search

salesforce / BiST

Code for the paper BiST: Bi-directional Spatio-Temporal Reasoning for Video-Grounded Dialogues (EMNLP20)
11 stars 5 forks source link

Questions about GPU resources #1

Closed yaoshentao closed 3 years ago

yaoshentao commented 3 years ago

Hello, Thank you for your open source code and your interesting work！There were some problems in the process of using it. When I execute code on the Action task in your branch tgif-qa, I get a cuda out of memory error. I have modified the batchsize from 32 to 16, but the same situation still occurs. For batchsize=16, this error occurred when epoch=40. I would like to know what the size of the GPU memory you are using is and the error I get are normal? My GPU memory is 12GB.I lost my error information. If you need this error information, I can run it again and give you a screenshot.

henryhungle commented 3 years ago

Hi @yaoshentao I can run the code on the Action task (batch size=16) on a single GPU, Tesla V100-SXM2-16GB. With batch size 32, I also had Cuda out of memory error.

Based on my experience, on Action and Transition tasks, you probably should use a small batch size to avoid memory issues (smaller than the batch size on FrameQA and Count tasks).

yaoshentao commented 3 years ago

Hi, Thanks for your suggestion!