This manual will cover the steps to execute the commands to perform the training and classification process.

• Linux, Ubuntu 16.04 LTS. There is no guarantee that this software, after having followed all the steps, can work correctly in a different version of the previous OS.
• Python 2.7.2 nativo.
• It is recommended that the computer has a Nvidia GPU for faster processing. If available, install Tensorflow for GPU, CUDA ToolKit 7.0 o greater and CuDNN.

The project folder called DeepLeishmaniaScan contains several files and folders that are important:

• models: contains initial configurations and model files
• datasource: where the input images are found
• fold [#]: subsets of datasource (from the K-Fold algorithm)
• fold-test [#] subset of test images.
• train.py: training script.
• classify.py: classification script.
• jsonreader.py: subroutine to read json files.
• InceptionV3_1.json: InceptionV3 base architecture.
• inceptionV3_1.h5: base weights of InceptionV3.

Within / models, other folders named with numbers serve as identification for all trained models. These folders contain the following files:

• [model-id] .json: file used to be read by Java application
• [model-id] -arch.json: architecture of modified model (different from InceptionV3_1.json)
• [model-id] .h5: weights of modified model
• runconfig.json: file that has the values of the hyperparameters used for a specific model.

The configurable hyperparameters in the runconfig.json file are:

• Number of times
• Learning rate
• Momentum
• Nesterov technique (True or False)
• Number of images per batch

To apply changes of any of these hyperparameterers you need to start another training session.

• To start training you must first open the terminal and go to the folder containing all folders previously mentioned (using cd).

$ cd /../DeepLeishmaniaScan/

• Once inside the folder, execute the following command

$ python train.py /models/[id de modelo disponible]/runconfig.json

The process will take some time depending on the number of epochs defined in the previous configuration file. At the end of the process, 3 files will be generated (or overwritten if they already exist):

• [modelID] -arch.json new model architecture.
• [modelID] .h5 contains new values of each of the weights in the trained network.
• output.txt containing several metrics, such as sensitivity, specificity, accuracy, precision, execution time, and others.

In the same folder, execute:

python classify.py /models/[id de modelo disponible]/runconfig.json [../../imagen.jpg|png|jpeg]

This command requires an additional parameter, which is the image's location. Output value will be the likelihood (%) of being infected with cutaneous leishmaniasis.