prepare.py

import pickle
import datasets
import os
import umap


if __name__ == "__main__":
    cache_file = "dataset_cache.pkl"
    if os.path.exists(cache_file):
        # Load dataset from cache
        with open(cache_file, "rb") as file:
            dataset = pickle.load(file)
        print("Dataset loaded from cache.")
    else:
        # Load dataset using datasets.load_dataset()
        ds = datasets.load_dataset("renumics/cifar10-outlier", split="train")
        print("Dataset loaded using datasets.load_dataset().")

        df = ds.rename_columns({"img": "image", "label": "labels"}).to_pandas()
        df["label_str"] = df["labels"].apply(lambda x: ds.features["label"].int2str(x))

        df = df.sample(10000, random_state=42).reset_index(drop=True)

        # precalculate umap embeddings
        df["embedding_ft_precalc"] = umap.UMAP(
            n_neighbors=70, min_dist=0.5, random_state=42
        ).fit_transform(df["embedding_ft"].tolist()).tolist()
        print("Umap for ft done")


        df["embedding_foundation_precalc"] = umap.UMAP(
            n_neighbors=70, min_dist=0.5, random_state=42
        ).fit_transform(df["embedding_foundation"].tolist()).tolist()

        print("Umap for base done")

        # Save dataset to cache
        with open(cache_file, "wb") as file:
            pickle.dump(df, file)

        print("Dataset saved to cache.")