yyc / cs3244

CS3244 (2018) Group 25

Component Regularization for Domain-Specific Image Classification

Set up & Quick Start

Baseline

Exp2a

python2 exp2a.py --num_epoch 40

Overview

Experiment1 Baseline model - baseline.py

Experiment2a Shi Yuan's - not yet created

Experiment2b Yu Chuan's - not yet created

Dataset (contains images and their classification)

• data.h5, can be 'train', 'val', or 'test'

  • key: 'ims_', value: array, i-th row contains i-th image of shape (3, 256, 256)

  • key: 'classes_', value: array, i-th row contains category (as int) of i-th recipe

  • key: 'impos_, value: array, i-th row contains list of image ids for the i-th recipe Note! image ids returned by this array are 1 more than the true image id

  • Others:

  • key: 'ids_', value: array, i-th row contains string id of i-th recipe

Helper package (/helper)

• classes.py to extract category labels from classes1M.pkl
• image.py to show image, has example code to show all images of a particular category

Others

• mnist.py, digits classifier created following a tutorial
• resnet50.py, extract_titles.py unused?

Details

If you are not too familiar with keras, you can try out the tutorial in https://elitedatascience.com/keras-tutorial-deep-learning-in-python which is done once in mnist.py. Some of the method calls are different due to changes in the keras api

We are trying to create 3 NN models:

1. Baseline model, normal image classifier
   • ResNet50 (an existing image classifier) to a fully connected layer
   • Analyze its performance
   • Refer to baseline.py
2. Shi Yuan's
3. Yu Chuan's

Dataset

hdf5 file which is like a key-value database containing images, classes etc.