scripts/dataset/dataset.py

# Copyright (c) Facebook, Inc. and its affiliates.

import glob
from itertools import chain
from os import path

import torch
from PIL import Image
from torch.utils.data import Dataset


class SegmentationDataset(Dataset):
    _EXTENSIONS = ["*.jpg", "*.jpeg", "*.png"]

    def __init__(self, in_dir, transform):
        super(SegmentationDataset, self).__init__()

        self.in_dir = in_dir
        self.transform = transform

        # Find all images
        self.images = []
        for img_path in chain(
            *(
                glob.iglob(path.join(self.in_dir, ext))
                for ext in SegmentationDataset._EXTENSIONS
            )
        ):
            _, name_with_ext = path.split(img_path)
            idx, _ = path.splitext(name_with_ext)
            self.images.append({"idx": idx, "path": img_path})

    def __len__(self):
        return len(self.images)

    def __getitem__(self, item):
        # Load image
        with Image.open(self.images[item]["path"]) as img_raw:
            size = img_raw.size
            img = self.transform(img_raw.convert(mode="RGB"))

        return {"img": img, "meta": {"idx": self.images[item]["idx"], "size": size}}


def segmentation_collate(items):
    imgs = torch.stack([item["img"] for item in items])
    metas = [item["meta"] for item in items]

    return {"img": imgs, "meta": metas}