`plot_model` bogus error

[khteh]

        plot_model(
            self._model,
            to_file="output/RNN_SignsLanguageDigits.png",
            show_shapes=True,
            show_dtype=True,
            show_layer_names=True,
            rankdir="TB",
            expand_nested=True,
            show_layer_activations=True)

"dot" with args ['-Tpng', '/tmp/tmpxa9l83z3/tmpxd55ygh6'] returned code: -6

stdout, stderr:
 b''
b"failed at node 1187[0]\ndot: maze.c:311: chkSgraph: Assertion `np->cells[0]' failed.\n"

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/usr/src/Python/pAIthon/RNN_SignsLanguageDigits.py", line 531, in <module>
    rnn.BuildModel(args.retrain)
    ~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^
  File "/usr/src/Python/pAIthon/RNN_SignsLanguageDigits.py", line 351, in BuildModel
    plot_model(
    ~~~~~~~~~~^
        self._model,
        ^^^^^^^^^^^^
    ...<5 lines>...
        expand_nested=True,
        ^^^^^^^^^^^^^^^^^^^
        show_layer_activations=True)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/khteh/.local/share/virtualenvs/pAIthon-GaqEDHQT/lib/python3.13/site-packages/keras/src/utils/model_visualization.py", line 510, in plot_model
    dot.write(to_file, format=extension)
    ~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/khteh/.local/share/virtualenvs/pAIthon-GaqEDHQT/lib/python3.13/site-packages/pydot/core.py", line 1730, in write
    s = self.create(prog, format, encoding=encoding)
  File "/home/khteh/.local/share/virtualenvs/pAIthon-GaqEDHQT/lib/python3.13/site-packages/pydot/core.py", line 1844, in create
    assert process.returncode == 0, (
           ^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: "dot" with args ['-Tpng', '/tmp/tmpxa9l83z3/tmpxd55ygh6'] returned code: -6

[innat]

This is one of the annoying error, hardly interpretable. There is probably no exact cause of it. Sometimes if the model arch becomes too complex, it happens to plot properly. For example, if I use UNETR with ViT backbone (24 layers), plot model will fail, but if I reduce the size, say 2/4 layers, then it works.

[khteh]

Yeah, that sucks!

[maitry63]

@khteh Hi, I reproduce the issue with UNETR with ViT backbone (24 layers) & the plot is coming fine on Keras: 3.12.0
ans keras_hub: 0.21.1. Attached gist file for the same.

[khteh]

It breaks when working on custom complex model: https://github.com/khteh/pAIthon/blob/master/Classification/RNN_SignsLanguageDigits.py

[innat]

@maitry63 I meant UNETR not ViT with UNet style. Also, the error message is hard to interpret.

I got the same error few days ago with UPerNet + EfficientNet backbone. But if I use UPerNet + Swin Transformer, that it works. There is high chance of this error for being complex graph (many connections).