Were both training and test splits from original datasets (e.g., AudioSet, VggSound) used to train CLAP?

LAION-AI / audio-dataset

Audio Dataset for training CLAP and other models

640 stars 53 forks source link

Were both training and test splits from original datasets (e.g., AudioSet, VggSound) used to train CLAP? #97

Open ttgeng233 opened 1 year ago

ttgeng233 commented 1 year ago

I want to know the exact splits of AudioSet or VggSound used to train the CLAP. Because many audio-related datasets for downstream tasks were collected from these two large-scale datasets, if all their test data were seen during the pre-training stage, the evaluation results would be unconvincing.

YuchenHui22314 commented 2 months ago

While evaluating, we manually eliminate those examples already seen in the pretraining stage. For example, while testing on ESC-50, we eliminated all overlaps with freesound and audioset.