my-unet-keras/notes

Save slice indexes at start when cutting up the image, so you can just
use those later when stitching them together.

Use a batch size of 1 on cpu.

Increase number of epochs when using larger batch size on GPU.

I did a few tests and found that increased batch size reduces accuracy on cpu. So I should increase the number of epochs to compensate on GPU. It's fine for GPU because of parallelization. 

imread_collection("*.png") seems to be a simple way to read all in a folder?

https://scikit-image.org/docs/dev/user_guide/getting_started.html  simpler

>>> mask = camera < 87
>>> # Set to "white" (255) the pixels where mask is True
>>> camera[mask] = 255

img.mean() returns the mean value. could be a better binarization metric?

to fill the whole bitfield.
>>> from skimage import exposure
>>> image = exposure.rescale_intensity(img10bit, in_range=(0, 2**10 - 1))
>>> image = exposure.rescale_intensity(img10bit, in_range='uint10')

https://scikit-image.org/docs/dev/user_guide/data_types.html
"You should never use astype on an image, because it violates these assumptions about the dtype range:"
>>> image = exposure.rescale_intensity(img_int32, out_range=(0, 2**31 - 1))
>>> img_uint8 = img_as_ubyte(image)


Using an image from OpenCV with skimage¶

If cv_image is an array of unsigned bytes, skimage will understand it by default. If you prefer working with floating point images, img_as_float() can be used to convert the image:
>>>

>>> from skimage.util import img_as_float
>>> image = img_as_float(any_opencv_image)

Using an image from skimage with OpenCV

The reverse can be achieved with img_as_ubyte():

These can replace my cutImage functions.
np.pad  !!! NOTE: skimage.util.pad is a wrapper for np.pad that appears in ancient code. !!!
skimage.util.view_as_windows(arr_in, window_shape, step=1)

numpy.vstack to take a list of arrays and stack them into one big array!