v-iashin/LearnablePINs

model

Opened this issue · 1 comments

Tengfei-Wang commented

Hi, thx for sharing your code, and have you fixed the code?
and which of main.ipynb and dev.ipynb is your final result?

v-iashin commented

Hi, no I haven't. Moreover, I have no clue what is the problem behind it. In particular, the loss is reducing a bit during 1-2 epochs but after that, it remains mostly the same. Even when $\tau$ is increased it only jumps up but doesn't decrease with time.

Another evidence that I think can be the reason is the fact that the original paper describes the Hard Negative Mining algorithm which fails to mine negatives as it should.

I would recommend starting implementing this paper from a scratch. I have no appetite to continue exploring the problems because I spent a lot of time for this implementation already. Please, let me know if you will see some conceptual differences between your version and mine.

The dev.. is the development notebook with experiments. Please use main.. I guess it is self-explanatory.