PSLA: Improving Audio Tagging With Pretraining, Sampling, Labeling, and Aggregation

Audio tagging is an active research area and has a wide range of applications. Since the release AudioSet, great progress been made in advancing model performance, which mostly comes from development novel architectures attention modules. However, we find that appropriate training techniques are equally important for building audio models with but have not received they deserve. To fill gap, this work, present PSLA, collection can noticeably boost accuracy including ImageNet pretraining, balanced sampling, data augmentation, label enhancement, aggregation their design choices. By EfficientNet these techniques, obtain single (with 13.6M parameters) ensemble achieve mean average precision (mAP) scores 0.444 0.474 on respectively, outperforming previous best system 0.439 81M parameters. In addition, our also achieves new state-of-the-art mAP 0.567 FSD50K.