Latent Aggregation

Aggregating seemingly different latent spaces.

Quickstart

Development installation

Setup the development environment:

git clone git@github.com:crisostomi/latent-aggregation.git
cd latent-aggregation
conda env create -f env.yaml
conda activate la
pre-commit install

Run the tests:

pre-commit run --all-files

We use HuggingFace Datasets throughout the project; assuming you already have a HF account (create one if you don't), you will have to login via

huggingface-cli login

which will prompt you to either create a new token or paste an existing one.

Update the dependencies

Re-install the project in edit mode:

pip install -e '.[dev]'

Experiment flow

Each experiment exp_name in part_shared_part_novel, same_classes_disj_samples, totally_disjoint has three scripts:

• prepare_data_${exp_name}.py divides the data in tasks according to what the experiment expects;
• run_${exp_name}.py trains the task-specific models and uses them to embed the data for each task;
• analyze_${exp_name}.py obtains the results for the experiment.

Each script has a corresponding conf file in conf/ with the same name. So, to run the part_shared_part_novel, you have to first configure the experiment in conf/prepare_data_part_shared_part_novel.yaml. In this case, you have to choose a value for num_shared_classes and num_novel_classes_per_task. Now you will prepare the data via

python src/la/scripts/prepare_data_part_shared_part_novel.py

this will populate the data/${dataset_name}/part_shared_part_novel/ folder. Then you'll embed the data by running

python src/la/scripts/run_part_shared_part_novel.py

so that now you will have the encoded data in data/${dataset_name}/part_shared_part_novel/S${num_shared_classes}_N${num_novel_classes_per_task}. Having all the latent spaces, you can now run the actual experiment and collect the results by running

python src/la/scripts/analyze_part_shared_part_novel.py

The results can now be found in results/part_shared_part_novel.