search

jerpelhan / DAVE

MIT License
40 stars 4 forks source link

About "base_3_shot.pth" #8

Closed Erica-Yang closed 2 months ago

Erica-Yang commented 4 months ago

Hi, how to generate "base_3_shot.pth"? or is it the same as "loca_few_shot.pth" in LOCA?

jerpelhan commented 4 months ago

We replaced the MAX operation with a softmax operation in LOCA (lines 101-109) for more efficient training. You have two options: either train/use the original LOCA version and modify lines 338-348 in dave.py to match the original implementation in loca.py, or retrain the loca with the softmax operation/use weights base_3_shot.pth. Either way, the performance should be similar.

Erica-Yang commented 4 months ago

Thanks a lot for your reply!