fix add graph func is not called bug

[shenmishajing]

Motivation

Fix add graph func is not called bug. See issue #621

Modification

• Call visualizer.add_graph after model build.
• Add watch_kwargs param to WandbVisBackend.

Checklist

4. The documentation has been modified accordingly, like docstring or example tutorials.

[HAOCHENYE]

Will it be more proper to call add_graph in NativeVisualizerHook.before_train, and this feature can be enabled by setting
watch_model=True (default False, not sure will it harm the performance)

[HAOCHENYE]

BTW, since add_graph has not been used by the downstream repo, could we rename it to a more proper name? @zhouzaida

[shenmishajing]

Will it be more proper to call add_graph in NativeVisualizerHook.before_train, and this feature can be enabled by setting watch_model=True (default False, not sure will it harm the performance)

I agree with that. But, does the NativeVisualizerHook have been added to runner now? I do not find it now, and that's why I add it in runner init directly.

[HAOCHENYE]

NativeVisualizerHook should be a custom hook. If users need to visualize something, they can inherit or copy NativeVisualizerHook to implement custom VisualizerHook. The code snippet of config to enable VisualizerHook is like this:

config_for_xxx.py

custom_hooks = [
    dict(type='NativeVisualizerHook')
    ...
]

[shenmishajing]

NativeVisualizerHook should be a custom hook. If users need to visualize something, they can inherit or copy NativeVisualizerHook to implement custom VisualizerHook. The code snippet of config to enable VisualizerHook is like this:

config_for_xxx.py

custom_hooks = [
    dict(type='NativeVisualizerHook')
    ...
]

Ok, got it. I think it's more proper to call add_graph in NativeVisualizerHook.before_train. By the way, should we move the add_config call to NativeVisualizerHook.before_train as well? @HAOCHENYE

[HAOCHENYE]

I think add_config is a more commonly used function for the general experiments, it is ok to be called in runner.__init__() .

[shenmishajing]

Ok, thanks for your reply.

[codecov]

Codecov Report

Base: 77.98% // Head: 77.98% // Decreases project coverage by -0.00% ⚠️

Coverage data is based on head (0a05e71) compared to base (aaba1d8).
Patch coverage: 66.66% of modified lines in pull request are covered.

❗ Current head 0a05e71 differs from pull request most recent head e2acbb2. Consider uploading reports for the commit e2acbb2 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #632      +/-   ##
==========================================
- Coverage   77.98%   77.98%   -0.01%     
==========================================
  Files         126      126              
  Lines        9059     9065       +6     
  Branches     1804     1804              
==========================================
+ Hits         7065     7069       +4     
- Misses       1681     1683       +2     
  Partials      313      313              

Flag	Coverage Δ
unittests	77.98% <66.66%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmengine/hooks/naive_visualization_hook.py	93.93% <50.00%> (-2.84%)	⬇️
mmengine/visualization/vis_backend.py	84.40% <75.00%> (-0.21%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[zhouzaida]

Hi @shenmishajing , thanks for your contribution. After having a discussion with WandB team, they have no plan to support visualizing models in wanbd.watch but we will merge this PR because it is a useful feature.