Histograms in Tensorboard

[AakashKumarNain]

I was training a few models, and for debugging something I wanted to use Tensorboard. For me, the use-case was to look for the histograms of the weights for each layer in the network. When I added the callback to the training loop, this is what I get:

Expected behavior: The histogram plot should display the distribution of weights, and biases(if any) separately for each layer in the model.

cc: @sampathweb

[AakashKumarNain]

Please let me know if you want help to fix this. I can work on this

[fchollet]

TensorBoard relies on variable names, so this is a name scope issue. We should introduce name spaces in layers / models as appropriate and in a cross-backend way (though I think it would only be meaningful in TF).

[AakashKumarNain]

Thanks for the inputs @fchollet

though I think it would only be meaningful in TF

Any specific reason behind this?

[fchollet]

Any specific reason behind this?

"Tensor names" are not a thing in JAX, and as far as I know also not a thing in PyTorch (which makes sense, being eager first).

[sachinprasadhs]

@AakashKumarNain , Feel free to close the issue if you do 't have any follow-up question.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[miticollo]

I would want to try to propose a solution.

IMO an easy and less error-prone solution could be the following:

1. Replace this line with

   weight_name = weight.path.replace(":", "_")

   Checking this TensorBoard guide this is possible.
2. Replace this line with

   histogram_weight_name, weight, step=epoch, description=f"This KerasVariable belongs to **{layer.name}**"

   or another description.

Anyway I don't know if this is a satisfying solution.