Error in computer_vision_fine_tuning.py example

[shamanez]

🐛 Bug

Give the following error

"AttributeError: 'MyLightningCLI' object has no attribute '_get'"

To Reproduce

python ./pl_examples/domain_templates/computer_vision_fine_tuning.py

Environment

• PyTorch Lightning Version: 1.4.9
• Python version: 3.9
• OS: Ubuntu

Additional context

[justusschock]

cc @mauvilsa any idea what causes this?

[mauvilsa]

I get a different error if I run python ./pl_examples/domain_templates/computer_vision_fine_tuning.py with lightning 1.4.9, python 3.8 and jsonargparse 3.19.4.

Traceback (most recent call last):
  File "./pl_examples/domain_templates/computer_vision_fine_tuning.py", line 292, in <module>
    cli_main()
  File "./pl_examples/domain_templates/computer_vision_fine_tuning.py", line 287, in cli_main
    MyLightningCLI(TransferLearningModel, CatDogImageDataModule, seed_everything_default=1234)
  File "/home/mvillegas/repos/thirdparty/pytorch-lightning/pytorch_lightning/utilities/cli.py", line 285, in __init__
    self.instantiate_classes()
  File "/home/mvillegas/repos/thirdparty/pytorch-lightning/pytorch_lightning/utilities/cli.py", line 341, in instantiate_classes
    self.instantiate_trainer()
  File "./pl_examples/domain_templates/computer_vision_fine_tuning.py", line 281, in instantiate_trainer
    finetuning_callback = MilestonesFinetuning(**self.config_init["finetuning"])
TypeError: ABCMeta object argument after ** must be a mapping, not MilestonesFinetuning

@shamanez if you update jsonargparse to this version do you still have the _get error?

That example needs to be updated. Now the classes added with parser.add_class_arguments are automatically instantiated. Thus this instantiation is no longer necessary computer_vision_fine_tuning.py#L281.

I can fix that example.

[shamanez]

Actually in the new example, line 281 is different.

https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pl_examples/domain_templates/computer_vision_fine_tuning.py#L281

This line also giving an error saying there's no such attribute.

[shamanez]

@mauvilsa yes versions are the same except I use python 3.9. Btw do you think this template is up to date inorder to use it as a base template in my new project ?

Because this is the only template I saw that uses CLI callbacks

[mauvilsa]

@mauvilsa yes versions are the same except I use python 3.9. Btw do you think this template is up to date inorder to use it as a base template in my new project ?

Because this is the only template I saw that uses CLI callbacks

No, it is not up to date. To make a callback configurable is now simpler. In add_arguments_to_parser the callback should be added with add_lightning_class_args as explained in lightning_cli.html#configurable-callbacks. With this there is no need to modify the instantiate_trainer method of LightningCLI thus these lines would be removed computer_vision_fine_tuning.py#L280-L283.

[shamanez]

I hope the following colab is an updated version.

https://pytorch-lightning.readthedocs.io/en/latest/notebooks/lightning_examples/text-transformers.html

But above Colab template is not using LightiingCLI. I feel with LightningCLI code gets much cleaner specially no need to initialize the trainer separately. Is my understanding correct @mauvilsa ?

[mauvilsa]

Created a pull request #9934 and tested it locally in python 3.8 and 3.9.

[mauvilsa]

But above Colab template is not using LightiingCLI. I feel with LightningCLI code gets much cleaner specially no need to initialize the trainer separately. Is my understanding correct @mauvilsa ?

Certainly LightningCLI simplifies more things. LightningCLI is still in beta and I am not sure what is the plan on updating more examples now or when it comes out of beta.