lautaroest / sucpa-convergence

SUCPA Convergence

This is the code to analyze the convergence of the SUCPA algorithm.

Requirements

Check that you have the packages listed in requirements.txt installed. If not, run:

conda create --name sucpa python=3.10
conda activate sucpa
pip install -r requirements.txt

Usage

Running SUCPA on NLP datasets

To run the SUCPA algorithm for a specific value of $\beta^{[0]}$ (for instance, $\beta^{[0]}=\begin{bmatrix} 0.1 & 0.2 \end{bmatrix}$) run:

python run_sucpa.py \
    --dataset=sst2 \
    --steps=10 \
    --beta_init=0.1,0.2

You can also run the SUCPA algorithm for a range of values of $\beta^{[0]}$ by defining a number of repetitions and an initial seed:

python run_sucpa.py \
    --dataset=sst2 \
    --steps=10 \
    --repetitions=10 \
    --random_state=3892

In all cases, the results will be saved in the results/dataset={dataset} folder. Supported datasets for these experiments are sst2 (two classes) and mnli (three classes). Logits for these datasets are contained in the data directory and were obtained with the code in this repository.

Running SUCPA on image dataset

The procedure tu run SUCPA on image datasets is the same as for NLP datasets, so you just need to run (for instance):

python run_sucpa.py \
    --dataset=cat-dog \
    --steps=10 \
    --beta_init=0.1,0.2

The logits for the image datasets are contained in the data directory and were obtained fine-tuning a ResNet18 model on the Cat-Dog dataset of this Kaggle competition. To obtain the logits and labels contained in the data/dogs-vs-cats/ directory you can download the dataset, open it on the ./data directory and run the run_resnet.py script.

Results

Citation