search

wellcometrust / grants_tagger

Tag grants with MeSH and other tags
MIT License
14 stars 4 forks source link

Paper experiments 📜 #188

Closed nsorros closed 1 year ago

nsorros commented 2 years ago

Message to @aCampello few weeks ago

Competition
* more years
* full text
* larger hidden size
* optimize threshold
* ensemble

Paper
* without multilabel attention
* xtransformers
* bilstm / cnn

Awaiting clarification

aCampello commented 2 years ago

Thanks for the issue, @nsorros. Whatever experiment we think is feasible would be appreciated. It would be really ace to get XTransformers working but this seems to have become a quite hard task, and the model is really resource consuming.

I think I'd add to that issue some non-coding tasks:

nsorros commented 2 years ago

I think most are feasible. Can we define feasible thought @aCampello just to ensure we are on the same page?

aCampello commented 2 years ago

By feasible I mean we'll have time to do it? Are we proposing to split the experiments among people mentioned in this issue ?

nsorros commented 2 years ago

By feasible I mean we'll have time to do it? Are we proposing to split the experiments among people mentioned in this issue ?

I would estimate that it takes about 1 week to run depending on the experiment and 1 day to develop max. Some might take less. It does not need much involvement from others but they could jump in if needed.