Is it possible to evaluate/finetune the model with batchsize > 1 case? about stereo-transformer HOT 3 CLOSED

Sounds good. Feel free to submit a pull request if you manage to fix it ;)

from stereo-transformer.

Hi @MinJunKang, yes you are right, I trained everything with batch size equal to 1. That's why I introduced the random cropping as an augmentation.

You should be able to run with larger batch by just removing random cropping or replacing with fixed size cropping here. However, you may run into bugs since I have never tried batch size larger than 1. If you see a bug let me know ;)

from stereo-transformer.

Hi! @mli0603 Thanks for your kind reply!

It seems that there is some issue running the model with batch size larger than 1 at positional encoder function and transformer function.

Now I will close this issue and I will let you know if there is some progress that can update these functions to be worked with more than batchsize 2 by opening new issue.

Again, thanks for your reply! It helps me a lot :)

from stereo-transformer.