search

GfellerLab / MixTCRpred

Predictor of TCR-epitope interactions
Other
16 stars 5 forks source link

code question #1

Closed phbradley closed 9 months ago

phbradley commented 9 months ago

Hey there,

Thanks for making this nice code available! I've just started poking around, and I know next to nothing about AI/transformers, but I'm wondering about this line:

https://github.com/GfellerLab/MixTCRpred/blob/main/src/models.py#L40

self.embedding_pos_TRB = PositionWiseEmbedding(self.vocab_size, self.embedding_dim, self.padding[1] + 2*self.padding[-1])

and whether it should be self.padding[2] rather than self.padding[1]. Of course, they are the same right now (20), so it doesn't matter! But maybe there could be unexpected behavior if one wanted to make cdr3a padding less (or cdr3b greater) at some future time? Also just trying to see if I understand what's going on.

Thanks, Phil

GiancarloCroce commented 9 months ago

Hi Phil,

thanks for your interest in our work. You're correct; it should be self.padding[2] (padding for the beta cdr3 chain). As you have noticed, in the GitHub version (for making predictions using the pre-trained MixTCRpred models) we set padding cdr3 beta = padding cdr3 alpha = 20 (line 147 in MixTCRpred.py), but thank you very much for pointing that out!

Best, Giancarlo

phbradley commented 9 months ago

Great, thanks Giancarlo for the speedy reply!