Andriy Mulyar

Samuel, Thank you for pointing this out. In the original datasets considered for evaluation, performance was calculated in the multi-class manner. I will be pushing out an update that will...

Yes - but it will of course use more memory. I'd suggest to try finetuning a pretrained longformer if your documents truly are that long.

Yes if you fix the number of chunks. On Mon, Nov 2, 2020, 10:27 AM Vipula Rawte wrote: > Hi, > > The papers describe four pooling functions: 1. Mean,...

Hi, the public codebase just hasn't been updated. You can change the pooling from max to mean in the implementation to replicate the stated results in the paper. Cheers

The code is up to the date with the correction. On Mon, Nov 16, 2020, 6:26 AM chokchou wrote: > hello, I noticed your new paper on arXiv submitted on...

1. Yes, separate for every chunk. 2. In our datasets we found it sufficient to fine-tune only the final transformer layer. On Fri, Oct 2, 2020, 11:12 AM Vipula Rawte...

Hi Arne, Thank you for your carefully laid out issue. What you tried seems reasonable. This repository contains code transferred from the original implemention - I will cross check with...

Arne, Thank you again for the continued detailed analysis. What you quoted indeed disagrees with our original implementation. This may very well be be an error in writing/interpretation on our...

Hi, these are label indicators.

Hello, I have also conducted some investigations in utilizing the Hellinger Distance as a DT impurity criterion. I have my own cython implementation incorporated into a fork of sklearn along...