Paper question: no use of validation set during U2Net training

[FraPochetti]

Hi,
in section 4.3 of the paper you write:

We train the network until the loss converges without using validation set

From that sentence, I understand that you wait for the training loss to flatten out (converge), without using any validation set to check for over/underfitting.
Am I getting this right?
Thanks and have a great one!

[xuebinqin]

Yes, that's true. Actually, we did use MSRA2.5k as the validation. But all the other test sets have different characteristics from MSRA2.5K. Good performance on MSRA2.5K won't guarantee better generalizations on different datasets. We then remove the validation set and train to converge. If you want, you can plot the -log(loss). The flat curve of -log(loss) should make more sense because the loss decreases very slight in the later training stage and -log(loss) is able to amplifying the decreasing trend, which provides you more recognizable changes in the loss.

…

On Mon, Apr 26, 2021 at 12:20 PM Francesco Pochetti < ***@***.***> wrote: Hi, in section 4.3 of the paper <https://arxiv.org/pdf/2005.09007.pdf> you write: We train the network until the loss converges without using validation set From that sentence, I understand that you wait for the training loss to flatten out (converge), without using any validation set to check for over/underfitting. Am I getting this right? Thanks and have a great one! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#194>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGOROIMOE4KLGX5QNH5D3TKUO5PANCNFSM43SLW72Q> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[FraPochetti]

Understood. Thanks!

[FraPochetti]

May I ask how big is the dataset you trained on?
I don't seem to find this info in the paper.

[xuebinqin]

Please refer to the Sec. 4.1 https://arxiv.org/pdf/2005.09007.pdf

…

On Mon, Apr 26, 2021 at 6:53 PM Francesco Pochetti ***@***.***> wrote: May I ask how big is the dataset you trained on? I don't seem to find this info in the paper. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#194 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORKP6RIX3GFESXT74TLTKV47DANCNFSM43SLW72Q> .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[FraPochetti]

Ouch sorry, that was hidden in plain sight!
Apologies for missing it and thanks