NVlabs / RVT

Official Code for RVT-2 and RVT
https://robotic-view-transformer-2.github.io/
Other
261 stars 31 forks source link

what is the minimum GPU mem required for training the RVT-2 model #47

Closed ruili3 closed 2 months ago

ruili3 commented 2 months ago

Hello,

Thanks for the nice work! I noticed that you use 8 V100 GPUs for model training. What are the minimum GPU resources to train the model based on the default setting (e.g., batch size)? I may not have as many as up to 8 V100, so I'd be grateful for your feedback on the computation cost. Thanks.

imankgoyal commented 2 months ago

Hello,

Sorry for the late reply. We used 16GB memory for each V100 (128GB in total). In general, RVT-2 should be the cheapest model to train for any computational resource. I would be very interested in what speed you observe on other GPU setups. Please make sure to scale the batch size given your GPU.

Best, Ankit