turn VPT into a python package

[verbose-void]

refactor such that lib is now inside a parent folder vpt. also moved the demo scripts, like run_agent.py into vpt/. created a setup.py file, which you can run pip install -e vpt/ such that it's easier to use VPT from external scripts in this way:

from vpt.run_inverse_dynamics_model import get_IDM_agent

get_IDM_agent("path/to/idm.model", "path/to/idm.weights")

i know you mentioned that VPT is for demo purposes only and the code shouldn't be used elsewhere, but the model files and weight files are built so propriatarily that using them without copy pasting the code (for example IDM) is a reverse-engineering task.

instead, i figured it would be easier to turn the VPT code into a package.

i modified the README with updated information on how to run the files as well. there are no tests though, so i can't 100% guarentee everything works but i tested some of the experiments and am fairly sure it should be fine. if not it will be a trivial import fix.

[verbose-void]

i thought it was going to be easier to create the module and then write custom code but i'm definitely having trouble. i created another issue regarding "reverse engineering": https://github.com/openai/Video-Pre-Training/issues/19

[Miffyli]

Hey. Thanks for the PR!

On surface I do agree with you: it would be much easier to use this library for productive things ("building up on it") if it were a Python library. I have already had the same headache with the BASALT baselines which use this code.

However, I still leaning towards keeping this library as a non-module thing for few reasons:

• It is not meant to be "production grade". Granted, it does work, but it was never designed that in mind which can bite back later. A proper package would imply that this code can be used in "production", and with that comes the implicit assumptions of decent documentation and tests, both which lack currently.
• This is tied to a research publication, and while it does not have the full source code, it has some importants parts on what the model architecture is like (but again, granted, it does not have scripts which train the model or evaluate them). Generally for this reason I would like to keep it free of the program language overhead if possible, which this PR adds.
• There are no plans to maintain this code now that it has been shared, which could be required if this was shared as a Python library (especially if it is put on pypi).

So as a TL;DR, I do think making this repo a python package would be a good idea. Now, if someone makes a new repository which implements the same stuff but as a nice module/library, I am more than happy add a link to it in the readme :)

[verbose-void]

@Miffyli understood, thanks for the thorough response