Currently only the code for PseudoTask is posted, a pull request for the Task Tangent Kernel aspects of our paper will be made in the coming weeks. As this code is newly released, it's only been tested on a limited number of machines/environments. Please let us know of any difficulties you encounter while using it and we'll be happy to help!

Most arguments are specified in the config files of config/, you can edit entries in these or pass paths as command line arguments to match to however your local machine is set up.

For main*py scripts, arguments are primarily handled by the config files. The only neccessary command-line arguments to replicate experiments are

dataset.task_id=`seq -s, 0 n` -m

where n is 24 for cub-only, 29 for cars, or 49 for cub-inat. The seq expands out into a comma-separated range and the -m flag tells the script to execute in multi-run mode.

LEEP/RSA/Cui et al.'s EMD do not require the task_id flag as the iteration is done automatically not using the hydra module. These can be run from the corresponding embed*py.

For accuracy_embeddings*py scripts, the primary argument is the location of the embeddings file. This is typically given as --root [path to file dir]. Note that the CUB Attribute evaluation requires two roots as attributes are only defined for CUB so the iNat embeddings from CUB-iNat are typically re-used.

LEEP/RSA/Cui et al.'s EMD have a hardcoded path that their corresponding embed script saves to (these are for cub-inat only). The TTK accuracy scripts have hardcoded paths to the example similarity matrices.(TODO: Add w/ TTK instructions)

To train new experts use train_experts.py, task_ids are as above. To train on CUB attributes use dataset.name=cub_attributes and dataset.attribute=[desired attribute] (e.g. breast_pattern). To train car experts use dataset.name=cars.

To transfer experts among each other use transfer_experts.py (with attribute differences as before) and transfer_experts_cars.py

print_pseudolabel_preds.py is run similarly to main*py scripts to display the pseudolabels versus actual class index for a given task.

Conda environment is in environment.yaml, run conda env create -f environment.yml

Cars: https://drive.google.com/file/d/1deqk9Ep3yf7ejwXV3mgQ1A6ZWIf5XyRS/view?usp=sharing

CUB: https://drive.google.com/file/d/1IwaFpQnbANlSjZNZ1NLZg7MxRcUblUhH/view?usp=sharing

iNat (info only): https://drive.google.com/file/d/1KqNc52ApY4axccVWdDOMBGL582WWzQsB/view?usp=sharing

iNat (images): https://github.com/visipedia/inat_comp/blob/master/2018/README.md#Data (train and val images)

Embeddings are available at: https://drive.google.com/file/d/1DkQQfQpvSLxv7p6C1n_vzXjth8jasyQh/view?usp=sharing

Note that there is slight variation between the numbers produced by the above embeddings and our paper.