import io from PIL import Image from datasets import GeneratorBasedBuilder, DatasetInfo, Features, SplitGenerator, Value, Array2D, Split import datasets import numpy as np import h5py from huggingface_hub import HfFileSystem import time import os import hashlib from pathlib import Path class CustomConfig(datasets.BuilderConfig): def __init__(self, extract, remove_tar, chunk, **kwargs): super().__init__(**kwargs) self.dataset_type = kwargs.pop("name", "all") self.extract = kwargs.pop("extract", extract) self.remove_tar = kwargs.pop("remove_tar", remove_tar) self.chunk = kwargs.pop("chunk", chunk) _metadata_urls = { "train":"https://huggingface.co/datasets/XingjianLi/tomatotest/resolve/main/train.txt", "val":"https://huggingface.co/datasets/XingjianLi/tomatotest/resolve/main/val.txt" } class RGBSemanticDepthDataset(GeneratorBasedBuilder): BUILDER_CONFIGS = [ CustomConfig(name="full", version="1.0.0", description="download and extract the dataset to h5 pairs (all tar files automatically removed by default, 160GB)", extract=True, remove_tar=True, chunk=1), CustomConfig(name="sample", version="1.0.0", description="load both segmentation and depth (for 1 tar file, 870MB)", extract=False, remove_tar=False, chunk=1), CustomConfig(name="depth", version="1.0.0", description="only load depth (sample)", extract=False, remove_tar=False, chunk=1), CustomConfig(name="seg", version="1.0.0", description="only load segmentation (sample)", extract=False, remove_tar=False, chunk=1), ] # Configs initialization BUILDER_CONFIG_CLASS = CustomConfig def _info(self): return DatasetInfo( features=Features({ "left_rgb": datasets.Image(), "right_rgb": datasets.Image(), "left_semantic": datasets.Image(), "left_instance": datasets.Image(), "left_depth": datasets.Image(), "right_depth": datasets.Image(), }) ) def _h5_loader(self, fileobj, type_dataset): # Reference: https://github.com/dwofk/fast-depth/blob/master/dataloaders/dataloader.py#L8-L13 file_bytes = fileobj.read() file_bytes = io.BytesIO(file_bytes) with h5py.File(file_bytes, "r") as h5f: left_rgb = self._read_jpg(h5f['rgb_left'][:]) if type_dataset == 'depth': right_rgb = self._read_jpg(h5f['rgb_right'][:]) left_depth = h5f['depth_left'][:].astype(np.float32) right_depth = h5f['depth_right'][:].astype(np.float32) return left_rgb, right_rgb, np.zeros((1,1)), np.zeros((1,1)), left_depth, right_depth elif type_dataset == 'seg': seg_left = h5f['seg_left'][:] left_semantic = seg_left[:,:,2] left_instance = seg_left[:,:,0] + seg_left[:,:,1] * 256 return left_rgb, np.zeros((1,1)), left_semantic, left_instance, np.zeros((1,1)), np.zeros((1,1)) else: right_rgb = self._read_jpg(h5f['rgb_right'][:]) seg_left = h5f['seg_left'][:] left_semantic = seg_left[:,:,2] left_instance = seg_left[:,:,0] + seg_left[:,:,1] * 256 left_depth = h5f['depth_left'][:].astype(np.float32) right_depth = h5f['depth_right'][:].astype(np.float32) return left_rgb, right_rgb, left_semantic, left_instance, left_depth, right_depth def _read_jpg(self, bytes_stream): img = Image.open(io.BytesIO(bytes_stream)) return img def _split_generators(self, dl_manager): if 'full' == self.config.dataset_type: dataset_names = self._get_dataset_filenames() if self.config.extract: dataset_chunk_list = [dataset_names[i:i+self.config.chunk] for i in range(0, len(dataset_names), self.config.chunk)] for dataset_chunk in dataset_chunk_list: print(self.config.chunk) print(dataset_chunk) archives = dl_manager.download({"train":dataset_chunk, "val":dataset_chunk}) for archive, tar_file_name in zip(archives["train"], dataset_chunk): extracted_dir = dl_manager.extract(archive) print(f"\tExtracted {archive} to {extracted_dir}") if self.config.remove_tar and os.path.exists(archive): os.remove(archive) print(f"\tDeleted tar file {archive}") blob_folder = '/'.join(archive.replace("snapshots","blobs").split('/')[:-3]) if self.config.remove_tar and os.path.exists(blob_folder): for filename in os.listdir(blob_folder): filepath = os.path.join(blob_folder, filename) os.remove(filepath) print(f"\tDeleted tar file {blob_folder}") print("All extracted. exiting") exit() archives = dl_manager.download({"train":self._get_dataset_filenames(), "val":self._get_dataset_filenames()}) else: archives = dl_manager.download({"train":self._get_dataset_filenames()[0:2], "val":self._get_dataset_filenames()[0:2]}) split_metadata = dl_manager.download(_metadata_urls) train_archives = [dl_manager.iter_archive(archive) for archive in archives["train"]] val_archives = [dl_manager.iter_archive(archive) for archive in archives["val"]] return [ SplitGenerator( name=Split.TRAIN, gen_kwargs={ "archives": train_archives, "split_txt": split_metadata["train"] }, ), SplitGenerator( name=Split.VALIDATION, gen_kwargs={ "archives": val_archives, "split_txt": split_metadata["val"] }, ), ] def _generate_examples(self, archives, split_txt): with open(split_txt, encoding="utf-8") as split_f: all_splits = set(split_f.read().split('\n')) for archive in archives: for path, file in archive: archive_start_time = time.time() if path.split('/')[-1][:-3] not in all_splits: # skip the image pairs not in train.txt or val.txt continue left_rgb, right_rgb, left_semantic, left_instance, left_depth, right_depth = self._h5_loader(file, self.config.dataset_type) yield path, { "left_rgb": left_rgb, "right_rgb": right_rgb, "left_semantic": left_semantic, "left_instance": left_instance, "left_depth": left_depth, "right_depth": right_depth, } def _get_dataset_filenames(self): fs = HfFileSystem() all_files = fs.ls("datasets/xingjianli/tomatotest/data") filenames = sorted(['/'.join(f['name'].split('/')[-2:]) for f in all_files]) return filenames