Works like a charm!
I used Bitfusion's g2.2x Ubuntu 14 TensorFlow image and it worked great out of the box... I also found the trial period very useful, since I had a problem similar to cifar10 and during the trial period I was able to make sure that this image pulls off cifar10 easily. All I had to run was a single line of command and it started training a cifar10 model, using the GPU. I was able to witness 8x speedup when compared to my weak CPU. Yesterday I came back for my actual problem with my dataset size being nearly 8 times greater than that of cifar10. With some minor changes to cifar10 codebase, I was able to train a model for my problem successfully. I think I can wait for some more time before buying myself a physical GPU. Infrastructure-as-Service rocks!
There are no comments to display