This is the official repository for the paper CompA: Addressing the Gap in Compositional Reasoning in Audio-Language Models accepted at ICLR 2024.

[Website] [Paper] [CompA-order] [CompA-attribute] [CompA-661k] [Stage2 CSV] [Stage3 CSV]

You are required to install the dependencies: pip install -r requirements.txt. If you have conda installed, you can run the following:

cd CompA && \
conda create -n compa python=3.10 && \
conda activate compa && \
pip install -r requirements.txt

1. For Vanilla Training:

Use the following command after updating the train and val file in /src/laion_clap/train.sh from the src-stage1/laion_clap/ directory:

sh train.sh

2. For training with compositionally-aware hard negatives:

Use the following command after updating the resume ckpt (the ckpt from vanilla training), train and val file in /src-stage2/laion_clap/resume.sh from the src-stage2/laion_clap/ directory

sh resume.sh

3. For training with modular contrastive learning:

To generate the training file for stage 3 training, run the following command. The demo files required to run the code have been shared as well.

cd audio_mix
python pos-neg-generation.py -audio ./data/final_audio.csv -sound_class ./data/global_sound.yml

Use the following command after updating the resume ckpt (the ckpt from training with compositionally-aware hard negatives), train and val file in /src-stage3/laion_clap/resume.sh from the src-stage3/laion_clap/ directory:

sh resume.sh

The evaluation files need hook.py from CLAP repository. Please place the files in the CLAP/src/laion_clap/ folder and run the below commands as required.

1. For Zero-Shot evaluation:

python ./evaluation/zshot.py <test_files_dir_path> <class_label_to_idx_file_path> <clap_ckpt_path>

test_files_dir_path - Path to the folder which contains audio files and their respective jsons. This format can be obtained by using the audio-dataset repo.
class_label_to_idx_file_path - Path to the file which contains a class label and its respective index in the format of a Python dictionary. These files can be found in CLAP/class_labels.

2. For CompA-Order and CompA-Attribute evaluations:

python ./evaluation/benchmark_eval.py <benchmark_file_path> <audio_dir_path> <clap_ckpt_path>

benchmark_file_path - Path to CompA-Order or CompA-Attribute benchmark file.
audio_dir_path - After downloading and extracting from the link provided, the path to CompA_order_files or CompA_attribute_files folder.

This repository benefits from CLAP. Thanks for their awesome work.

@inproceedings{
ghosh2024compa,
title={CompA: Addressing the Gap in Compositional Reasoning in Audio-Language Models},
author={Sreyan Ghosh and Ashish Seth and Sonal Kumar and Utkarsh Tyagi and Chandra Kiran Reddy Evuru and Ramaneswaran S and S Sakshi and Oriol Nieto and Ramani Duraiswami and Dinesh Manocha},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=86NGO8qeWs}
}