Hello world

• This project provides scripts to handle audio data processing & training a TensorFlow Danger Classification Model using the YAMnet audio embedding extraction layer (transfer learning)

Enviroment (Conda based)

• Python Version:
  • dev: 3.12.9 (github_codespace)
  • training: 3.11.11 (google_colab)
• Target OS: Linux (No support for windows)
• Operate: Ubuntu 22.04.4 LTS x86_64 (Google colab Q1 2025)
• Develop: Debian GNU/Linux 12 x86_64 (See /.devcontainer/devcontainer.json)
• CUDA Version: 12.*

Dependencies

• See /environment.yml for conda
• See /requirements.txt for pip

Setup

Using conda

1. install conda/ miniconda into your distribution
2. create conda environment from environemnt.yml

conda env update --file ./environment.yml --prune

3. activate conda environment

conda activate sra-env

4. Run setup

python setup.py

5. Training

python workflow.py

Using pip

1. run setup script

python setup.py

2. run setup script

python workflow.py

Insights

/workflow.py procedure

• For each dataset:

  • Download dataset
  • Filter sound file based on defined sound classes at /classes.csv
  • Move filtered sound files into the main dataset path /dataset/
  • Normalize:
    • Rename filtered sound files in /dataset/ to system rule f"{class_name}_{original_dataset}_{original_idx}"
  • Save filtered dataset meta into a .csv file in /ds/meta
  • Append filtered dataset meta to the main dataset meta in /ds/meta

• After pre-processing each dataset

  • Augmentation
    • Calculate mean datapoint count for each label in classes.csv
    • For each label in classes.csv
      • if count < mean
        • duplicate random data in label to match with mean
      • if count > mean
        • remove random data in label to match with mean
  • Normalize:
    • For each .wav file in /dataset/
      • Convert to PCM 16 wav format
      • Convert bit depth from any to 16
  • Assign fold label for each data point
  • Save augmented, folded dataset state into a .csv file

• Training: See details at workflow.py:train()

How to add & implement new dataset for training?

0. Register the new dataset info object to datasets.json (based on the previous format)
1. Human class mapping for added dataset to system-defined labels id in config.json:classmapping.default
2. Create a new dataset .py file inside /ds/, filename must be identical to the dataset key field value at the registered object in step 0
3. Create the dataset class, extending DataSet from ds.dataset.py
4. Implement necessary methods
5. Add the dataset class constructor to datasets_registry in workflow.py
6. Run script

workflow.py --process_data_only

7. Run script :D

workflow.py --use_processed