Kaggle In-Class Project for Object Recognition in tiny images:
===============================================================

Team: Katch22
members: Harsh Upadhyay(hu3) and Suguman Bansal(sb55)

This Readme file is just an introduction to the various files present in this project that we made use of
from time to time during our search for the best classifier for the CIFAR10 image set.

The folder also contains a directory called Output which contains codes, Logs and weights of our models generated as we
worked.

Output Folder:
==============
1. Folders named attempt<X> which contain some results that we didn't submit to kaggle due to their poor accuracy on the validation set
2. Folders named kaggle_sub<X> contain data and codes from our kaggle submissions
3. O1 and O2 are older data
4. The "Underfit" folders contain data for when our CNN had too many layers.

Root Folder:
=============
> cnn.py : this is code for a CNN created using the scikit-neuralnetworks library. We gave up on it since we never got more than 10% validation accuracy
> kerasCnn.py : this one trains a CNN using the training data and has options to augment the input data
> utils.py : contains the util functions used by kerasCnn.py, ensemble.py, visualize_weights.py, predict.py and predict_single.py
> image_utils.py : contains a couple of functions reused from hw3 and only used by the cnn.py file. Not used for our final model that had keras
> visualize_weights.py : script to generate visualizations of the Convolution Layer weights in the CNN.
> predictions#.csv : contains csv prediction files uploaded to kaggle at different points. They are also there in the Output folders
> Other files are just there.

The Latest Model with 85.25% accuracy was generated using ensemble.py and The latest copy of the code is in the root directory.