ASPP_Project

Project for Graduate Course in Advanced Scientific Programming in Python - Uppsala University

MNIST_Classification.ipynb is a Jupyter notebook to run a simple convolutional neural network classification on the MNIST dataset (http://yann.lecun.com/exdb/mnist/). The dataset consists of 60,000 samples for training and a test set of 10,000 samples of handwritten digits. The digits have been size-normalized and centered in a fixed-size image. The input images have the dimensions [28,28,1]

It uses torch as a backend for this machine learning task and uses the following packages:

from future: print_function

argparse

torch

from torchvision: datasets, transforms

The architecture and paramenters for the model look as follows:

Conv2d Layer (3,8,'Padding',1)

BatchNormalization Layer

ReLU Layer

MaxPool Layer (2,'Stride',2)

Conv2d Layer (3,16,'Padding',1)

BatchNormalization Layer

ReLU Layer

MaxPool Layer (2,'Stride',2)

Conv2d Layer (3,32,'Padding',1)

BatchNormalization Layer

ReLU Layer

Fully Connected Layer (10)

Softmax Layer

Classification Layer (CrossEntropy)

Training parameters:

ADAM optimizer

GradientDecayFactor (beta1): 0.9000

SquaredGradientDecayFactor (beta2): 0.9990

Epsilon: 10^(-8)

LearningRate: 0.01

Epochs: 10

L2-reg: 10^(-4)

MiniBatchSize: 128

Already after the first epoch an accuracy of 98% is reached on the test data.