2018 UIdaho Student Data Science Competition

The machine learning competition will ask participants to grade 20,000 handwritten math quizzes. Can you develop a model that performs better than a random guess? Will you climb to the top of our Leaderboard?

Three prizes will be awarded to the top scorers:

• Golden Vandal: a University of Idaho scholarship of $500.
• Silver Vandal: a University of Idaho scholarship of $300.
• Iron Vandal: a University of Idaho scholarship of $200.

This repo

Here is the code for one of my more successful models for the student data science competition. It is built on PyTorch, and uses the PyTorchNet basic framework to load a malleable configuration and training/testing structure. That makes it a bit harder to dive into than a single long file unfortunately.

This model is based on VGG net.

Methodology

The idea is pretty simple, use a large VGG like network with some minor randomly applied transformations to the train images.

• Train single network on full set of numbers
• Segment testing images into even fifths (i.e. [1] [+] [2] [=] [4])
• Predict numbers (generate a probability vector)
• For positions 2,4 only take max probability of symbols (i.e. >= 10)
• For other positions use only numbers (i.e. < 10)
• Use predicted numbers/symbols to check if equation is valid
• Set answers to 1 if true, else 0

Files

The most important files are below:

• models/vgg.net - contains the model architecture used
• datasets/uisdsc.py - Automatically downloads the dataset and performs transformations for the model
• main.py - simply grabs the dataloader and runs the testing and training in a loop.
• train.py - the bulk training function.
• test.py - how the test data gets evaluated

Where is the full model?

For my final solution I trained three models:

• this VGG model
• A ResNet-like model
• A densenet-like model

However, they were each trained individually using a similar framework and in the end I wrote a sloppy script that took the output files I had copied/pasted into another directory to merge their weighted results. The whole framework was pretty poorly done and obviously more than a little bit of a hackjob. If anyone really wants it, I can take some time to clean it up and post it.

Usage

1. Spin up visdom python -m visdom.server
2. Edit/view args.txt to change any preferences
3. Run it: python main.py

Other solutions

Camden Clark - The winner behind us all.