This is the companion code for the paper "Error-correcting output codes in the framework of deep ordinal classification". The code is prepared to automatically initialize and run all experiments as described in the paper.

The experimentation results from the original paper are available in file results.xlsx.

Authors are:

• Javier Barbero-Gómez (@javierbg)
• Pedro Antonio Gutiérrez (@pagutierrez)
• César Hervás-Martínez ([email protected])

The following has been tested to run on an up-to-date Linux installation (Debian 10 buster).

You should first install the latest version of conda. The small version known as miniconda is recommended.

For instructions on how to install miniconda check this link.

Then, create a new environment named "ordinal-cnn-ecoc" populated with all the needed packages using the requirements.txt file:

conda create --name ordinal-cnn-ecoc --file requirements.txt

Once finished, activate the environment with the command:

conda activate ordinal-cnn-ecoc

• Diabetic Retinopathy dataset:
  • Download files train.zip.001-005 and decompress
  • Download trainLabels.csv.zip and decompress
  • Run the partitioning script as:

python partition_retinopathy.py <path to trainLabels.csv> <path to train images folder> <path to folder to contain transformed images> <path to folder to contain all partitions>

• Adience dataset:
  • Download files fold_0_data.txt-fold_4_data.txt and place in a common folder
  • Download aligned.tar.gz and decompress
  • Run the partitioning script as:

python partition_adience.py <path to folder containing folds> <path to images folder> <path to folder to contain transformed images> <path to folder to contain all partitions>

First you should initialize all jobs to submit. Just run the init.py script:

python init.py <retinopathy_dataset_partitions_path> <adience_dataset_partitions_path>

A "workspace" folder should have been created containing all project data, as well as other project related files.

The experiments are designed to be ran in a GPU cluster using a workload manager such as Slurm, PBS, LSF or HTCondor. Refer to the signac project documentation for instructions on how to configure your environment.

Once configured, all experiments can be submitted simply by using the command:

python project.py submit -o experiment

When all jobs have been completed, the results can be extracted in tabular format using the extract_results.py script:

python extract_results.py results.xlsx