hydra-lightning
hydra-lightning copied to clipboard
romesco
Configuration classes enabling Hydra to configure and manage Pytorch Lightning projects.
In _hydra_configs/pytorch_lightning/callbacks.py_ in ModelCheckpointConf there is no default value for feeld `dirpath`. Should be: `dirpath: Any = None` according to https://pytorch-lightning.readthedocs.io/en/latest/generated/pytorch_lightning.callbacks.ModelCheckpoint.html
It is impossible to instantiate a `Trainer` without this modification on the new `ptl` version. See: https://github.com/PyTorchLightning/pytorch-lightning/pull/6163
This working properly currently depends on @awaelchli 's fix here: https://github.com/PyTorchLightning/pytorch-lightning/pull/5155 Ironing out a few details to clean it up. Testing is a good idea too.
It looks like there's no direct support for Loggers (or other nested structures) used by the Trainer. Is there any plan to suppport that? Is it as simple as running...
This has been coming up frequently in different groups. Provide a possible MWE and a centralized place to have community members provide suggestions.
Probably add a simple one until recursive instantiation is fully supported in hydra.
Include a basic example of configuring callbacks.
Modify `mnist_00.py` to use lightning example with lightning configuration classes.
Ensure basic MNIST Lightning example runs with latest hydra / latest lightning.
Quoting @omry: Something like: ``` trainer: _target_: Trainer callback: _target_: hydra.utils.get_method path: my_callback ``` Another possibility using existing hydra methods: ``` _target_ : foo.Foo callback: ${get_function:my_callback} ```