plantorgans.py

import datasets
import pandas as pd
import glob
from pathlib import Path
from PIL import Image, ImageOps

_DESCRIPTION = """Photos of various plants with their major, above ground organs labeled. Includes labels for stem, leafs, fruits and flowers."""

_HOMEPAGE = "https://huggingface.co/datasets/jpodivin/plantorgans"

_CITATION = """"""

_LICENSE = "MIT"

_BASE_URL = "https://huggingface.co/datasets/jpodivin/plantorgans/resolve/main/"
_TRAIN_URLS = [_BASE_URL + f"sourcedata_labeled.tar.{i:02}" for i in range(0, 8)]
_TEST_URLS = [_BASE_URL + f"sourcedata_labeled.tar.{i:02}" for i in range(8, 12)]
_MASKS_URLS = [_BASE_URL + f"masks.tar.0{i}" for i in range(0, 2)]
_SEMANTIC_MASKS_URLS = [_BASE_URL + f"semantic_masks.tar.0{i}" for i in range(0, 2)]

_SEMANTIC_METADATA_URLS = {
    'train': 'https://huggingface.co/datasets/jpodivin/plantorgans/resolve/main/metadata_semantic_train.csv',
    'test': 'https://huggingface.co/datasets/jpodivin/plantorgans/resolve/main/metadata_semantic_test.csv'
}

_PANOPTIC_METADATA_URLS = {
    'train': 'https://huggingface.co/datasets/jpodivin/plantorgans/resolve/main/metadata_train.csv',
    'test': 'https://huggingface.co/datasets/jpodivin/plantorgans/resolve/main/metadata_test.csv'
}

class PlantOrgansConfig(datasets.BuilderConfig):
    """Builder Config for PlantOrgans"""
 
    def __init__(self, data_urls, metadata_urls, splits, **kwargs):
        """BuilderConfig for PlantOrgans.
        Args:
          data_urls: list of `string`s, urls to download the zip files from.
          metadata_urls: dictionary with keys 'train' and 'validation' containing the archive metadata URLs
          **kwargs: keyword arguments forwarded to super.
        """
        super().__init__(version=datasets.Version("1.0.0"), **kwargs)
        self.data_urls = data_urls
        self.metadata_urls = metadata_urls
        self.splits = splits


class PlantOrgans(datasets.GeneratorBasedBuilder):
    """Plantorgans dataset
    """
    BUILDER_CONFIGS = [
        PlantOrgansConfig(
            name="semantic_segmentation_full",
            description="This configuration contains segmentation masks.",
            data_urls=_BASE_URL,
            metadata_urls=_SEMANTIC_METADATA_URLS,
            splits=['train', 'test'],
        ),
        PlantOrgansConfig(
            name="instance_segmentation_full",
            description="This configuration contains segmentation masks.",
            data_urls=_BASE_URL,
            metadata_urls=_PANOPTIC_METADATA_URLS,
            splits=['train', 'test'],
        ),
    ]

    def _info(self):
        features=datasets.Features(
                {
                    "image": datasets.Image(),
                    "mask": datasets.Image(),
                    "image_name": datasets.Value(dtype="string"),
                    "class": datasets.ClassLabel(
                        names=['Fruit', 'Leaf', 'Flower', 'Stem']),
                })
        if self.config.name == 'instance_segmentation_full':
            features['score'] = datasets.Value(dtype="double")
        else:
            features['class'] = datasets.ClassLabel(
                    names=['Fruit', 'Leaf', 'Flower', 'Stem'])
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            supervised_keys=("image", "mask"),
            homepage=_HOMEPAGE,
            citation=_CITATION,
            license=_LICENSE,
        )


    def _split_generators(self, dl_manager):
        
        train_archives_paths = dl_manager.download_and_extract(_TRAIN_URLS)
        test_archives_paths = dl_manager.download_and_extract(_TEST_URLS)
      
        train_paths = []
        test_paths = []

        for p in train_archives_paths:
            train_paths.extend(glob.glob(str(p)+'/sourcedata/labeled/**.jpg'))
        for p in test_archives_paths:
            test_paths.extend(glob.glob(str(p)+'/sourcedata/labeled/**.jpg'))
        
        if self.config.name == 'instance_segmentation_full':
            metadata_urls = _PANOPTIC_METADATA_URLS
            mask_urls = _MASKS_URLS
            mask_glob = '/_masks/**.png'
        else:
            metadata_urls = _SEMANTIC_METADATA_URLS
            mask_urls = _SEMANTIC_MASKS_URLS
            mask_glob = '/semantic_masks/**.png'

        split_metadata_paths = dl_manager.download(metadata_urls)

        mask_archives_paths = dl_manager.download_and_extract(mask_urls)
            
        mask_paths = []
        for p in mask_archives_paths:
            mask_paths.extend(glob.glob(str(p)+mask_glob))

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "images": train_paths,
                    "metadata_path": split_metadata_paths["train"],
                    "masks_path": mask_paths,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "images": test_paths,
                    "metadata_path": split_metadata_paths["test"],
                    "masks_path": mask_paths,
                },
            ),
        ]


    def _generate_examples(self, images, metadata_path, masks_path):
        """
        images: path to image directory
        metadata_path: path to metadata csv
        masks_path: path to masks
        """

        # Get local image paths
        image_paths = pd.DataFrame(
            [(str(Path(*Path(e).parts[-3:])), e) for e in images], columns=['image', 'image_path'])

        # Get local mask paths
        masks_paths = pd.DataFrame(
            [(str(Path(*Path(e).parts[-2:])), e) for e in masks_path], columns=['mask', 'mask_path'])
        
        # Get all common about images and masks from csv
        metadata = pd.read_csv(metadata_path)

        # Merge dataframes
        metadata = metadata.merge(masks_paths, on='mask', how='inner')
        metadata = metadata.merge(image_paths, on='image', how='inner')

        # Make examples and yield
        for i, r in metadata.iterrows():
            # Example contains paths to mask, source image, certainty of label,
            # and name of source image.
            example = {
                'mask': r['mask_path'],
                'image': r['image_path'],
                'image_name': Path(r['image_path']).parts[-1],
                'class': r['class']
            }
            if self.config.name == 'instance_segmentation_full':
                example['score'] = r['score']
            else:
                example['class'] = r['class']
            yield i, example