search

viscom-ulm / Net2Vis

Automatic neural network visualizations generated in your browser!
MIT License
140 stars 18 forks source link

Select shape for layers #33

Open lleuven opened 2 years ago

lleuven commented 2 years ago

Describe the request

It would be very nice, if the shape of a layer can be customized by the user. In my case, I am using a concatenate layer which unfortunately changes its shape depending on the total number of layers in the network. I personally would prefer a block, for example, in this place.

How can it be tackled?

When clicking on a layer or group, I would like to see the option to either change the shape of the selected layer or to change at least its properties.

Sparkier commented 2 years ago

I am not sure I understand. Could you maybe post a link to your network explaining the problem in more detail?

lleuven commented 2 years ago

You can have a look at my networks, but you can see the issue in the opening U-net in your puplication on Net2Vis on arxiv too.

https://viscom.net2vis.uni-ulm.de/BhzmR90B6a2Pe0KIiIp8YqAYfZll45n5ugJz59GgCPWW8cfSNo https://viscom.net2vis.uni-ulm.de/NXEzhamclMlR3UE1wQdYBZtwN43cgIdPyxuQwv0USoGPI7s4yW

grafik (copied from https://arxiv.org/pdf/1902.04394.pdf )

Here you can see, that the split for the skip connections and the concatenate layers look different for each successive layer but also for the counterparts (split and concat on same layer level). At least to me, the green layers would initially appear to do different things because the shape is different. Refering to my examples, I would like to see an option that I can harmonise the shape of the concatenate layers (or replace the ">" shape with an rectangular box). I hope my idea is clearer now.

Sparkier commented 2 years ago

Thank you for the clarification, I think I understand. The current design is intentional as it communicates the spatial resolution and number of features just as any other layer.

Having said that, adding an option to replace the concat layers with something different would technically totally be possible. My problem is that I currently do not have the time to implement this. I could probably take a look at it late May. I'm sorry about this but sadly I don't get paid for this.

If you want to have these figures without the > right now, there are two options:

  1. Contribute to the project yourself by implementing this (I could tell you where to touch things).
  2. Export the figure, then open the svg in your favorite vector graphics tool (affinity designer, illustrator, ...) and replace the svg elements for the layers.

Sorry that I can't help more right now.