fashion-mnist-tf

Multi-class classification for Fashion-MNIST in tensorflow

Assignment 3 code for Deep Learning, CS60010.

MNIST data provides us very high accuracy with simple models, so we will be using fashion-MNIST.

The neural network has 3 hidden layers, with 50 epochs/iterations. Refer to Report for more details.

Performance.

• Training accuracy: 96.83% (Max accuracy in an iteration: 100%)

• Testing accuracy: 89.46%

• Loss as a function of iterations

• Accuracy as a function of iterations

Layers

We apply Logistic Regression at every hidden layer. Here are the results:

• Layer 1: 88.87%
• Layer 2: 89.33%
• Layer 3: 89.46%

The first layer seems to provide enough accuracy, which proves further layers might not be needed.

Usage

python train.py --train

Run training, save weights into weights/ folder.

python train.py --train iter=5

Run training with specified number of iterations. Default iterations are 50.

python train.py --test

Load precomputed weights and report test accuracy.

python train.py --layer=1

Run Logistic Regression on hidden layer's output and report the accuracy. Allowed options : 1, 2, 3.

Code structure

• data_loader is used to load data from zip files in data folder.
• module defines the neural network parameters, and network related code.
• train handles input and states the model.

License

The MIT License (MIT) 2018 - Kaustubh Hiware.