broadinstitute/ml4ht_data_source

Multimodal data loader compatible with pytorch and tensorflow

ml4ht_data_source

ml4ht_data_source is a library that allows you to streamline the process of modeling multi-modal health data.

It makes it easy to load and model on data from different storage formats with complex QC and date time selection logic. The data can be used in both tensorflow and pytorch.

Library functionalities

The library can...	Example use case
Load data from different storage formats	Modeling on MRIs stored in hd5 files named by sample id with labels in a pandas data frame
Have clear and easy to use selection of data by sample id	Selecting a set of patients with a specific condition
Have clear and easy to use selection of data by date-time	Selecting ECGs that have a heart attack at most 10 days prior
Allow flexible data transformations	Comparing different augmentation strategies
Allow flexible data filtering	Comparing different QC strategies
Make data exploration convenient	Comparing distributions of labels after different QC and date selection strategies
Allow a random state to be shared across modalities	Selecting a random chunk of an MRI to segment

Setup

ml4ht_data_source uses python 3.6 or higher. Setup can be done using venv.

python3.8 -m venv env
source env/bin/activate
pip install -r requirements.txt
pre-commit install
pip install .

Tests

ml4ht_data_source is thoroughly tested using pytest.

source env/bin/activate
pip install .
pytest

ml4ht_data_source's Issues

Easy way to use the output of exploration as date and sample id selector

Once you run explore_data_descriptions, you should be able to use the valid sample ids and dates for the later PipelineSampleGetter easily.

Once you run explore_pipeline_sample_getter you should be able to use the output to define a new PipelineSampleGetter which skips the failed sample ids and dates.

Repo rename to ML4HT ?

To alliterate with other ml4h repos: ml4hs, ml4h, ml4h_ccds...

`DataDescription` that can return multiple modalities

Currently if you get multiple modalities from the same source, e.g. multiple columns from the same data frame, you have to have multiple DataDescriptions. It may be simplifying to have a DataDescription extension which can return a dictionary of arrays.

Add examples folder

There should be a folder of examples.
The most important example is define and running a full data exploration pipeline.

multimodal multi-date inference

Currently all-loading-option inference is possible for single modality, single task models. It should be possible for multi-modal models