Issues with aray shape when using dask_image.imread (and dask.array.image.imread) vs. imageio.imread

[habi]

I'm loading a big bunch of tomographic data in a preview/analysis notebook.
As we keep scanning samples, the dataframe I put the preview images in gets larger and larger.
I've been using imageio.imread to load the preview images (middle axial slices and MIPs) from disk.
I'd like to switch to dask_image.imread for this, as I'm loading the full datasets with it and generate the preview files from the full stacks loaded like this.

I now saw that loading an image with imageio.imread returns an image with two coordinates (size of the image), while dask_image.imread (and dask.array.image.imread return an image with three coordinates, the first one being 1, the second and third being the size of the image.
I'm very well aware that I can just .squeeze() the array before displaying with matplotlib, but expect that all the imread functions return the same kind of array.

Minimal Complete Verifiable Example:

I've made a gist which shows my issue fully self-contained, it can be found here and can be started in Binder:

It boils down to

imgio = imageio.imread('random.png')
imgdask = dask.array.image.imread('random.png')
imgdaskimg = dask_image.imread.imread('random.png')

returning different shapes.

This is closely related to #229 :)

[GenevieveBuckley]

Thank you for the report (and binder example!) @habi

My best suggestion is to squeeze the array to remove the singleton dimension(s) if they're causing you problems.

import numpy as np

squeezed_imgdask = np.squeeze(imgdask)
squeezed_imgdask.shape
# (100, 100)

Since this has no effect if no singleton dimensions are present, you would be able to add this generally to your code. Then you'll get the same output, regardless of whether you happen to be using dask or not.

[habi]

My best suggestion is to squeeze the array to remove the singleton dimension(s) if they're causing you problems.

I did squeeze the array in the end, so it's all good :)
I just expected the same return as imageio, maybe that's something to keep in mind for the work in #229.

[GenevieveBuckley]

That's good to hear, thanks

[habi]

I'm again having an issue with this, with a fresh installation of imageio and dask_image in a new conda environment.
The versions are

dask-image                2021.12.0          pyhd8ed1ab_0    conda-forge
imageio                   2.16.1             pyhcf75d05_0    conda-forge

When I load an image (one of thousands :) with

img_imgio = imageio.imread(filename)
img_dask = dask_image.imread.imread(filename)
print(img_imgio.shape)
print(img_dask.shape)

I get (3072, 3072) for imageio and (1, 3072, 3072, 4) for dask_image.
Is there any way to force dask_image to read 'simple' PNGs as 8bit gray images?

[jakirkham]

Guessing that we are getting some RGBA or similar uint8 splitting of the last dimension. This can be fixed by viewing it as uint8. It will leave a singleton dimension behind (so (1, 3072, 3072, 1)), but we can use squeeze for both this and the first dimension as Genevieve had suggested above.

img.view(np.uint32).squeeze()

More broadly we are looking at moving over to imageio. Some discussion in issue ( #181 ) about this.

[habi]

Thanks for the comment @jakirkham!

The underlying issue is more that I'm using

for c, sample in Data.iterrows()):
    Reconstructions[c] = dask_image.imread.imread(os.path.join(sample['Folder'], '*rec*.png'))

to lazily load +10000 of images from disk (several samples with each a folder of +1000 reconstructions).

From these I then generate files as necessary (axial views and MIPs), but do not view them directly.
It seems to me that I have to switch everything to the 'pure dask' way mentioned in issue #181 above.