The Code Repository for "HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection", in ICASSP 2022.
In this paper, we devise a model, HTS-AT, by combining a swin transformer with a token-semantic module and adapt it in to audio classification and sound event detection tasks. HTS-AT is an efficient and light-weight audio transformer with a hierarchical structure and has only 30 million parameters. It achieves new state-of-the-art (SOTA) results on AudioSet and ESC-50, and equals the SOTA on Speech Command V2. It also achieves better performance in event localization than the previous CNN-based models.
pip install -r requirements.txt
We do not include the installation of PyTorch in the requirment, since different machines require different vereions of CUDA and Toolkits. So make sure you install the PyTorch from the official guidance.
Install the 'SOX' and the 'ffmpeg', we recommend that you run this code in Linux inside the Conda environment. In that, you can install them by:
sudo apt install sox
conda install -c conda-forge ffmpeg
- config.py
change the varible "dataset_path" to your audioset address
change the variable "desed_folder" to your DESED address
change the classes_num to 527
./create_index.sh #
// remember to change the pathes in the script
// more information about this script is in https://github.com/qiuqiangkong/audioset_tagging_cnn
python main.py save_idc
// count the number of samples in each class and save the npy files
Open the jupyter notebook at esc-50/prep_esc50.ipynb and process it
Open the jupyter notebook at scv2/prep_scv2.ipynb and process it
python conver_desed.py
// will produce the npy data files
The script config.py contains all configurations you need to assign to run your code. Please read the introduction comments in the file and change your settings.
IMPORTANT NOTICE
Similar to many transformer structures, the HTS-AT needs warm-up otherwise the model will underfit in the beginning. To find a proper warm-up step or warm-up epoch, please pay attention to these two hyperparameters in the configuration file. The default settings works for the full AudioSet (2.2M data samples). If your working dataset contains different size of samples (e.g. 100K, 1M, 10M, etc.), you might need to change a proper warm-up step or epoch.
For the most important part: If you want to train/test your model on AudioSet, you need to set:
dataset_path = "your processed audioset folder"
dataset_type = "audioset"
balanced_data = True
loss_type = "clip_bce"
sample_rate = 32000
hop_size = 320
classes_num = 527
If you want to train/test your model on ESC-50, you need to set:
dataset_path = "your processed ESC-50 folder"
dataset_type = "esc-50"
loss_type = "clip_ce"
sample_rate = 32000
hop_size = 320
classes_num = 50
If you want to train/test your model on Speech Command V2, you need to set:
dataset_path = "your processed SCV2 folder"
dataset_type = "scv2"
loss_type = "clip_bce"
sample_rate = 16000
hop_size = 160
classes_num = 35
If you want to test your model on DESED, you need to set:
resume_checkpoint = "Your checkpoint on AudioSet"
heatmap_dir = "localization results output folder"
test_file = "output heatmap name"
fl_local = True
fl_dataset = "Your DESED npy file"
Notice: Our model is now supporting the single GPU.
All scripts is run by main.py:
Train: CUDA_VISIBLE_DEVICES=1,2,3,4 python main.py train
Test: CUDA_VISIBLE_DEVICES=1,2,3,4 python main.py test
Ensemble Test: CUDA_VISIBLE_DEVICES=1,2,3,4 python main.py esm_test
// See config.py for settings of ensemble testing
Weight Average: python main.py weight_average
// See config.py for settings of weight averaging
CUDA_VISIBLE_DEVICES=1,2,3,4 python main.py test
// make sure that fl_local=True in config.py
python fl_evaluate.py
// organize and gather the localization results
fl_evaluate_f1.ipynb
// Follow the notebook to produce the results
We provide the model checkpoints on three datasets (and additionally DESED dataset) in this link. Feel free to download and test it.
@inproceedings{htsat-ke2022,
author = {Ke Chen and Xingjian Du and Bilei Zhu and Zejun Ma and Taylor Berg-Kirkpatrick and Shlomo Dubnov},
title = {HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection},
booktitle = {{ICASSP} 2022}
}
Our work is based on Swin Transformer, which is a famous image classification transformer model.