It looks like pytorch by default keeping a lot of GPU memory and there isn't any memory available for the data to fit. I tried to train on two different Gpus, RTX 2080 Super (8 GB) and Titan Xp (12 GB) but same error. In case of RTX 2080 Super pytorch used to occupy around 6.5 GB of memory and in case of Titan Xp it occupy around 11 GB of memory. In conclusion I am not able to train the network for batch size > 1.
What could be the solution. There must be someway like it is in TF2.0 to limit the GPU memory used by the framework. Any help would be appreciated.
AloveDL
commented
4 years ago
Hi,
It looks like pytorch by default keeping a lot of GPU memory and there isn't any memory available for the data to fit. I tried to train on two different Gpus, RTX 2080 Super (8 GB) and Titan Xp (12 GB) but same error. In case of RTX 2080 Super pytorch used to occupy around 6.5 GB of memory and in case of Titan Xp it occupy around 11 GB of memory. In conclusion I am not able to train the network for batch size > 1.
What could be the solution. There must be someway like it is in TF2.0 to limit the GPU memory used by the framework. Any help would be appreciated.
Regards, Maaz