Is it possible to use both CPU and GPU to train convolution neural network (CNN)?

Hello, I have designed a CNN with 25 layers for medical image classification. The dataset consists of 10000 images of size 227X227X3. The PC has 8GB RAM and GPU (2GB). I am facing out of memory problem for the designed network but it works well for smaller networks. Is is possible to use CPU and GPU in parallel to solve the memory issue? Kindly suggest to solve the problem. I am using MATLAB version 2017b. Thank you

Answer 1

[Joss Knight]

Yes it is. Are you averse to reading documentation? If not, try:

• Setting the ExecutionEnvironment option to predict and classify or the ExecutionEnvironment input to trainingOptions.
• Lowering the MiniBatchSize option to predict and classify or the MiniBatchSize input to trainingOptions. This will decrease the load on your GPU memory.

[sotiraw sotiroglou]

He asks if BOTH cpu and gpu can be used together. Are you averse to reading documentation? Would you please show us this option that is for BOTH using at the same time cpu and gpu? Cause i really want to know

[Joss Knight]

Hiya. You've got to admit that the question is ambiguous at least!

There's no option to do some computation on the CPU and some on the GPU, although of course loads of things are in fact running on the CPU, like the whole of the MATLAB interpreter for instance. But there would be very little advantage in doing so anyway. Either you run synchronously, which means the GPU workers are waiting (an exceedingly long time) for the CPU workers to provide a result, or you run asynchronously, in which case the CPU workers have models that are getting very out of date with the rest of the cluster. So there isn't really a situation in which I expect the extra CPU workers to be doing anything useful.

[sotiraw sotiroglou]

the question seems clear, both in title and in context. Next time please take a minute to actually read it before you try to insult with your answers . this way you will also avoid embarass yourself by not understanding what people asks

[Joss Knight]

Well it wasn't clear to me and unless you are the original poster I suggest you not put words into their mouth. I certainly did not intend any offence, but I would happily apologise to the original poster if I was careless with my words.