Meta-learning with Latent Embedding Optimization (LEO)

CS-502: Deep learning in Biomedicine at EPFL. Project Option 2. Implementation of the method called Latent Embedding Optimization (LEO). Original paper can be found here.

Files and Their Descriptions:

• methods/leo.py: Actual file that contains implementation of LEO.
• conf/method/leo.yaml: File that contains configurations and fixed hyperparameters for LEO.
• pretrained_weights/*.tar: Files that are contain pretrained backbone weights for the algorithm.
• Other already existing files are updated such as run.py and utils/io_utils.py for compatibility with new algorithm and reproducibility of the results.
• Grid Search results for Swissprot can be found in branch grid-search-swissprot under folder checkpoints\exp_name\results.txt and for Tabula Muris in branch boris\grid-search-tabula-muris under folder output\exp_name_results.txt.
• In addition to this, detailed outputs for all experiments/training can be found under output\exp_name.txt in branches grid-search-swissprot and boris\grid-search-tabula-muris.
• Note: Each experiment name in the results files is self-explanatory, which contains the used options and their values.

How to Run

For the new method (LEO), execute run.py with method=leo command line argument, and with the configuration wanted. E.g. running command

python run.py exp.name={exp_name} method=leo dataset=swissprot n_shot=1 method.latent_space_dim=64 method.enable_finetuning_loop=True method.optimize_backbone=True

will train LEO model with latent space dimension set to 64, with inner finetuning loop enabled, with outer optimization of backbone enabled and with random initialization for backbone (default). It will get the results on Swissprot for 5-way 1-shot case.

Authors

• Abdurrahman Said Gürbüz
• Boris Zhestiankin
• Sergi Blanco-Cuaresma