HDN minimum example #396
HDN minimum example #396
Conversation
| data_type: Literal["array", "tiff", "custom"], | ||
| axes: str, | ||
| patch_size: list[int], | ||
| batch_size: int, | ||
| num_epochs: int, | ||
| augmentations: Optional[list[Union[XYFlipModel, XYRandomRotate90Model]]] = None, | ||
| independent_channels: bool = True, | ||
| loss: Literal["mae", "mse"] = "mae", | ||
| n_channels_in: Optional[int] = None, | ||
| n_channels_out: Optional[int] = None, | ||
| logger: Literal["wandb", "tensorboard", "none"] = "none", | ||
| model_params: Optional[dict] = None, | ||
| dataloader_params: Optional[dict] = None, |
There was a problem hiding this comment.
Unused parameters, should be passed to relevent configs.
Also this function needs more thought with how it will work with the _create_configuration function above that currently by default creates a UNet configuration.
| algorithm_params = create_algorithm_configuration() | ||
| data_params = create_data_configuration() | ||
| training_params = create_training_configuration() |
There was a problem hiding this comment.
These functions don't exist, I think you are looking for algorithm_factory, data_factory, and the training configuration is just implemented by initialising the TrainingConfig class directly since there is only one type.
| """ | ||
| return [ | ||
| "restoration", | ||
| "UNet", |
| from .transformations.n2v_manipulate_model import ( | ||
| N2VManipulateModel, | ||
| ) | ||
| from .vae_algorithm_model import VAEAlgorithmConfig |
There was a problem hiding this comment.
| from .vae_algorithm_model import VAEAlgorithmConfig | |
| from algorithms.vae_algorithm_model import VAEAlgorithmConfig |
| from .data_model import DataConfig | ||
| from .fcn_algorithm_model import FCNAlgorithmConfig |
| if summary.status == "failed": | ||
| raise ValueError(f"Model description test failed: {summary}") | ||
| # raise ValueError(f"Model description test failed: {summary}") | ||
| print("I don't give a flying fuck!") | ||
| # TODO BMZ expect just one output but our VAE outputs a whole stack of things |
There was a problem hiding this comment.
Can we just have at the start of this function
if isinstance(model, VAEModule):
raise NotImplementedError(
"Exporting a VAE model to the Bioimage Model Zoo format has not been implemented yet, we are working on it!"
)| r"""Compute the log-probability. | ||
|
|
||
| at `x` of a Gaussian distribution |
There was a problem hiding this comment.
I think this additional full stop and new line might have been an accident
| elif isinstance(self.data_config, N2VDataConfig): | ||
| return self.patch_transform(patch=patch) | ||
| else: | ||
| raise ValueError( | ||
| "Something went wrong! No target provided (not supervised training) " | ||
| "while the algorithm is not Noise2Void." | ||
| ) | ||
| return self.patch_transform(patch=patch) |
There was a problem hiding this comment.
I think we still want to raise an error here if the user does not provide a target for a supervised algorithm
|
Also if the LVAE model, LVAE losses and anything else related are not going to be sandboxed from the rest of CAREamics any more we need to edit pre-commit config to allow mypy to check the relevant code again careamics/.pre-commit-config.yaml Line 33 in 3c1abfd |
Description
Note
tldr:
MVE of HDN
Background - why do we need this PR?
Training HDN with careamist
Overview - what changed?
Changes Made
New features or files
Modified features or files
Removed features or files
How has this been tested?
Related Issues
Breaking changes
Additional Notes and Examples
--- BMZ doesn't work
Please ensure your PR meets the following requirements: