Distributed training without GPU

DeepSpeech
#1

Hi,

I have 3 32GB high end VMs but no GPU. Is it possible to use them for distributed re-training(transfer learning) process.
I went through other similar post but was unable to understand much. Also saw run-cluster.sh which is provided, but feel that GPU is needed for it.
Could someone help by telling how to setup distributed training for improving results.

Thanks

#2

There’s no reason it should not, but you need to be aware that the speedup from a GPU is way more important that you can get on CPU, so even three high-end VMs will still be much slower.