Libtorch: A custom (e.g. NOT mnist) Image dataset in C++

dambo · June 14, 2019, 6:49am

Dear All,
I am trying to migrate the python based ImageList and PilImageList from here to Libtorch C++:

vita-epfl/openpifpaf/blob/master/openpifpaf/datasets.py

import copy
import logging
import os
import torch.utils.data
import torchvision
from PIL import Image

from . import transforms, utils


ANNOTATIONS_TRAIN = 'data-mscoco/annotations/person_keypoints_train2017.json'
ANNOTATIONS_VAL = 'data-mscoco/annotations/person_keypoints_val2017.json'
IMAGE_DIR_TRAIN = 'data-mscoco/images/train2017/'
IMAGE_DIR_VAL = 'data-mscoco/images/val2017/'


def collate_images_anns_meta(batch):
    images = torch.utils.data.dataloader.default_collate([b[0] for b in batch])
    anns = [b[1] for b in batch]
    metas = [b[2] for b in batch]

This file has been truncated. show original

Most of the examples I found in C++ are for MNIST and not a custom dataset.
Has anyone done something like this before? or is there a build in ImageFolder in C++?

Thanks,

dambo · June 15, 2019, 2:00pm

Anyone … breaking my head on this.

This is what I am looking for in C++:

class ImageList(torch.utils.data.Dataset):
    def __init__(self, image_paths, preprocess=None, image_transform=None):
        self.image_paths = image_paths
        self.image_transform = image_transform or transforms.image_transform
        self.preprocess = preprocess

    def __getitem__(self, index):
        image_path = self.image_paths[index]
        with open(image_path, 'rb') as f:
            image = Image.open(f).convert('RGB')

        if self.preprocess is not None:
            image = self.preprocess(image, [])[0]

        original_image = torchvision.transforms.functional.to_tensor(image)
        image = self.image_transform(image)

        return image_path, original_image, image

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

Thanks.