When training a sequence to sequence model in CNTK, what do the "samples per second" mean in the logs? Does it mean the number of tokens processed per second, or the number of sequences processed. Here's an example from the log:
01/24/2017 21:56:03: Epoch[ 1 of 50]-Minibatch[7701-7800]: ce = 3.13816951 * 93986; errs = 82.246% * 93986; time = 75.3008s; samplesPerSecond = 1248.1 01/24/2017 21:57:19: Epoch[ 1 of 50]-Minibatch[7801-7900]: ce = 3.13674664 * 94825; errs = 82.424% * 94825; time = 75.7230s; samplesPerSecond = 1252.3
In this case number of samples is the number of tokens. If however you had a single label per sequence it would be the number of sequences. So it is really the number of labels that is responsible for this number.