Skip to content

Latest commit

 

History

History
131 lines (91 loc) · 5.1 KB

readme.md

File metadata and controls

131 lines (91 loc) · 5.1 KB
Raw

Audio Classification SNP

This repository contains a Deep Learning project focused on the classification of audio data. It includes various scripts for data preprocessing, model training, prediction, and hyperparameter tuning. Additionally, a pipeline script is provided for executing the entire processing flow with data storage in MongoDB.

This project is based on the repository Audio-Classification by Seth Adams.

Getting Started

To get started, follow these steps:

  1. Create a conda environment with Python 3.9: conda create -n audio python=3.9

  2. Activate the conda environment: conda activate audio

  3. Install the required packages: pip install -r requirements.txt

  4. Run the scripts as needed. Pass the function arguments when executing the scripts, for example: python clean.py --src_root audiofiles/wav.

Training Model

Use the following scripts to develop your models and test them. Select a samplerate and desired delta time which remain consistent throughout the process.

1. Audio Preprocessing

The clean.py script can be used to preview the signal envelope at a threshold and remove low magnitude data. Uncommenting split_wavs will create a clean directory with downsampled mono audio split by delta time.

  • clean.py - Purpose: Preprocesses audio data. - Inputs: - src_root: Source directory of raw audio files. - dst_root: Destination directory for preprocessed files. - delta_time: Time frame for processing. - sr: Sampling rate. - fn: Feature normalization option. - threshold: Threshold for filtering.

2. Model Parameters

In modles.py various model parameters depending on the selected model. These parameters can be adjusted to customize the model architecture and behavior.

  • models.py - Purpose: Contains model architectures (Conv1D, Conv2D, LSTM). - Inputs: Various model parameters depending on the selected model.

3. Model Training

In train.py the model training is done using the specified inputs in modles.py.

  • train.py - Purpose: Trains the model on preprocessed audio data. - Inputs: - model_type: Type of the model (Conv1D, Conv2D, LSTM). - src_root: Directory of preprocessed data. - batch_size: Size of each training batch. - delta_time: Time frame for processing. - sample_rate: Sampling rate.

Use PlotHistory.ipynb to visualize the training process.

4. Prediction

In predict.py a trained modell can be tested doing inference.

  • predict.py - Purpose: Performs inference using a trained model. - Inputs: - model_fn: Filename of the trained model. - pred_fn: Filename for saving predictions. - src_dir: Directory of data for prediction. - dt: Delta time for processing. - sr: Sampling rate. - threshold: Threshold for filtering.

5. Hyperparameter Tuning

The following scripts are used for hyperparameter tuning using random search. This step is optionl but can be useful in finding the best parameters.

  • hyperparametertuning_conv1d.py

  • hyperparametertuning_conv2d.py

  • hyperparametertuning_lstm.py

  • Inputs: - objective: Objective function for optimization. - max_trials: Maximum number of trials. - executions_per_trial: Number of executions per trial. - directory: Directory for saving results. - project_name: Name of the tuning project.

  • Parameters for Optimization: n_mels, n_fft, win_length, hop_length, dropout_rate, batch_size.

Run Pipeline

Use the following script to classify your data with pretrained models and save the results in MongoDB:

pipeline/00_Pipeline.py

  • Purpose: Script to execute the entire processing pipeline with storage in MongoDB.

  • MongoDB Connection Information:

    • MONGO_URI: MongoDB connection URI.
    • DB_NAME: Name of the MongoDB database.

  • Collection Names:

    • AUDIO_COLLECTION_NAME: Name of the collection to store audio classification results.
    • AGGREGATED_COLLECTION_NAME: Name of the collection to store aggregated audio data.
    • AGGREGATED_PER_DAY_COLLECTION_NAME: Name of the collection to store aggregated daily audio data.

  • Model Paths:

    • MODEL_PATHS: Dictionary containing the paths to the pre-trained models for different audio classes.

  • Directory Path:

    • DIRECTORY_PATH: Path to the directory containing the audio files.

  • Confidence Threshold:

    • CONFIDENCE_THRESHOLD: Threshold value for the confidence score of audio classification.

  • Maximum Threads:

    • MAX_THREADS: Maximum number of threads to use for processing the audio files.

Contact

For any questions or contributions, please open an issue in the repository.