-
Notifications
You must be signed in to change notification settings - Fork 6.8k
plot_network issue with LRN #10282
Comments
Seems |
I'll fix the LRN op. @reminisce do we really need to provide output names for convolution and fullyconnected? It seems the old interface doesn't provide output names, except for the operators that have multiple outputs. BTW, it seems your fix has a bug. |
I checked all other operators. It seems others are fine. |
@zheng-da The legacy interface has |
@zheng-da The funny thing is that even with the typo of writing |
@reminisce you are right. nnvm::FListInputNames defines the signature of the functor, so your code is fine. it's just a little confusing. I'm still confused if we have to define |
@ThomasDelteil my PR (#10296) has been merged. Can you try it again? |
@zheng-da There is no hard requirement. Doing this is for two reasons:
As you can see, there are multiple places in the code base following this convention. In quantization, we also use this to determine which layer should be calibrated or not. I will submit a PR to fix the cosmetic issue with unit tests covering convolution, fc, and lrn. |
@ThomasDelteil We have open-sourced MXBoard which enables users to visualize MXNet data (including network structures) in TensorBoard. The current version on PyPI is 0.1.0rc3. You can either install from source or |
I will try it next week for sure, looks great! |
@zheng-da the fix is working the graph is displaying |
There is an issue in
master
that is not present in1.1.0
Plotting the network on this tutorial:
http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-10165/10/tutorials/onnx/inference_on_onnx_model.html
Fails with this error:
The issue was introduced before
mxnet==1.2.0b20180221
It is related to the fact that the LRN layer does not have an output, whilst in
1.1.0
I can dosym.get_internals()['lrn0_output']
The issue was introduced by this PR and this commit c3e3a83:
#9677
The text was updated successfully, but these errors were encountered: