andrea.pedrotti
/
gfun_multimodal
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

todo update

This commit is contained in:
Andrea Pedrotti 2023-03-02 18:20:43 +01:00
parent 0c9454cdd4
commit 25fd67865d
1 changed files with 3 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
main.py
View File

@ -13,16 +13,13 @@ from gfun.generalizedFunnelling import GeneralizedFunnelling
""" """
TODO: TODO:
- [!] logging
- add documentations sphinx - add documentations sphinx
- zero-shot setup - [!] zero-shot setup
- load pre-trained VGFs while retaining ability to train new ones (self.fitted = True in loaded? or smt like that)
- test split in MultiNews dataset
- when we load a model and change its config (eg change the agg func, re-train meta), we should store this model as a new one (save it)
- FFNN posterior-probabilities' dependent - FFNN posterior-probabilities' dependent
- re-init langs when loading VGFs? - re-init langs when loading VGFs?
- there is a mess about sigmoid in the Attention aggregator + and evaluation function (predict). We were applying sig() 2 times on the outputs (at pred and at eval)...
- [!] loss of Attention-aggregator seems to be uncorrelated with Macro-F1 on the validation set! - [!] loss of Attention-aggregator seems to be uncorrelated with Macro-F1 on the validation set!
- aligner layer (suggestion by G.Puccetti) - [!] experiment with weight init of Attention-aggregator
""" """