search

leokarlin / LaSO

LaSO: Label-Set Operations networks for multi-label few-shot learning - official implementation
BSD 3-Clause "New" or "Revised" License
86 stars 14 forks source link

About 1-shot and 5-shot (from more detailed implementation) #14

Closed zhangyikaii closed 3 years ago

zhangyikaii commented 3 years ago

Hi. Thank you for sharing the code, It's wonderful.

The difference between 1-shot and 5-shot setting's sampling is only here. Why in the case of 5-shot, divide by 5 in each batch of fakeVectorsPairs?

(It doesn't matter that your implementation is different from the Episodic Sampling in Few-Shot classical methods like Prototypical Networks, but 5-shot means five samples per class. So how to explain division 5 here?)

Thank you again and I am looking forward to your reply.

alfassy commented 3 years ago

Hi, We don't work in episodes here, we train with our 5 shot labeled samples, then we just check performance on the entire validation set, if I recall correctly I changes the batch size in the 5 shot case because of CUDA memory limitations, I was using K40 gpu at that time. Regards, Amit

zhangyikaii commented 3 years ago

Thank you very much!