Training modern deepnets can take an inordinate amount of time even with the best GPU hardware available. Inception-3 on ImageNet 1000 using 8 NVIDIA Tesla K40s takes about 2 weeks (Google Research Blog).

Even when a large network is trained successfully, the memory footprint and the prediction latency (due to the number of its parameters) can make it challenging to put it into production.

One way to keep…

Read More →