pest-management-opendata.py

import functools as ft
import itertools as it
import collections as cl
from pathlib import Path
from dataclasses import dataclass, asdict
from urllib.parse import ParseResult, urlparse, urlunparse

import pandas as pd
from datasets import (
    Split,
    Image,
    Value,
    Features,
    Sequence,
    ClassLabel,
    DatasetInfo,
    SplitGenerator,
    GeneratorBasedBuilder,
)
from shapely.wkt import loads

__version__ = '20230327-1214'
__citation__ = '''
@inproceedings{iclr-2023,
 author = {White, Jerome and Agrawal, Chandan and Ojha, Anmol and
                  Agnihotri, Apoorv and Sharma, Makkunda and Doshi,
                  Jigar},
 title = {{BOLLWM}: A real-world dataset for bollworm pest monitoring
                  from cotton fields in {India}},
 booktitle = {International Conference on Learning Representations},
 year = {2023},
 series = {ICLR},
 publisher = {Workshop on Practical Machine Learning for Developing
                  Countries},
 doi = {10.48550/arXiv.2304.00763},
}
'''

SplitInfo = cl.namedtuple('SplitInfo', 'dtype, basename, split')
Payload = cl.namedtuple('Payload', 'source, target, df')

#
#
#
def readsp(path, split):
    return (pd
            .read_csv(path, compression='gzip')
            .query(f'split == "{split}"'))

#
#
#
@ft.singledispatch
def bucket2virtual(url):
    raise TypeError(type(url))

@bucket2virtual.register
def _(url: ParseResult):
    netloc = '.'.join((
        url.netloc,
        url.scheme,
        'ap-south-1',
        'amazonaws',
        'com',
    ))

    return urlunparse(url._replace(scheme='https', netloc=netloc))

@bucket2virtual.register
def _(url: str):
    return bucket2virtual(urlparse(url))

#
#
#
@dataclass
class SplitPayload:
    split: str
    metadata: Path
    images: dict

    def __iter__(self):
        df = readsp(self.metadata, self.split)
        for (i, g) in df.groupby('url', sort=False):
            source = urlparse(i)
            target = Path(self.images[i])

            yield Payload(source, target, g)

#
#
#
class SplitManager:
    _splits = tuple(it.starmap(SplitInfo, (
        (Split.TRAIN, 'dev', 'train'),
        (Split.VALIDATION, 'dev', 'val'),
        (Split.TEST, 'test', 'test'),
    )))

    @property
    def labels(self):
        path = bucket2virtual(self.metaname('dev'))
        df = pd.read_csv(path, compression='gzip')
        yield from df['label'].dropna().unique()

    def __init__(self, bucket):
        self.bucket = bucket
        self.path = Path('metadata', __version__)

    def __call__(self, dl_manager):
        for i in self._splits:
            name = self.metaname(i.basename)
            info = Path(dl_manager.download(bucket2virtual(name)))

            images = self.images(i.split, info)
            ipaths = dl_manager.download(dict(images))

            payload = SplitPayload(i.split, info, ipaths)
            yield SplitGenerator(name=i.dtype, gen_kwargs=asdict(payload))

    @staticmethod
    def images(split, info):
        df = readsp(info, split)
        for i in df['url'].unique():
            yield (i, bucket2virtual(i))

    def metaname(self, split):
        path = self.path.joinpath(split).with_suffix('.csv.gz')
        return self.bucket._replace(path=str(path))

#
#
#
class ExampleManager:
    # _Pest = cl.namedtuple('_Pest', 'label, geometry')
    # _Feature = cl.namedtuple('_Feature', 'image, pests')

    @staticmethod
    def features(labels):
        return Features({
            'image': Image(),
            'pests': Sequence({
                'label': ClassLabel(names=labels),
                'geometry': Value('binary'),
            }),
        })

    @staticmethod
    def pests(df):
        if 'geometry' in df.columns:
            for i in df.dropna().itertuples(index=False):
                geometry = loads(i.geometry)
                yield {
                    'label': i.label,
                    'geometry': geometry.wkb,
                }

    def __init__(self, payload):
        self.payload = payload

    def __iter__(self):
        for i in self.payload:
            key = urlunparse(i.source)
            with i.target.open('rb') as fp:
                raw = fp.read()
            value = {
                'image': {
                    'path': i.target,
                    'bytes': raw,
                },
                'pests': list(self.pests(i.df)),
            }

            yield (key, value)

#
#
#
class PestManagementOpendata(GeneratorBasedBuilder):
    _bucket = urlparse('s3://wadhwaniai-agri-opendata')

    def _info(self):
        data = SplitManager(self._bucket)
        labels = sorted(data.labels)
        features = ExampleManager.features(labels)

        return DatasetInfo(
            homepage='https://github.com/WadhwaniAI/pest-management-opendata',
            citation=__citation__,
            # description=_DESCRIPTION,
            license='CC-BY 4.0',
            features=features,
        )

    def _split_generators(self, dl_manager):
        splits = SplitManager(self._bucket)
        return list(splits(dl_manager))

    def _generate_examples(self, **kwargs):
        payload = SplitPayload(**kwargs)
        examples = ExampleManager(payload)
        yield from examples