In Stardist2D number_of_steps is calculated assuming all training data has the same size

Describe the bug
Although the training works even if the input images are of different widths or heights, number_of_steps is calculated using the dimensions of a randomly selected image.

To Reproduce
Steps to reproduce the behavior:

1. open StarDist_2D_ZeroCostDL4Mic.ipynb
2. Scroll down to 'section 4.1 "Prepare the training data and model for training"
3. See code:
   number_of_steps = int(Image_XImage_Y/(patch_sizepatch_size))*(int(len(X)/batch_size)+1)
   if (Use_Data_augmentation):
   augmentation_factor = Multiply_dataset_by
   number_of_steps = number_of_steps * augmentation_factor

Where Image_X and Image_Y are the dimensions of a random image

Expected behavior
I suggest to replace it with the following simple fix that loop through the input images and exactly calculates the number of steps:

number_of_steps = calc_number_of_steps(X)

Where calc_number_of_steps is:
import math
def calc_number_of_steps(X):
num_steps = 0

for x in X:
assert x.shape[0] >= patch_size and x.shape[1] >= patch_size , "patch size must not be greater that smallest image dimension"
num_steps += int(x.shape[0]/patch_size) * int(x.shape[1]/patch_size)

num_steps = math.ceil(num_steps/batch_size)
if (Use_Data_augmentation):
augmentation_factor = Multiply_dataset_by
num_steps *= augmentation_factor
return num_steps

Open StarDist_2D_ZeroCostDL4Mic.ipynb and go to section 4.1, then inspect how number_of_steps uses Image_X and Image_Y. Try training data with different image dimensions, update the calculation so all input images are accounted for, and verify that the resulting step count and patch-size validation behave as expected.