BoO-18/CNN-MNIST-Pytorch

mnist classification with pytorch using Convolutional NeuralNetwork and NeuralNetwork

CNN-MNIST-Pytorch

a simple mnist classification with pytorch using Convolutional NeuralNetwork and NeuralNetwork.

Dependencies

• python 3.6 +

• pytorch 1.0 +

• logging

• torchvision

• matplotlib

How to use

• Train MNIST models:

  • For CNN model(you should choose a model in the code): python CNN_MNIST.py
  • For NN model: python NN_MNIST.py

• Test our models:

  • To draw a diagram for our models: python plot_linechart.py
  • To test our models on realistic image: python USE_CNN.py

Some results

Performance

Results of some comparative tests.

Model	Accuracy	Loss
NN(NeuralNetwork)	95.33%	0.1687
CNN with 5*5 kernel	98.75%	0.0065
CNN with 3*3 kernel + padding	98.20%	0.0038
CNN with 5*5 kernel + padding	98.90%	0.0167
CNN with 5*5 kernel + BatchNorm2d	99.15%	0.0001
CNN with 5*5 kernel + dropout	98.50%	0.0302

Test on realistic image