Open MichaelCurrie opened 9 years ago
Most are okay, though:
(Plots from HistogramManager.plot_information()
)
Cumulative invalid histograms in a 10-video sample (including features where some videos had valid histograms and some did not):
(i.e. for all 726 features to be invalid we'd have 7260 invalid histograms.)
Does this still happen? Are you using a cached version of the histograms? I fixed a bug about a weak ago in which one feature was no longer being computed. This is related to the issue I setup regarding the comparison being too lenient when merging nans #152
It's definitely not because of the cached (pickled) histograms in generate_stats.py
; I've definitely deleted my pickle file this past week.
I agree this is related, and as you say in that issue it means we may be fooling ourselves when we think we are agreeing with the Schafer code in the cases where we are generating all NaNs for the feature.
In investigating the histogram output on our sample of 10 Schafer feature files, it appears some features are never calculated.
This comes from the following lines of
generate_stats.py
: