How long did it take to train the model?

[karunk]

I've been wanting to experiment with the SVHN dataset myself and I loved your project. I wanted to run this on my computer. Do you think it can be run on a 8gb macbook air?

What were the specifications of the machine you ran on? How long did it take?

I'm asking this because, I've been reading that a GPU was necessary while dealing with SVHN..

I know this is not an issue, but I opened it because there is no other way to ask this question.

Thanks.

[thomalm]

I guess you would have to try to find out, but I'm guessing that it would be possible if you make some minor alterations to the code and scale the images to 32x32 and train and validate your network using small batches of e.g. 64 images. I would also suggest that you could experiment with smaller network architectures (fewer layers) to make it run on your machine.

I ran the project on a MacBook Pro 15" 2014 with 16GB of ram and i mostly ran into memory issues when dealing with larger images (64x64) and trying to feed the entire test-set at once into my network. In my final implementation i had to train the network for ~24 hours to get a full transcription accuracy of a little over 90%.

Optimally you would like to use one or more GPUs as you can build a much larger network and train it much faster. If you go through the trouble of learning how to use AWS and rent a GPU instance i'm guessing you could improve your results by a few %.