search

lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

Where is the test code? #2

Closed TyroneLi closed 3 years ago

TyroneLi commented 3 years ago

Where is the test code? Could u release this part. By the way, do u consider compute the reco loss across the whole dataset or the past few mini-batch instead of the current mini-batch.

lorenmt commented 3 years ago

Hello, the released code already includes testing and every component to reproduce the results in the paper. Could you specify which other implementation you need?

The reco loss is only computed by the current batch for simplicity. And it actually works the best compared to the whole dataset/history batch in my own experience.

TyroneLi commented 3 years ago

Hello, the released code already includes testing and every component to reproduce the results in the paper. Could you specify which other implementation you need?

The reco loss is only computed by the current batch for simplicity. And it actually works the best compared to the whole dataset/history batch in my own experience.

So you have already tried using the whole dataset/ history batch to compute reco loss? Why don't u just update this part code?

lorenmt commented 3 years ago

Hello, the released code already includes testing and every component to reproduce the results in the paper. Could you specify which other implementation you need? The reco loss is only computed by the current batch for simplicity. And it actually works the best compared to the whole dataset/history batch in my own experience.

So you have already tried using the whole dataset/ history batch to compute reco loss? Why don't u just update this part code?

As I said, the current implementation achieves higher performance compared to using the whole dataset.