Should these be padded out to be the left and right edges of the first and last segments so that the mid points of the pixels are the midpoint of the time bins?

@tacaswell, I'm not quite sure, if I understand what you mean by padding: The length in of the spectrogram is always shorter then the signal (by nfft-1 samples in my example). With the fix, the point in time is exactly at the midpoint of the segment (i.e., window function).

The extent kwarg sets where the edges of the image will be placed. If t is the midpoints, then this starts the first pixel at the mid point of the first time window and the last pixel ends at the midpoint of the last bin. I think the correct behavior should be that each pixel starts and stops on the edges of the time window represents (or is at least centered on the time window if there is overlap).

Ah, now I understand:


Never noticed that, because I usually try to maximize the overlap.

Current coverage is 62.07% (diff: 100%)

Merging #7692 into master will increase coverage by <.01%

@@             master      #7692   diff @@
==========================================
  Files           174        174          
  Lines         56021      56022     +1   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
+ Hits          34773      34776     +3   
+ Misses        21248      21246     -2   
  Partials          0          0          

Powered by Codecov. Last update ab98852...c3a8574

Padding Fixed:

Thanks, LGTM to me!

This is another case where we would have asked you to squash and re-base (to only have one set of the regenerated images in the repository history), but in general you do not have to squash down to a single commit.

Thanks @vollbier 🎉

I think this is your first contribution to mpl, hopefully there will be many more!

I hope so too.

@tacaswell added this to the 2.1 (next point release) milestone 4 days ago

#7666 says 2.0.1?

Bumped the issue to 2.1, I think this change is too big to put into 2.0 (at least on top of the rc). I am very open to be convinced otherwise.

@tacaswell I have no idea, how such changes disturb matplotlib's release cycles. From the user's point of view, I would tend to include it: I stumbled upon this issue while comparing spectrograms to Wigner-Ville distributions for certain signals. It lead me at first to the wrong conclusion, that spectrograms give biased estimates, which can be a quite serious roadblock for further investigations.

@vollbier sold.

backported as v2.x 6afa9ce