and audio. Recent work has shown that transformers can outperform Convolutional Neural Networks (CNNs) on vision and audio tasks. However, one of the main shortcomings of transformer models, compared to the well-established CNNs, is the computational complexity. In transformers, the compute and memory complexity is known to grow quadratically with the input length. Therefore, there has been extensive work on optimizing transformers, but often at the cost of degrading predictive performance. In this work, we propose a novel method to optimize and regularize transformers on audio spectrograms. Our proposed models achieve a new state-of-the-art performance on Audioset and can be trained on a single consumer-grade GPU. Furthermore, we propose a transformer model that outperforms CNNs in terms of both performance and training speed. Source code: https://github.com/kkoutini/PaSST

PDF Abstract

Results from the Paper


 Ranked #1 on Audio Classification on FSD50K (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Audio Classification AudioSet PaSST (Ensemble) Test mAP 0.496 # 3
Audio Classification AudioSet PaSST-S (Single) Test mAP 0.471 # 11
Audio Tagging AudioSet PaSST mean average precision 0.496 # 2
Audio Classification FSD50K PaSST-S mAP 65.55 # 1
Audio Classification FSD50K PaSST-N-S mAP 64.2 # 2
Instrument Recognition OpenMIC-2018 PaSST mean average precision 0.843 # 1

Methods


No methods listed for this paper. Add relevant methods here