Fix typo in QKVBimodalCSRPool

drprojects / DeepViewAgg

[CVPR'22 Best Paper Finalist] Official PyTorch implementation of the method presented in "Learning Multi-View Aggregation In the Wild for Large-Scale 3D Semantic Segmentation"

Other

222 stars 24 forks source link

Fix typo in QKVBimodalCSRPool #16

Closed altaykacan closed 1 year ago

altaykacan commented 1 year ago

Hello @drprojects ,

as I was going through the QKVBimodalCSRPool code I noticed that the code overwrites the keys when use_mod_q is true, then on line 506 queries is referenced without being defined which causes an error on my end. And as far as I understand the transform self.Q should be used for getting the query embeddings.

I thought this small change would fix the bug. Thank you once again for your work and making it available!

drprojects commented 1 year ago

Hi @altaykacan,

Good catch, thanks for the PR !

This QKVBimodalCSRPool was indeed intended for Query-Key-Value attention for the multi-view fusion. Results were not very satisfying in my experiments but maybe this (ugly :scream:) error could have played a role in this !

I'd be interested to hear if you saw an improvement between the GroupBimodalCSRPool and the QKVBimodalCSRPool in your own experiments :wink: