SUR - audio and visual person verification

Authors: Skuratovich Aliaksandr, Tikhonov Maksim

Date: 25.4.2022

Structure

src/CNN/: implementation of a Binary clasification based on a Convolutional Neural Network
src/MAP/: Maximum A-posteriori Classifier based on the Bayesian Gaussian Mixture models.
src/NEURAL_PCA/: Binary classification using feed-forward neural network wirh preprocessing and feature extraction with Principal Component Analysis.
src/models/: Directory with ready models.
src/main.py: The main script.
src/hyperparams.yaml: hyperparameters to perform the training and/or evaluating.
dataset/: Small dataset with .wav and .png files to train on.

Running

NOTE: your version of python has to be >= 3.10.0

Install the requirements

pip install -r requirements.txt

All running options listed in config hyperparams.yaml

python3 main.py hyperparams.yaml

Here are the parameters from hyperparams.yaml:

default: &default
  train: False  # Train or used pretrained model
  eval: True  # evaluation is set create reports with the scores for each file from the test set. 
  model_dir: ./models  # Directory to load/store the trained model
  dataset_dir: ./dataset  # Path to a dataset to train on.
  eval_dir: ./tests  # Directory with the test set.
  GPU: 0  # If you train on the GPU machine, change this cell
  root_dir: .
  wandb_entity: <your_wandb_username_here>

CNN:
  <<: *default
  train: False
  eval: False
  model_name: CNNKyticko.pt

MAP:
  <<: *default
  dataset_dir: {non_target: ./dataset/non_target_train, target: ./dataset/target_train}
  model_name: {target: bgmm_target.pkl, non_target: bgmm_non_target.pkl}
  dev_dataset: {non_target: ./dataset/non_target_dev, target: ./dataset/target_dev}

Neural_PCA:
  <<: *default
  model_name: NeuralPCA.pt
  u_mean: ./NEURAL_PCA/u_mean.npy

SkuratovichA/SUR

SUR - audio and visual person verification

Structure

Running

Install the requirements

All running options listed in config hyperparams.yaml

TODOS