Diabetic retinopathy

Environment

Download miniconda: wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
Install miniconda: bash Miniconda3-latest-Linux-x86_64.sh
To create the environment to run the model, simply do conda env create -f <path to>/environment.yml.

You can find the appropriate environment.yml here.

Cluster Environment

To have an efficient environment on SLURM clusters, you can pack your environment and unpack it at the beginning of you job on the local SSD.

Assuming you installed the environment in the section above and then activated it. The following command will generate retino.tar.gz which will contain your entire environment.

conda pack -n retino

Then add the following snippet at the beginning of your SLURM launch script and you should be all set.

# Make sure no previous conda env is activated
if command -v conda; then
  conda deactivate
fi

# Create env folder
ENV_PATH=$SLURM_TMPDIR/retino
mkdir -p $ENV_PATH

# Extract packed env
tar -xzf <path to>/retino.tar.gz -C $ENV_PATH

# Activate the environment. This adds `$ENV_PATH/bin` to your path
source $ENV_PATH/bin/activate

# Cleanup prefixes from in the active environment.
conda-unpack

For more information see the conda-pack documentation.

Git Hooks Activation

Make sure you have git version > 2.9. After cloning, run the following command once to activate the proper hooks.

From the root of the repo run git config core.hooksPath .githooks.

Config file

The hyperparameters to use are defined in a yaml config file, see config.yml for an example, and are passed to the model using the --config flag.

Note that in the model/base sub-config the normalization layer norm_layer hyperparameter is fully dynamical and assumes you know which parameter to use with which norm_layer/model combination. No checks are made. For example, if you use group_norm with resnet, you need to also pass num_groups and it needs to be a divider of 64.

Hyperparameter search (Orion)

Setup

Install orion in your env. pip install orion

Config

You need to specify the search spaces in the config file. See orion.example.conf.yml for an example.

The templating used in the config file is manually and only partially supported.

Run

Then to launch a job in the specified search space using the following command.

orion hunt -c .orionconfig.yml -n kaggle --worker-max-trials 1 python main.py --data-folder data/kaggle2015/ --experiment-folder results-eye --config orion.example.conf.yml

To run orion on a slurm cluster see the documentation of the included example.

Museau/Retino