Community

How to reduce the GPU memory load for lower GFX specs in cnnexample.py (ResourceExhaustedError OOM ) ?

(copy/paste from YouTube "Installing the GPU version of TensorFlow for making use of your CUDA GPU")

Hi Harrison,
Love your videos, really great to watch a true "step-by-step" tutorial on ML, very user-friendly.
FYI, I recovered a Nvidia 750 Ti 2go (current is RX 480 8go with FreeSync monitor, too bad AMD sucks at ML) and installed it today on Win 10. Everything works fine after install, did several runs on TF 1.0 + TFLearn examples.

But when I try to run your "cnnexample.py", I keep running into the following error after Epoch9 completed (so during the last Epoch 10, grrr...).

'''ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[10000,32,28,28]'''

Obviously the 750TI is "Out Of Mana/Memory", can't afford your Titan X Pascal 12Go :-(

Google searches didn't bring any working solution.
I tried to cut "batch_size=128" to 64 then 32, no success.
Any idea how to reduce the GPU memory load for lower GFX specs in your example ?

Cheers,