dataset preparation

[SreenijaK]

Can anyone clearly explain how do we prepare the dataset. I've refered many issues but dataset preparation is still unclear.

[jszgz]

Do you know what does the second dimension in embeddings represent? Why do we pick a randint embedding_ix in the second dimension?

        key = self.filenames[index]
        if self.bbox is not None:
            bbox = self.bbox[key]
            data_dir = '%s/CUB_200_2011' % self.data_dir
        else:
            bbox = None
            data_dir = self.data_dir
        # captions = self.captions[key]
        embeddings = self.embeddings[index, :, :]
        img_name = '%s/images/%s.jpg' % (data_dir, key)
        imgs = get_imgs(img_name, self.imsize,
                        bbox, self.transform, normalize=self.norm)

        wrong_ix = random.randint(0, len(self.filenames) - 1)
        if(self.class_id[index] == self.class_id[wrong_ix]):
            wrong_ix = random.randint(0, len(self.filenames) - 1)
        wrong_key = self.filenames[wrong_ix]
        if self.bbox is not None:
            wrong_bbox = self.bbox[wrong_key]
        else:
            wrong_bbox = None
        wrong_img_name = '%s/images/%s.jpg' % \
            (data_dir, wrong_key)
        wrong_imgs = get_imgs(wrong_img_name, self.imsize,
                              wrong_bbox, self.transform, normalize=self.norm)

        embedding_ix = random.randint(0, embeddings.shape[0] - 1)
        embedding = embeddings[embedding_ix, :]
        if self.target_transform is not None:
            embedding = self.target_transform(embedding)

        return imgs, wrong_imgs, embedding, key  # captions```