Probabilistic Ratio Cut (PRCut) Implementation

PyTorch implementation

Environment setup

• Create a python environment and make sure that poetry package is installed
• Activate the python environment and from within the code folder call poetry install. This should install all the dependencies and the package
• For producing the Turtle results, we use the author's implementation, which should be cloned to the same folder for run_turtle.py to work

git clone https://github.com/mlbio-epfl/turtle
## This step is important to be able to import turtle as a package
touch turtle/__init__.py

Computing representations

There are 2 representation types we can extract:

• the ones that come from Dino and CLIP:

python extract_vit_representation.py --dataset cifar10 --model dinov2 --root-dir /buckets/ml --device-num 1

This will download the dataset to /buckets/ml/ if it doesn't already exist, then creates the folder /buckets/ml/representations/dinov2/ that contains cifar10_[feats/y]_[train/val].npy

• the ones that come from solo Learn package

python extract_solo_representation.py method=dino --config-name solo_cifar100

This will behave similarly to the previous call. The available representations are in prcut.data.solo_checkpoints

Training the models

Once the different representations are extracted, we can call the following to run PRCut:

python run_prcut.py dataset=fashionmnist repr=dinov2 root_dir=/buckets/ml/

or Turtle:

python run_turtle.py --dataset fashionmnist --phis dinov2 --root_dir /buckets/ml/