Awesome

CTAL: Pre-training Cross-modal Transformer for Audio-and-Language Representations

Installation

• Python >= 3.6
• PyTorch version >= 1.7.0
• For pre-traininng, please prepare the GPU with at least 16GB Memory (V100, RTX3080Ti)
• To develop locally, please follow th instruction below:

    git clone https://github.com/Ydkwim/CTAL.git
    cd CTAL
    pip install -r requirements.txt

Preprocess

• Semantic Feature: please refer to the jupyter notebook: nontebook/preprocess_text.ipynb

• Acoustic Feature: please refer to the jupyter notebook: notebook/preprocess_audio.ipynb

Upstream Pre-training

After you prepare both the acoustic and semantic features, you can start to pre-training the model with executing following shell command:

    python run_m2pretrainn.py --run transformer \
    --config path/to/your/config.yaml --name model_name

The pre-trained model will be saved to the path: result/transformer/model_name. For the convenience of all the users, we make our pre-trained upstream model available:

• CTAL-Base: https://drive.google.com/file/d/1erCQplU9it9XBNrWDyLutekKthsZIi0q/view?usp=sharing

• CTAL-Large: https://drive.google.com/file/d/1L_QIZVRybJiiG2NywcX5xQQw8Y-3Vq5I/view?usp=sharing

Downstream Finetune

It is very convient to use our pre-trained upstream model for different types of audio-and-language downstream tasks, including Sentiment Analysis, Emotion Recognition, Speaker Verification, etc. We prepare a sample fine-tuning script m2p_finetune.py here for everyone. To start the fine-tuning process, you can run the following commands:

• Sentiment Regression:

    python m2p_finetune.py --config your/config/path \
    --task_name sentiment --epochs 10 --save_path your/save/path

• Emotionn Classification:

    python m2p_finetune.py --config your/config/path \
    --task_name emotion --epochs 10 --save_path your/save/path

• Speaker Verification:

    python m2p_finetune.py --config your/config/path \
    --task_name verification --epochs 10 --save_path your/save/path

Contact

If you have any problem to the project, please feel free to report them as issues.