Wrapper to handle image inputs both from local paths and urls Args: path_or_url (str): path or link to image.

Takes in a batch of images and returns a batch of embeddings of the same dimensions as the LM's word embeddings. :param config: Multimod

Dataset which loads image caption data from our standard format and transforms them into tensors that can be input to the model. Images are e

Randomly crops, then randomly resizes, then randomly crops again, an image. Mirroring the augmentations from https://arxiv.org/abs/2102.12092