Download datasets:

Download and decompress tsv file from here: https://github.com/google-research-datasets/wit/blob/main/DATA.md
Use prepare_wit.py to download images from Wikipedia as annotated on each TSV file.
Use scale_converter.py to remove corrupt images and resize suitable images to 224x224
Use join_datasets_custom_split.py to group all JSONs from different subsets of the dataset together
Use discard_incorrect_files.py to filter out images that we were not able to convert.
Finally, use run-clip.sh to train.