MoPoE-MIMIC
This repository contains the code for the framework in Multimodal Generative Learning on the MIMIC-CXR Database (see paper).
It is based on the framework used in Generalized Multimodal ELBO (see paper, code).
Installation
git clone https://github.com/Jimmy2027/joint_elbo
cd joint_elbo
git checkout hendrik_mimic
path/to/conda/environment/bin/python -m pip install .
For development, install with:
git clone https://github.com/Jimmy2027/joint_elbo
cd joint_elbo
git checkout hendrik_mimic
path/to/conda/environment/bin/python -m pip install -e .
to enable testing:
git clone https://github.com/Jimmy2027/joint_elbo
cd joint_elbo
git checkout hendrik_mimic
path/to/conda/environment/bin/python -m pip install -e .[test]
Note
If pip throws an SSL Error, create first a new conda environment with conda env create -f environment.yml, and then install mimic using the steps above.
Usage
Run the main training workflow with:
cd mimic
python main_mimic.py
A json config in configs can be used to give arguments to main.py with the flag --config_path. Note that the parameters in the config will be overwritten by the arguments passed through the command line.
cd mimic
python main_mimic.py --config_path path_to_my_json_config
Otherwise an additional condition can be added in mimic.utils.filehandling.get_config so that the config is found automatically.
Training the classifiers
See here for instructions on how to train the classifiers.
Testing
run unittests with:
cd mimic
python -m pytest tests/
or more specifically:
cd mimic
python -m unittest tests/test_that_you_want_to_run.py
Creating the tensor dataset
The tensor dataset can be created with the script dataio/create_tensor_dataset.py.
The creation of the tensor dataset consists of two steps. In a first step, the images of the original dataset are resized to a wanted size and stored as jpg in a folder.
The first step is only executed if the folder of the resized images, or a zipped version of it is not found.
During the second step, the jpg images are read into a torch tensor and saved as such.