The performance of pre-trained model about factorizablenet HOT 6 CLOSED

Hello, could you tell me how much memory did your GPU use while testing with the trained model?
I have got an error CUDA out of memory while testing.

from factorizablenet.

Hello, my gpus are titan xp which has 12Gb memory for each gpu.
The exp was run on single gpu.

from factorizablenet.

Thank you for your reply! :)
I have solved the problem by adding with torch.no_grad() before inference.
By the way, I got the same testing result as you.

from factorizablenet.

@shuni1001 , @Freya-Cui , I think you are not using "use_region" option in the model options file:
(VG-MSDN.yaml). It is set to false by default.
Since you are not using this external (phrases for regions) knowledge base, I think we are not getting the results as reported in README.

from factorizablenet.

Hi @sandeep-ipk we didn't use the external captions, like caption regions, during the training in this job.

The provided pre-trained model is trained using warm-up training. We trained the model for several epochs with learning rate decay, then reset the learning rate and train again. It will help to get a better result. I have provided the information in the last block of the "Training" section in README.

from factorizablenet.

@Freya-Cui Where did you add the 'with torch.no_grad()' ? Could you tell me the file name and the line number?

from factorizablenet.