Understanding histogram summary plot in Tensorboard

[jao...@gmail.com]

Hi all,

How to read the histogram summary plot in tensorboard ? For example, in the cifar10 tutorial, it plots the histogram of the convolution weight and biase but I don't understand the meaning of all these lines. Especially, what is meaning of the Y-axis of the plot ?

Jao

[Dan Mané]

Hi Jao,

Great question, it's not well documented at the moment :) Right now, the way we visualize histograms is by showing where different %iles are distributed in the data. So the darkest line in the center shows the median of the data, the next line up shows the 68th percentile, the next line up shows the 95th percentile, the next line up shows the 99.7th percentile, and the final line (with the very faint shading) shows the maximum of the data (the 100th percentile).

We're planning to add a crosshair display that will annotate this so it's easier to understand. In the medium term, we want to shift away from showing percentile summary statistics, and show the actual probability density function instead. The percentile approach has a few disadvantages, for example if your data distribution is bimodal it does a bad job visualizing it.

Dan

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/070537be-4b2d-4c71-b985-b778874a7c0c%40tensorflow.org.

[qin...@gmail.com]

I am sorry that I still don't understand what you say, is there a documnet now. I can't find it, so you can tell me more about it?

在 2015年11月12日星期四 UTC+8上午11:08:16，danmane写道：

[Jaonary Rabarisoa]

Hi,

This is now explained in the tensorboard readme https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tensorboard

Jao

[qin...@gmail.com]

thanks!!!

在 2016年5月19日星期四 UTC+8下午3:07:57，Jaonary Rabarisoa写道：

[Charlos Parker]

Is it possible to request a picture example of what your talking about? What does "the darkest line" mean? What plots are you referring to?

[lethi...@gmail.com]

Hi,

I have another question about the quantile "histogram". I see that we can plot almost everything (W, dW, b, db,...).

But what is in it actually ? Knowing that b is a vector, W is a matrix. I assumed that Tensorflow has calculated the norm (L2 et Frobenius) of these tensors to view their evolution through time, right ?

Of course, a probability density function is more than welcome than just a quantile region....

Thanks,
Hoa

[Martin Wicke]

While you can do that, a histogram summary will simply show the mean and quantiles for the values in whatever you're summarizing.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/1fd504a5-ba60-4f63-b186-8f7bfe7a7544%40tensorflow.org.

[lethi...@gmail.com]

sorry but W is a matrix, multidimension. How can you just plot something like mean and quantiles of a matrix ?

[lethi...@gmail.com]

I think the essential point is that W and b in the Histogram TensorBoard points out just the final step Regression of Neural Network.

W and b here are just scalars.

It is not at all the same thing W matrix and b vectors like in the literatures (Yoshua Bengio, G. Hinton, Y. Lecun, Mikolov,....)

[Martin Wicke]

You take the values of the matrix and compute a distribution. Then you bin that distribution and you have your histogram.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/80a0acab-7e92-4b81-a00c-fb03ca1f85b5%40tensorflow.org.