Description

A classical ML project for classifying digits from 0-9 based on a vector image with size 28x28. Dataset is the famous MNIST dataset (http://yann.lecun.com/exdb/mnist/).

ML model

Different ML models shown in this notebook work a little different. In general, ML models accept N*784 array with N elements (digits to predict); models outputs either an array N*1 specifying which number it predicts (Sklearn), or N*10.

Pipeline

• Load data for MNIST;
• Pre-processing:
  • Convert N*28*28 3D array of N images to N*784 2D array of N images;
  • Normalize: / 256
  • (depending on the model) Process labels: from 5, 4, etc. to [0,0,0,1,0,0,0,0,0,0]
• Train models:
  • Sklearn models: SGDClassifier, Random Forest, NN
  • Keras (Tensorflow): NN
  • Accuracy assessed on train set by cross-validation method
• Test models:
  • On the test holdout set
  • On the custom local images