Closed bmilde closed 7 years ago
great, thanks - will look at this ASAP. For now, it is not so much a concern for us because when we train with e.g. a 30ms frame shift, we also do 3-fold over-sampling (with offset of 0, 10, and 20 milliseconds), so that overall the duration reported after all the data has been processed is actually correct without this patch.
Only a minor addition, the reason I've added it is related to my setup of the 30ms frame shift. I've randomized the offsets, so that overall amount of training data stays the same, but the model will still learn from different offsets. But this results in a framerate of 33.3 without oversampling.
Confident these won't break anything, testing in progress
so that the corpus size estimates (in h) are correct in the log when a different frame rate than 100 is used in a setup.