Is it your idea to add batch normalization and replace upconv with bilinear upsample?

[fanfanda]

No description provided.

[milesial]

Hi, yes, the bilinear upsample is a way to have the model use less memory, even if the results are not as good. You can use upconv with bilinear=False in the net construction.
About the batch norm, I believe there was some discussion on the Kaggle forums and kernels about it, so I decided to add it.

[fanfanda]

Thanks for your reply. In my recent paper experiment, I retained two of your changes. Do I need to cite your project? I hope you can contact me by email. My email is (fanfanda@ict.ac.cn). Thank you again.

[milesial]

If you wish to cite me you can link the project URL, but I have to tell you that these changes are not backed by any theoretical study, I did it mainly because my GPU did not have enough memory :)
Make sure to follow the GNU GPL v3 license.

[fanfanda]

ok, thanks~