Cold PAWS: Unsupervised class discovery and addressing the cold-start problem for semi-supervised learning

Repository to reproduce label selection results

This is python code for the label selection strategies from the paper. For the model fitting code see this repo.

---

First, download the sample data from here. Then, to initialise t-SNE clusters and the data files, run

python main_setup.py

The next step is to select indices using the methods defined the config files. In this instance, 'finetune' is the mini-max approach and 'repulsive' is the maxi-min approach. To run these methods on CIFAR-10, for a budget of 40 labels, use

python main_results.py --config config/test.yaml --dataset data_processed/cifar10.pickle

Here are some further examples to generate the benchmarking runs, or the results for the imagenette dataset

python main_results.py --config config/benchmark.yaml --dataset data_processed/sw24708.pickle
python main_results.py --config config/base-extra-class-disc.yaml --dataset data_processed/imagenette.pickle

To take the output files from a particular run and put them all together in a CSV file, run

python unsupervised_class_detection.py --config config/benchmark.yaml --processed_data 'data_processed/sw24708.pickle'

To make t-SNE plots from the encoding files in data, run

python vis_clustering.py

The output folder contains some example outputs for reference which need to be downloaded seperately from the link above.