Cifar 100 challenge is implemented via deep neural networks in theano using the theanet package. This repository is the basis for the simulations shown in Chapter 18 of the book Computer Age Statistical Inference by Bradley Efron and Trevor Hastie.

We have managed to get error rates as low as 35% on the test data, which is was state-of-the-art according to these benchmarks. To reproduce the 35% test error rate, you need to use the 55M.prms parameter file, then train, retrain and retrain.

Get the [tar.gz file] (http://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz) from the official site and unzip it to the root directory.

python train.py params/0baby.prms [1]     # For a test run with a small network
python train.py params/55M.prms [1]       # For a serious run with a huge network

This will train on the training data, (like for ever if you are using the bigger configurations) and output a pkl file with the parameters in it, that can be loaded later on. You also get a text file, if you specify the optional 1 at the end of the command, like 0baby_555555.txt. Here 555555 is a random random-seed. (This can be specified in the prms file.) You can edit the theanet params file to your liking. Sample params files are given in the params folder.

python train.py 55M_555555_42.pkl 

You could also give an existing pkl file as an input in stead of a prms file. This will initialize the network with those parameters and pick up the training from there. Although the learning rate could be low. The file edit_pkl.py will tell you how to edit the pkl files before re-training.

python test.py 55M_555555_42.re_37.re_35.pkl # This file has been obtained by triple retraining

This only opens the test dataset and saves the intermediate images as the input goes through the CNN.

• You need to use Python 3.0 (Come on over now, in to the new century).
• Numpy
• PIL
• Theano
• Theanet