Skip to content

Latest commit

 

History

History
50 lines (39 loc) · 1.91 KB

README.md

File metadata and controls

50 lines (39 loc) · 1.91 KB

MDL-By-MetaLearning

This is the repository for the paper "Multi-Domain Learning by Meta-Learning: Taking Optimal Steps in Multi-Domain Loss Landscapes by Inner-Loop Learning" published in ISBI 2021. Available at: https://arxiv.org/abs/2102.13147

Note, the data used in the paper cannot be made public. In any case, if you would like to run the code on a different dataset, we provide some details on how this might be done.

setup_experiments.py may be used to generate config files required to run experiments. File paths declared at the start should be followed or altered appropriately. We provide a (made-up) example to illustrate the appropriate contents of these file paths in fake-paths/example-fold.

Following this, the run.sh bash script contained in the generated experiment folder can be run to do all experiments listed in the folder (note, each experiment is associated with its own config subdirectory). We provide an example of the default experiment folder output by this script in example-experiment-config. Depending on the number of available GPUs in your system, you may need to modify the <GPU_IDX> scheme assumed in this bash script when calling train.py. We give more details on how to modify calls to train.py next.

If you want to run experiments using your own config files, you can do so by calling the following in your terminal:

python3 train.py <CONFIG_DIRECTORY> <GPU_IDX> <NUM_TRIALS>

You can follow the appropriate subcalls if your research requires modifying our pipeline. Most of the logic is contained in train.py and runner.py with other python files containing the necessary helper functions.

If you have any questions, don't hesitate to post an issue here or reach out at our emails (available in the paper). We will not actively maintain the codebase, but are happy to help in this capacity.

Finally, if you make use of our code or techniques, please cite our paper. Thanks.