import datasets import numpy as np import os import glob from tqdm import tqdm # Optional for progress _DESCRIPTION = "EEG dataset with X and y numpy arrays." _CITATION = "" # Add citation if available _HOMEPAGE = "https://huggingface.co/datasets/prastab/eeg-processed-train-dataset" # Or your project page _LICENSE = "" # Specify license, e.g., 'apache-2.0' class EEGProcessedDataset(datasets.GeneratorBasedBuilder): """EEG Dataset Builder.""" VERSION = datasets.Version("1.0.0") def _info(self): # Define the features of your dataset # IMPORTANT: Specify correct shapes and dtypes! features = datasets.Features( { "X": datasets.Array3D(shape=(256, 62), dtype="float32"), # Adjust shape! (seq_len, channels) "y": datasets.Array3D(shape=(64, 62), dtype="float32"), # Adjust shape! (out_seq_len, channels) } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def _split_generators(self, dl_manager): # Data is already in the repo, so we just need paths relative to the repo root repo_path = os.path.dirname(__file__) # Gets the directory of this script within the repo return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={"split_dir": os.path.join(repo_path, "train")}, ), datasets.SplitGenerator( name=datasets.Split.VALIDATION, # or 'val' if you prefer gen_kwargs={"split_dir": os.path.join(repo_path, "val")}, ), datasets.SplitGenerator( name=datasets.Split.TEST, gen_kwargs={"split_dir": os.path.join(repo_path, "test")}, ), ] def _generate_examples(self, split_dir): """Yields examples.""" x_files = sorted(glob.glob(os.path.join(split_dir, "*_X_*.npy"))) y_files = sorted(glob.glob(os.path.join(split_dir, "*_y_*.npy"))) if len(x_files) != len(y_files): raise ValueError(f"Mismatch in X/Y file count for {split_dir}") if not x_files: raise FileNotFoundError(f"No data files found in {split_dir}") global_idx = 0 for xf, yf in tqdm(zip(x_files, y_files), total=len(x_files), desc=f"Processing {os.path.basename(split_dir)}"): # Basic check for pairing x_chunk_id = os.path.basename(xf).split('_X_')[-1] y_chunk_id = os.path.basename(yf).split('_y_')[-1] if x_chunk_id != y_chunk_id: print(f"Warning: Potential file mismatch? {xf} vs {yf}") # Add better checking if needed x_data = np.load(xf) # Loads (chunk_size, seq_len, channels) y_data = np.load(yf) # Loads (chunk_size, out_seq_len, channels) if x_data.shape[0] != y_data.shape[0]: raise ValueError(f"Sample count mismatch in chunk {xf}/{yf}") # Iterate through samples within the loaded chunk for i in range(x_data.shape[0]): yield global_idx, { "X": x_data[i].astype(np.float32), # Ensure correct dtype "y": y_data[i].astype(np.float32), # Ensure correct dtype } global_idx += 1