git Feature/frame list creation

[uoefeb]

Description

This will output a list of commands, one per video, to run inference. run python build_inference_command.py -h for instructions on how to use this.

A few notes:

• The README mentions the model GIT_BASE_MSRVTT_QA, but doesn't talk about a non-QA one. I'm not sure if simply removing the prefix (as mentioned in the readme) is enough to convert it to a captioning task or not.
• 'type': 'test_git_inference_single_image' this also seems suspect, but it's what the README says to use.

TSV files vs many many calls to the program

The README mentions that a TSV file can be used to designate multiple images for inference tasks. This feels like it would be the best solution rather than calling this program thousands of times but...

• It's not clear training supports TSV files. I looked at the code a bit, and it appears like the answer might be no? This seems odd and impractical though
• It's also not clear if a TSV file can be used for inference for video captioning... it seems like it talks about it in terms of single image captioning. I have not had a chance to look at this code
• The TSV structure they describe would have us base64 encoding ALL of the images into tsv file... this would be huge... again, this seems impractical.

[uoefeb]

Added training command builder. Current implementation is based on the incorrect assumption that we can represent a list of videos as a list-of-lists-of-frames. As discussed over teams, this won't actually work, but hopefully this code will give us somewhere to start from when we know how to actually feed videos in.