Fix transposed dimensions

[anuprulez]

This PR fixes #1608

Tested with all the models in https://training.galaxyproject.org/training-material/topics/imaging/tutorials/process-image-bioimageio/tutorial.html

In the above tutorial, "NeuronSegmentationInEM" does not have any Pytorch model available at the mentioned link. I propose to change it to https://bioimage.io/#/?id=10.5281%2Fzenodo.5874741

ping @kostrykin

[kostrykin]

Thanks!

Can you maybe please use something like <output name="output_predicted_image" file="result.tiff" ftype="tiff" compare="image_diff" /> for the tests?

galaxytools/tools/bioimaging/bioimage_inference.xml

Lines 67 to 82 in 973836f

	<test>
	<param name="input_imaging_model" value="input_imaging_model.zip" location="https://zenodo.org/api/records/6647674/files/weights-torchscript.pt/content"/>
	<param name="input_image_file" value="input_nucleisegboundarymodel.png"/>
	<param name="input_image_input_size" value="256,256,1,1"/>
	<param name="input_image_input_axes" value="bcyx"/>
	<output name="output_predicted_image" ftype="tiff">
	<assert_contents>
	<has_size size="524846" delta="110" />
	</assert_contents>
	</output>
	<output name="output_predicted_image_matrix" ftype="npy">
	<assert_contents>
	<has_size size="524416" delta="110" />
	</assert_contents>
	</output>
	</test>

This will be more reliable.

[kostrykin]

You also need to increment the @VERSION_SUFFIX@:

galaxytools/tools/bioimaging/bioimage_inference.xml

Line 5 in 973836f

<token name="@VERSION_SUFFIX@">2</token>

[kostrykin]

And another question, this is beyond the scope of this PR, but maybe we can discuss this :)

I tried to wrap my head around the input_axes parameter in this tool, but I just don't get it. Why is this input required?

galaxytools/tools/bioimaging/bioimage_inference.xml

Lines 40 to 44 in 973836f

	<param name="input_image_input_axes" type="select" label="Axes of the input image" optional="false" help="Provide the input axes of the input image. See the chosen model's RDF file to find the correct axes. For example: for the BioImage.IO model MitochondriaEMSegmentationBoundaryModel, the input axes is 'bczyx'">
	<option value="bczyx">bczyx</option>
	<option value="bcyx">bcyx</option>
	<option value="byxc">byxc</option>
	</param>

The only place that I could find where it is used is in

galaxytools/tools/bioimaging/main.py

Line 123 in 973836f

target_dimension = len(input_axes)

so it looks like you only need the number of axes?

If this is the case, I have two follow-up questions:

1. Wouldn't only two options be enough then, something like

   <option value="bcyx">Four axes (e.g., bcyx, byxc)</option> 
   <option value="bczyx">Five axes (e.g., bczyx)</option> 

2. Can't this information be extracted from the input image directly? I mean, if it's just the number of axes, it should be the number of dimensions of the numpy array of the image, shouldn't it?

[anuprulez]

so it looks like you only need the number of axes?

Yes, it is required to get the exact number of dimensions that the model needs to process the image.

1. Wouldn't only two options be enough then, something like

   <option value="bcyx">Four axes (e.g., bcyx, byxc)</option> 
   <option value="bczyx">Five axes (e.g., bczyx)</option> 

Yes, it should work. I will make this change

2. Can't this information be extracted from the input image directly? I mean, if it's just the number of axes, it should be the number of dimensions of the numpy array of the image, shouldn't it?

We are not currently taking input numpy arrays for processing as inputs, only TIFF and PNG. However, the numpy arrays might contain the correct number of dimensions that the models need but the TIFF files do not have this information. For example, the RDF file for this model "Mitochondria EM Segmentation Boundary Model" (https://bioimage.io/#/?id=10.5281%2Fzenodo.5874841) says that the input size is 256 x 256 x 32 x 1 but the input axes are bczyx. This difference I am not sure I understand it. The mode requires TIFF image to be represented as 5 dimensional to be processed.

[anuprulez]

Thanks!

Can you maybe please use something like <output name="output_predicted_image" file="result.tiff" ftype="tiff" compare="image_diff" /> for the tests?

galaxytools/tools/bioimaging/bioimage_inference.xml

Lines 67 to 82 in 973836f

	<test>
	<param name="input_imaging_model" value="input_imaging_model.zip" location="https://zenodo.org/api/records/6647674/files/weights-torchscript.pt/content"/>
	<param name="input_image_file" value="input_nucleisegboundarymodel.png"/>
	<param name="input_image_input_size" value="256,256,1,1"/>
	<param name="input_image_input_axes" value="bcyx"/>
	<output name="output_predicted_image" ftype="tiff">
	<assert_contents>
	<has_size size="524846" delta="110" />
	</assert_contents>
	</output>
	<output name="output_predicted_image_matrix" ftype="npy">
	<assert_contents>
	<has_size size="524416" delta="110" />
	</assert_contents>
	</output>
	</test>

This will be more reliable.

I agree but with file="results.tiff", we have to keep results.tiff files in the test_data, no? These tiff files are big (> 1 MB) and the related tests may not pass. For smaller files (<1MB) we can do it

[kostrykin]

Yes, it should work. I will make this change

Awesome!

I agree but with file="results.tiff", we have to keep results.tiff files in the test_data, no? These tiff files are big (> 1 MB) and the related tests may not pass. For smaller files (<1MB) we can do it

At least for this particular test, the file is just about 500 kB: result.tiff.zip

[anuprulez]

Also, opened a PR for updating the associated GTN tutorial to replace the unsupported TensorFlow model to a Pytorch one: galaxyproject/training-material#5957

[kostrykin]

With the image_diff check in place, I think we can remove this?

The has_size test is very heuristical, it might even raise a false-negative check failure.

[anuprulez]

sure!

[kostrykin]

Thanks!

[kostrykin]

For example, the RDF file for this model "Mitochondria EM Segmentation Boundary Model" (https://bioimage.io/#/?id=10.5281%2Fzenodo.5874841) says that the input size is 256 x 256 x 32 x 1 but the input axes are bczyx. This difference I am not sure I understand it.

Alright, we're on the same page here :)