mnist_tutorial

A tutorial for MNIST handwritten digit classification using sklearn, PyTorch and Keras.

Code structure

• numpy_matplotlib_sklearn.ipynb: for numpy, matplotlib and sklearn.
• pytorch.ipynb: for pytorch.
• keras.ipynb: for keras.
• Reference solution: (not published yet)
  • numpy_matplotlib_sklearn_solution.ipynb
  • pytorch_solution.ipynb
  • keras_solution.ipynb

Requirements

Code tested on following environments, other version should also work:

• linux system (ubuntu 16.04)
• python 3.6.3
• numpy 1.13.3
• matplotlib 2.1.0
• sklearn 0.19.1
• pytorch 0.4.1
• keras 2.1.2

Used Deep Learning structure

• numpy_matplotlib_sklearn.ipynb: for numpy, matplotlib and sklearn. [done]
• pytorch.ipynb: for pytorch. [done]
• keras.ipynb: for keras. [done]

Some results for Q1~Q6

Q1:

Training accuracy: 97.55%
Testing accuracy: 87.90%

Q2:

Training accuracy: 82.03%
Testing accuracy: 79.90%

Q3:

Training accuracy: 97.85%
Testing accuracy: 83.60%

Q4:

Training accuracy: 97.90%
Testing accuracy: 83.70%

Q5:

training accuracy：	 0.998046875
testing accuracy：	 0.9893830128205128

Q6:

Training accuracy: 1.00%
Testing accuracy: 0.99%