Skip to content

Latest commit

 

History

History
38 lines (35 loc) · 3 KB

README.md

File metadata and controls

38 lines (35 loc) · 3 KB
Raw

ClassiLearn: Streamlined classical classification

This repository modularizes the familiarization with classical, non-ML/AI classifiers using user-provided data, accepting audio recordings of speech and videos of human faces.

High-Level Usage

  1. Feature extraction: choose 1 of 2 available data modalities and provide input files to their respective characterization (feature extraction) script.
  2. Optional dimensionality reduction: apply PCA (principle component analysis) and/or LDA (linear discriminant analysis) to the extracted features.
  3. Classification: select reduced or original features and perform classification with cross validation (CV) and grid search, with various fusion strategies available.

1. Feature Extraction

2. Dimensionality Reduction

Handled by dim_reduce.py. For individual argument description, run python dim_reduce.py --help

  • Input: manually specify the path to the outputted features of 1 of the modalities
  • Configurations: option to choose 1 of PCA or LDA; specification of feature type to preserve from original extraction. option of early fusion
  • Output: appropriately selected and simplified features. Whether any configurations are set, output matches default format of classification input.

3. Classification

Run python classification.py --help for details

  • Default input: features file formatted by dim_reduce.py
  • Custom input: manually select original features extracted by 1 modality (omit step 2)
  • Configurations: option of classifier late fusion, feature late fusion, and probability fusion
  • Output (stored and overwritten in classification_output directory):
    1. classification scores, as both mean & stdev and raw scores across CV folds
    2. graphic of class probabilities by classifiers
    3. graphic of train & test split and correctness within feature space, if 2-dimensional.

Dependencies

Both the audio and video features require specific manually installed dependencies. Please refer to their details linked above.

  • disvoice (which requires specific versions of the following)
    • python=3.10
    • numpy=1.22.4
    • scipy=1.11.4
  • opencv-python
  • matplotlib
  • mediapipe
  • pandas
  • scikit-image
  • scikit-learn