Caffe performance on CPU

[Nikhil Sharma]

Hi Guys,

  I am trying to extract features from my data set which is a collection of 230,000 images. I am running this on my CPU and not in GPU mode. It is a windows machine and I am running it on a VMware virtualbox Ubuntu installation. Since the dataset is huge, it has been running since 24 hrs now. Could you please let me know if there is any way I could find out for how long it might go on running. Is it feasible to run such huge data set on just 4GB ram CPU memory. Should i switch to a faster computation platform? Any inputs would really help me.

Regards,

Nikhil. 

[Charles West]

Hey Nikhil,

I am pretty new, but I think that is a lot of images to try to train on with a CPU.  It might be worthwhile to switch to a GPU machine (preferably with a Solid State Drive).

Thanks,
Charlie

[Domenico Curro]

After 24 hours what iteration is it on?

An epoch is when the network has seen all of your data once, and it can be computer by #-of-images/batch-size.

You always at the very least want your network to have seen your data once. And if an epoch falls within a reasonable time, then I wouldn't fret.

GPU is always faster. An SSD will only help you speed up when the network is pre fetching all of the data (lots of disk reads).