Improve type coercion somehow

[ctrueden]

>>> img = io.imread('http://samples.fiji.sc/_malaria_parasites.tif')
>>> import numpy as np
>>> img = np.mean(img, axis=2)
>>> vessels = np.zeros(img.shape, dtype=img.dtype)
>>> ij.op().filter().frangiVesselness(imglyb.to_imglib(vessels), imglyb.to_imglib(img), [1, 1], 20)
<net.imglib2.RandomAccessibleInterval at 0x10b379048 jclass=net/imglib2/RandomAccessibleInterval jself=<LocalRef obj=0x7f8a3ec44e08 at 0x136257cb0>>
>>> ij.scifio().datasetIO().save(imglyb.to_imglib(vessels), 'out.png')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "jnius/jnius_export_class.pxi", line 906, in jnius.JavaMultipleMethod.__call__ (jnius/jnius.c:29616)
  File "jnius/jnius_export_class.pxi", line 632, in jnius.JavaMethod.__call__ (jnius/jnius.c:25965)
  File "jnius/jnius_conversion.pxi", line 68, in jnius.populate_args (jnius/jnius.c:8922)
  File "jnius/jnius_utils.pxi", line 207, in jnius.check_assignable_from (jnius/jnius.c:6212)
jnius.JavaException: Invalid instance of 'net/imglib2/python/ReferenceGuardingRandomAccessibleInterval' passed for a 'net/imagej/Dataset'

[hanslovsky]

DatasetIOService does not have a method DatasetIOService.save( RandomAccessibleInterval, String ), does it? The only way to make this work would be to add (on the python side) a method like imagej.to_dataset that accepts

• numpy.ndarray, or
• RandomAccessibleInterval

and returns a Dataset.
Is that what you have in mind?

[ctrueden]

I think what I meant when I filed this issue is that the pyjnius objects are statically typed like Java. So e.g. if a method return a java.util.Collection<String>, even though in practice the returned object will be some concrete type such as java.util.ArrayList<String>, you won't be able to pass that object to another function requiring e.g. a java.util.List<E>, unless you cast the object with myListProcessingFunction(jnius.cast(List, my_collection)). This is consistent with how Java's static typing works (i.e., in Java you'd need to write myListProcessingFunction((List) my_collection)), but is inconsistent with Python's less strict typing approach.

That said, my example above does not illustrate this issue well, since as @hanslovsky points out, the DatasetIOService.save method does not accept any old RAI.

[mpinkert]

I've been trying to write a helper function for saving a numpy array through ImageJ, and I think I've also hit upon this issue. The imglyb.to_imglib function, and correspondingly our new ij.py.to_java function, returns a net/imglib2/python/ReferenceGuardingRandomAccessibleInterval class. So far I haven't been able to save it using ij.io.save() or the corresponding scifio IO.

Casting it to net.imglib2.img.Img makes the command do nothing, while casting to net.imagej.DefaultDataset throws this error: java.lang.IncompatibleClassChangeError

I was able to save an ImgPlus the ij.IJ.save(). Opening an image in IJ2, which gave a net.imageJ.DefaultDataset, and saving it with ij.io().save() did work though.

[ctrueden]

@mpinkert You can use the DatasetService to turn a RAI into a Dataset:

from skimage import data
coins = data.coins()
java_coins = ij.py.to_java(coins)
dataset = ij.dataset().create(java_coins)
ij.scifio().datasetIO().save(dataset, '/Users/curtis/Desktop/coins.png')

[mpinkert]

@ctrueden I just tested that out - I saw two possible pain points.

1.) Python generated numpy arrays hit pixel type problems. For example, trying to pass a double gives big errors. With tifs, ImageJ cannot open the result because Unsupported BitsPerSample: 64. With .pngs, the data doesn't save in the first place due to Unsupported image type 'double'. Example:

from pathlib import Path

img = numpy.zeros([512, 512])
jimg = ij.py.to_java(img)
jdataset = ij.dataset().create(jimg)
jpath = str(Path(Path.home(), 'documents', 'test.tif'))
ij.io().save(jdataset, jpath)

You can get around this by specifying an appropriate datatype using img.astype(), but we should probably discuss if we want the user to specify this, or if we want to force it into an appropriate type when they pass it in.

2.) Color images are not converted properly. Numpy is normally reverse indexed, e.g. [z, y, x]. However, for RGB it is [y, x, C]. Turning it into a dataset makes it save as a 3 column / x row / y slice image. Fixing the index order is easy, but the image is still saved as a 3D image instead of RGB. Is there an easy way to specify that the dataset should be saved as RGB?

from pathlib import Path
from skimage import io

url = 'https://github.com/hanslovsky/imglyb-examples/raw/master/resources/butterfly_small.jpg'
img = io.imread(url)
jimg = ij.py.to_java(img)
jdataset = ij.dataset().create(jimg)
jpath = str(Path(Path.home(), 'documents', 'test.tif'))
ij.io().save(jdataset, jpath)

[mpinkert]

@ctrueden

Problem 1, the numpy type errors, occurs both ways. Converting an 8-bit image to a ndarray makes it dtype('float64'). This is a memory issue, and also a problem for RGB images with matplotlib, as matplotlib only plots int8 or dtype in range [0, 1]

Problem 2 is now part of issue #17.

[ctrueden]

Four things being discussed here:

1. The pixel type issue with writing float64 using SCIFIO. It's true that float64 TIFFs are not super compatible with things, including ImageJ1. And you cannot write float64 PNGs. But this issue is unrelated to the issue at hand here.

2. Color image conversion is another separate issue; see Add axis convention handling to conversion functions #17 for now to track that.

3. Static typing in Python is counterintuitive to Python programmers, who expect dynamic typing. But that's just how pyjnius is. We don't have the bandwidth to change that.

4. Automagical conversion between Python and Java would be amazing, especially when passing a Python object to a Java method call e.g. WindowManager.setTempCurrentImage(my_numpy_array) that could implicitly expand to WindowManager.setTempCurrentImage(ij.py.to_java(my_numpy_array)). But this is tricky for a couple of reasons: A) how to inject the ij context?; B) how to alter pyjnius's Java method invocation behavior at a low level like that?. For now, we'll not attempt to make this work.