Mukh-O-Mukhosh (Bengali: মুখ ও মুখোশ)

• A Convolutional Neural Network model for Face Mask Detection
• The name Mukh-O-Mukhosh is inspired from the name of the the first Bengali language feature film 'Mukh O Mukhosh' (Bengali: মুখ ও মুখোশ, lit. 'The Face and the Mask') which was released back in 1956. https://en.wikipedia.org/wiki/Mukh_O_Mukhosh
• This model will be trained on the dataset created by [prajnasb] [https://github.com/prajnasb/observations]

Roadmap:

• Background Study:

  • Artificial Neural Network (ANN) basics: perceptron, neural network, activation function, cost function, gradient descent, back propagation
  • Simple learning:
    • neural network with single input & single output feature
    • example: linear regression
  • Basic architecture of ANN:
    • multiple input features , hidden layers & multiple output features
    • data loading & splitting into train/test set
    • example: multiclass classification from continuous data (IRIS dataset)
  • General architecture of ANN:
    • concept of feature engineering, continuous & categorical data, embedding, batch normalization, dropout layer
    • example: regression from a mix of continuous & categorical data (NYC taxi dataset)
    • example: multiclass classification from a mix of continuous & categorical data (NYC taxi dataset)
  • Convolutional Neural Network:
    • getting familiarized with a basic image dataset (MNIST)
    • example: MNIST with ANN
    • motvation behind choosing CNN over ANN
    • basic concepts of CNN
    • example: MNIST with CNN

• Implementation:

  • dataset preparation
    • read images from directory
    • data preprocessing (resize, toTensor etc)
  • data loading
    • creating data loader (attr: batch size, shuffle/randomize etc)
  • defining model, loss function & optimizer
  • train the model
  • plot loss & accuracy (for training & validation)
  • experimenting with different hyperparams
  • training & validating using gpu/cuda