Discard init_args after class_path change causes error

[samvanstroud]

🐛 Bug report

When composing configs, it appears that overwritten class paths do not properly drop the old init_args namespace, causing an error when the new class does not have the same init_args.

Here is the error message:

$ main fit --config config_1.yaml --config configs_2.yaml
/unix/atlastracking/svanstroud/miniconda3/envs/salt/lib/python3.10/site-packages/jsonargparse/typehints.py:909: UserWarning: Due to class_path change from 'models.GATv2Attention' to 'models.ScaledDotProductAttention', discarding init_args: {'num_heads': 8, 'head_dim': 16, 'activation': Namespace(class_path='torch.nn.SiLU')}.
  warnings.warn(
usage: main [-h] [-c CONFIG] [--print_config[=flags]] {fit,validate,test,predict,tune} ...
main: error: Parser key "model": Problem with given class_path 'models.Tagger':
    - Parser key "gnn": Does not validate against any of the Union subtypes
      Subtypes: (<class 'torch.nn.modules.module.Module'>, <class 'NoneType'>)
      Errors:
        - Problem with given class_path 'models.Transformer':
          - Parser key "attention": Problem with given class_path 'models.ScaledDotProductAttention':
              - 'Configuration check failed :: No action for destination key "num_heads" to check its value.'
        - Expected a <class 'NoneType'>
      Given value type: <class 'jsonargparse.namespace.Namespace'>
      Given value: Namespace(class_path='models.Transformer', init_args=Namespace(embd_dim=128, num_heads=8, num_layers=8, attention=Namespace(class_path='models.ScaledDotProductAttention', init_args=Namespace(num_heads=8, head_dim=16, activation=Namespace(class_path='torch.nn.SiLU'))), activation='SiLU', residual=True, norm_layer='LayerNorm', dropout=0.1, out_proj=False))

To reproduce

Expected behavior

The overwriting class should not be required to have the same init_args as the overwritten class.
It would also be nice to have an option to suppress the warning about overwritten args.

Environment

• jsonargparse 4.18.0
• Python 3.10.6

[speediedan]

@samvanstroud Thanks for posting this issue! (and to @mauvilsa et al. for this valuable package!). As I've observed it with Fine-Tuning Scheduler as well, I'm pinning finetuning-scheduler's jsonargparse dependency to >=4.15.2, <4.18.0 until it's resolved.

[samvanstroud]

@mauvilsa is this related to Lightning-AI/pytorch-lightning#15796? I guess I was surprised that it was not fixed by #199, but I was composing yaml files so perhaps another fix is needed for this case?

[mauvilsa]

Actually what happened is that this is a regression and was caused by #199. The original issue Lightning-AI/pytorch-lightning#15796 is about discarding init_args when using individual command line arguments, not configs. It wasn't noticed in #199 because there is no unit test for the discard from config files. I will fix this and add a unit tests to prevent future regressions.

[mauvilsa]

Created a pull request hopefully fixing the issue. Please test it out, e.g.

pip install "jsonargparse @ https://github.com/omni-us/jsonargparse/zipball/issue-205-discard-init-args"