Metric reporting consistency #4
Description
Current scripts report in ms. Seems like as soon as we scale to multiple GPUs with weak and strong scaling that we likely should report in something like items/sec. (Current time reporting * batch size). Especially if people are going to be running different batchsizes on different sized memory. We should probably also print out batch size and iterations used.
Some scripts report total, some scripts report forward and backward differently. For example, the image+video tests report differently than the audio ones.
We are trying to line up the work we do for NV-Caffe support and get more consistent reporting so reporting tools work.