possible issue with rewards to go? about faster-trajectory-transformer

[Howuhh]

Also, @grahamannett, if you use gpt in your code, pay attention to the existing bug from #4 which is not merged yet.

[grahamannett]

@Howuhh damn, i originally had forked your repo and am using it too much/in process of running a fair amount of experiments and dont want to mess with what i have. seems like i cant make a new fork until i detach my current project from yours and requires submitting a github support ticket. i kicked off the runs that you have listed with the fix though and will PR once I get the fork issue sorted out. thanks a lot for the codebase, was very helpful!

[Howuhh]

Hi @grahamannett! Indeed, seems like this is a bug from incorrect numpy broadcasting which I missed unfortunately 😞. You can correct it with .squeeze(), .T or just [t + 1:, 0]. It's amazing that it trained so nicely in spite of this!

[grahamannett]

thanks for the heads up. its probably worth offering @Howuhh would you like me to retrain the models for the checkpoints with that fix? I am using the models to benchmark against anyway and will probably retrain them since i did not have that pre-norm fix.

[Howuhh]

@grahamannett yeah, that would be really cool, as I'm mostly busy with other things that have nothing to do with the TT at the moment. You can make a separate PR for this (or whatever you feel more comfortable with)

[Howuhh]

@grahamannett It's okay, do as you please. My project for which I wrote this code failed, glad the code is useful to someone anyway