Update TAGAN-axons

[bioimageiobot]

This is an automatic PR created by the @bioimageiobot regarding changes to the resource item 10.5281/zenodo.7872357.
The following version(s) will be added:

• 7876764 (version 2)

Please review the changes and make sure the new item or version(s) pass the following check list:

• Passed the BioImage.IO CI tests: static (and dynamic) validations
• The meta information for the RDF item is complete
  • The tags are complete and describe the model
  • Naming is intuitive and descriptive, example: Multi-Organ Nucleus Segmentation (StarDist 2D)
  • Authors are provided
  • Documentation is complete
    • For models, include an overview, describe how the model is trained, what is the training data, how to use the model, how to validate the results and list the references. TODO: Model documentation template.
• Approved by at least one of the BioImage.IO admin team member.

Maintainers: @cbouchard22

Note: If you updated or re-uploaded another version for the current item on Zenodo, this PR won't be changed automatically. To proceed, you can do the following:

1. Block this version, but keep looking for future versions: Edit the current resource.yaml and keep the top-level status field as accepted, but change the status under the current version to blocked.
2. Accept this version and keep looking for future versions: Merge this PR for now.
3. Keep proposed version(s) (and this resource in general if it is new) as pending: Close this PR without merging.

Then wait for the CI on the main branch to complete. It should detect the new version(s) and create another PR for the new version(s).

Previous PRs of this resource: #601

[esgomezm]

Hi @FynnBe you changed the requirements file regarding the input numpy but the CI keeps failing. Do you know why? Is there a way we can help to correct it?
@IvanHCenalmor @mariana-gferreira

[FynnBe]

I changed test-input.npy to test-input-cpu.npy, because "test-input.npy" is not present in the record, while "test-input-cpu.npy" is...

The error "Could not run 'aten::empty_strided' with arguments from the 'CUDA' backend." looks like the something is hardcoded to live in GPU memory.
To debug I'd try to test the model on cpu locally.
In a suitable conda env bioimageio test-model rdf.yaml should produce the same error (once test-input.npy is changed to test-input-cpu.npy or the file is renamed).

The files gpu.pt, cpu.pt and TAGAN_Axons_G.pth are not referenced in the rdf.yaml at all, which is a bit confusing.

[esgomezm]

Hi @cbouchard22 and @FynnBe
So, good news: The model we needed to add in the spec was the cpu.pt one. With that, everything works well. To enable the model to run in deepImageJ as well, I included some extra info in the specs (rdf.yaml) and it works nicely also in fiji <3

The bad news: this gan is producing a different output each time, so we cannot test that the output test image provided is the same as the one that the network outputs when we test it. We need to discuss this (I'm opening an issue) and update the CI.

Meanwhile, @cbouchard22 could you please upload the model corrected and update it? Please, find here the corrected model: https://drive.google.com/file/d/1ujdaAZwXFRYofxyoDrpX0kD6HgsO1TM9/view?usp=sharing