The official code repo of "HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection"

Last update: Jan 01, 2023

Overview

Hierarchical Token Semantic Audio Transformer

Introduction

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.

Classification Results on AudioSet, ESC-50, and Speech Command V2 (mAP)

Localization/Detection Results on DESED dataset (F1-Score)

Getting Started

Install Requirments

pip install -r requirements.txt

Download and Processing Datasets

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

AudioSet

./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

ESC-50

Open the jupyter notebook at esc-50/prep_esc50.ipynb and process it

Speech Command V2

Open the jupyter notebook at scv2/prep_scv2.ipynb and process it

DESED Dataset

python conver_desed.py 
// will produce the npy data files

Set the Configuration File: config.py

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. 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"

Train and Evaluation

Notice: Our model is run on DDP mode and requires at least two GPU cards. If you want to use a single GPU for training and evaluation, you need to mannually change sed_model.py and main.py

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

Localization on DESED

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

Model Checkpoints:

We provide the model checkpoints on three datasets (and additionally DESED dataset) in this link. Feel free to download and test it.

Citing

@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.

The official code repo of "HTS-AT: A Hierarchical Token-Semantic Audio Transformer for Sound Classification and Detection"

Related tags

Overview

Hierarchical Token Semantic Audio Transformer

Introduction

Classification Results on AudioSet, ESC-50, and Speech Command V2 (mAP)

Localization/Detection Results on DESED dataset (F1-Score)

Getting Started

Install Requirments

Download and Processing Datasets

Set the Configuration File: config.py

Train and Evaluation

Localization on DESED

Model Checkpoints:

Citing

Owner

Knut(Ke) Chen

Hyperbolic Image Segmentation, CVPR 2022

PyTorch Kafka Dataset: A definition of a dataset to get training data from Kafka.

A rough implementation of the paper "A Steering Algorithm for Redirected Walking Using Reinforcement Learning"

On Generating Extended Summaries of Long Documents

Latex code for making neural networks diagrams

Send text to girlfriend in the morning

Generative Art Using Neural Visual Grammars and Dual Encoders

Multi-Modal Machine Learning toolkit based on PaddlePaddle.

This program creates a formatted excel file which highlights the undervalued stock according to Graham's number.

Tgbox-bench - Simple TGBOX upload speed benchmark

This code is for our paper "VTGAN: Semi-supervised Retinal Image Synthesis and Disease Prediction using Vision Transformers"

An LSTM for time-series classification

Source code for TACL paper "KEPLER: A Unified Model for Knowledge Embedding and Pre-trained Language Representation".

Implementation of our NeurIPS 2021 paper "A Bi-Level Framework for Learning to Solve Combinatorial Optimization on Graphs".

This repository contains a PyTorch implementation of the paper Learning to Assimilate in Chaotic Dynamical Systems.

Attention over nodes in Graph Neural Networks using PyTorch (NeurIPS 2019)

MCMC samplers for Bayesian estimation in Python, including Metropolis-Hastings, NUTS, and Slice

Dataset Condensation with Contrastive Signals

This is Official implementation for "Pose-guided Feature Disentangling for Occluded Person Re-Identification Based on Transformer" in AAAI2022

Awesome Weak-Shot Learning