search

heng-hw / SpaCap3D

[IJCAI 2022] Spatiality-guided Transformer for 3D Dense Captioning on Point Clouds (official pytorch implementation)
https://spacap3d.github.io/
20 stars 5 forks source link

glove embedding not used in the code? #2

Closed ch3cook-fdu closed 2 years ago

ch3cook-fdu commented 2 years ago

I've notice that the Transformer Captioner does not use glove embedding as pretrained word embeddings. Would this hurt the captioner's performance?

heng-hw commented 2 years ago

Hi @ch3cook-fdu! Thanks for your interest! I've used the glove embedding at some point during the experiments but it did not improve the performance very much. So I did not use it in the later development of the model. You are welcome to add it to see if there is an improvement.

ch3cook-fdu commented 2 years ago

Thanks a lot! BTW, could you upload the training log? I found it slow to converge with CIDEr while training, I don't know if this is normal.

heng-hw commented 2 years ago

@ch3cook-fdu Yes, it did take some time to converge (the best CIDEr usually appears at epoch 30+). Please see this log.txt for reference. Hope it helps!

ch3cook-fdu commented 2 years ago

The cider performance has large fluctuation during training actually 🤔. Thx! That helps a lot!