AudioTagging Done Right: 2nd comparison of deep learning methods for environmental sound classification

• Introduction
• Citing
• AudioSet Recipe
• Neural Models
• Contact

Introduction

This repository is the implementation of our submission to InterSpeech 2022: https://arxiv.org/abs/2203.13448 with the same title of the repository.

This implementation's is a new iteration of our previous effort: https://github.com/lijuncheng16/AudioSetDoneRight

We still maintain our previous repo for testing/probing the model. This repository is mainly for training models. Thanks to the other opensourced project AST (https://github.com/YuanGongND/ast), where we used their data loading pipeline, and DeiT adaptation. The pre-trained DeiT model is loaded from the timm libraray: https://github.com/rwightman/pytorch-image-models (timm version: 0.4.5) newer version of timm changed the API for base-384 of deit, but should still be one liner to load the equivalent or better model. Without these opensourced efforts, it would be way less efficient for researchers like us to make things happen.

Getting Started

Parameters:
label_dim : The number of classes (default:527).
fstride: The stride of patch spliting on the frequency dimension, for 16*16 patchs, fstride=16 means no overlap, fstride=10 means overlap of 6 (used in the paper). (default:10)
tstride: The stride of patch spliting on the time dimension, for 16*16 patchs, tstride=16 means no overlap, tstride=10 means overlap of 6 (used in the paper). (default:10)
input_fdim: The number of frequency bins of the input spectrogram. (default:128)
input_tdim: The number of time frames of the input spectrogram. (default:1024, i.e., 10.24s)
imagenet_pretrain: If True, use ImageNet pretrained model. (default: True, we recommend to set it as True for all tasks.)
audioset_pretrain: IfTrue, use full AudioSet And ImageNet pretrained model. Currently only support base384 model with fstride=tstride=10. (default: False, we recommend to set it as True for all tasks except AudioSet.)
model_size: The model size of AST, should be in [tiny224, small224, base224, base384] (default: base384). mean: The mean value of the features, note this has to be recomputed for different data augmentation. std: Standard deviation of the features. freqm, timem: time and frequency mask length for SpecAugmentation. For TALtrans Models embedding_size: embedding size n_conv_layers: number of conv layers before the bottleneck, default set to 10 n_pool_layers: number of pooling layers before the bottleneck, default set to 5 n_trans_layers: number of transformer block layers kernel_size: conv kernel size

Input: Tensor in shape [batch_size, temporal_frame_num, frequency_bin_num]. Note: the input spectrogram should be normalized with dataset mean and std. Output: Tensor of raw logits (i.e., without Sigmoid) in shape [batch_size, label_dim].

Audioset Recipe

Audioset is the largest general Audio dataset you can find as of 2022, you will need to prepare your data json files (i.e., train_data.json and eval_data.json) using the scripts provided here. The raw wavefiles of Audioset is not released by Google due to Copyright constraints, and you need to download them by yourself. We are considering releasing a version here just for research purpose (yet to be done). Once you have downloaded the .wav files of audioset.(balanced train, unbalanced train, and eval). You can use our prep_AudioSet.py to generate a sample json file in ~/egs/audioset/data/datafiles, please keep the label code consistent with ~/egs/audioset/data/class_labels_indices.csv. You can see the prep_audioset.ipynb for the brief logic.

Once you have the json files, you will need to generate the sampling weight file of your training data.

cd ~/egs/audioset
python gen_weight_file.py ./data/datafiles/train_data.json

Then you just need to change the tr_data and te_data in ~/egs/audioset/run.sh and then cd ~/egs/audioset ./run-{modelrecipe_name}.sh

3. In ~/src/run.py, line 33-34, you need to add the normalization stats, the input frame length, and if noise augmentation is needed for your dataset. For normalization stats, you need to compute the mean and std of your dataset (check ast/src/get_norm_stats.py) or you can try using our AudioSet normalization input_spec = (input_spec + mean) / (std * 2). See our explanation about why is it 2 sigma in our paper.

Also, please note that we use 16kHz audios for the pretrained model, so if you want to use the pretrained model, please prepare your data in 16kHz.

Models

Currently, you will see implementations of CRNN, CNNTransformer, Transformer, ViT(AST), PSLA, ResNet in this repository. We will include more models while we make updates to our project.

Citing

Please cite our paper(s) if you find this repository useful.

@article{li2022ATbaseline,
  doi = {10.48550/ARXIV.2203.13448},
  url = {https://arxiv.org/abs/2203.13448},
  author = {Li, Juncheng B and Qu, Shuhui and Huang, Po-Yao and Metze, Florian},
  title = {AudioTagging Done Right: 2nd comparison of deep learning methods for environmental sound classification},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

Contact

Feel free to shoot me an email at junchenl@cs.cmu.edu in case you have long questions. For short questions, Twitter DM @JunchengLi Or just submit an issue here.