CUDA out of memory when running on single GPU about hydranet-wikisql

[lyuqin]

Hi José,
Yes, this is OOM error. Please change batch_size in wikisql.conf to smaller one. The current 256 setting is trained on my GPU cluster with 4x48GB memory, for 1x32GB I recommend to try batch_size=32 first.
Note that smaller batch size could result in lower accuracy, please refer to my reply here #9

[lyuqin]

Sorry, one correction:

1. If you specify --gpu 0, then it uses Nvidia RTX 2070 (8GB gpu memory?), so batch_size should be ~8.
2. If you don't specify --gpu, then it uses ram (32GB), so batch_size should be 32.
   If there is still OOM issue, please keep decreasing batch_size until no error

[JoseGuilhermeCR]

Thank you for your promptness in answering me!

I'm actually not quite sure about all of this, I'm just helping someone run on my computer because they thought it would work out of the box.

I'll try running with a small batch size, and I'd actually like to note that not specifying --gpu was not enough to make it use CPU and RAM, not as far as I noticed it.