wrong histograms about tensorboardx HOT 6 CLOSED

Thanks. For now I guess I will just stick to 'doane' since I don't want to hand set the binning for each parameter. Actually, I am finding the distributions tab more useful than the histograms tab for my puproses anyway (basically I am interested in monitoring spiking gradients that require clipping and maybe also gradients going to zero). We can close this issue if you want. There is definitely some buggy behavior with the default binning, but if tensorboard is doing the binning by default I guess it is probably a tensorboard issue.

from tensorboardx.

http://tensorboard-pytorch.readthedocs.io/en/latest/tensorboard.html#tensorboardX.SummaryWriter.add_histogram

Hi, will additional parameter bins=range(...) help?

from tensorboardx.

I didn't try this exactly, but I tried some of the string values for bins listed in the numpy docs. Setting bins='auto' causes the program to crash with MemoryError coming from a call to np.zeros (even though the machine has plenty of memory). Setting bins='sturges' fixes the problem and leads to nice histograms. But it is pretty strange; I can't think of anything really unusual about my usage that would make the default and 'auto' histograms fail for me if they aren't broken for everyone else.

from tensorboardx.

Actually 'sturges' doesn't do that great of a job. It uses too few bins and doesn't zoom in enough. But at least it is not completely wrong like the default histograms.

from tensorboardx.

I also tried 'doane' and 'fd'. doane seems about the same as sturges, and fd crashes with a memory error like auto. Maybe the sturges/doane behavior is reasonable in that my distributions are very sharply peaked about zero..

from tensorboardx.

Using auto will explode the memory (should be numpy bug). see #1 . From you last graph, I think bins=np.arange(-0.003, 0.003, 0.0001) might do the work.

from tensorboardx.