Hacky attempt at integrating this with FaRL LAION checkpoints

[evancasey]

Hi there! First of all thank you for the work on this awesome project.

I'm looking for an easy way to run the FaRL LAION checkpoints on a directory of images. The code structure in the main FaRL repo is not user friendly for inference uses so I tried to get it to work in here. However, since the LAION trained checkpoints are not JIT compiled I had to manually copy-paste the torch modules over and integrate this with the old inference code you had written in this repo.

My implementation is likely wrong because it does not output correct looking predictions. Maybe you can tell me what I'm doing wrong :) Or we can brainstorm a better solution to getting this repo working with the FaRL LAION checkpoints (maybe JIT compile them?). It could be very valuable for the research community!

https://gist.github.com/evancasey/eea749d7186e92670fca728ddb384212

Sample output:

[jyangliu]

Hello， Do you find the bug in your code implementation?