This repository contains some experiments in which we search for good ways of converting data to images capable of class distinction. For more information, please refer to this partial report.

We have three experiments repositories: conv_iris/, conv_leaf/ and conv_titanic/. Let's choose conv_iris/ as an example.

• First, we have to create the images, which is done by the Mathematica notebook create_images.nb.
• Then, we have to create two lists: one with the names of the images in our training set followed by the class (in the same line) and one similar for the test set. This is done by the script make_names_list.py.
• Now we have to create the leveldb data that Caffe will use for train and test. This is done by the script create_data.sh.
• Finally, we have to train and test our CNN. We have just to run the script train.sh. Of course we can play with the parameters in the Caffe models configurations files.

There is also a script experiment.sh that will do almost all steps above 50 times, which is a good way to train a model with multiple dataset partitions. This script will not create the images, so we must create them before calling it.