How to run 'powerful-fish' on my image?

[tibuch]

Hi everyone,

I wanted to try one of the available models to enhance membranes in one of my images. The image is 3D with shape = (91, 1024, 1024). I tried to use the CLI with bioimageio predict-image but I get an ERROR: Could not consume arg: predict-image.

Then I thought to try it via Python, but I encounter this:

prediction_pipeline.predict_sample_with_blocking(sample)
---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Cell In[85], line 1
----> 1 prediction_pipeline.predict_sample_with_blocking(sample)

File ~/Gitrepos/gliberal_crypt-morphogenisis/.pixi/envs/modelzoo/lib/python3.12/site-packages/bioimageio/core/_prediction_pipeline.py:200, in PredictionPipeline.predict_sample_with_blocking(self, sample, skip_preprocessing, skip_postprocessing, ns, batch_size)
    197     self.apply_preprocessing(sample)
    199 if isinstance(self.model_description, v0_4.ModelDescr):
--> 200     raise NotImplementedError(
    201         "predict with blocking not implemented for v0_4.ModelDescr {self.model_description.name}"
    202     )
    204 ns = ns or self._default_ns
    205 if isinstance(ns, int):

NotImplementedError: predict with blocking not implemented for v0_4.ModelDescr {self.model_description.name}

What is the recommended way to run tiled prediction with a v0.4 model?

[esgomezm]

Hi Tim!

The issue is basically that we cannot perform tiling with the new version of bioimageio.core library on "old models" (with a format version < 0.5).

For simply testing the inference, I would recommend using the test run online in the website. It will do the tiling and it's straightforward to drag&drop the image and click on run model. I would reduce slightly the Z-dimension from 91 to 64 and if possible the (X,Y) dimensions. If not, it will just take more time.

Otherwise, to use the Python code, here are some alternatives:

• Download the previous version of the bioimageio.core library (bioimageio.core==0.5.11) and try running the following scripts in this example notebook: https://github.com/HenriquesLab/DL4MicEverywhere/blob/main/notebooks/Bespoke_notebooks/bioimageio-pytorch/load_bioimageio_pytorch.ipynb

• Update the model yourself. Basically, load the model with the current bioimageio.core, as explained here, export the model (it will export the same model with an updated format) and run the prediction again.

I hope this helps. Please, let us know how it goes

[FynnBe]

Hi @tibuch I'm currently working on restoring some of the convenience functionality from before the update of bioimageio.spec (and its incorporation into bioimageio.core): #406 So this should become more accessible in the next release.