Float precision

[lacan]

Hi all,

I notice on things like care, Stardist and n2v that my images are opened as 64bit numpy arrays.
Whick leads me to the question: is TensorFlow in there using 64-bit float precision? Wouldn't 32 bit be enough and reduce the memory footprint by a factor of 2?
Thanks for the discussion!

Oli

[maweigert]

Hi Oli,

my images are opened as 64bit numpy arrays

Imo csbdeep/stardist should and will convert everything to float32 by default, as double precision is not necessary (and would be slower as you observe). Where do you see 64 bit floats generated?

E.g.

import numpy as np
from csbdeep.data import RawData, create_patches                       

x = np.random.randint(0,10,(10,100,100))                               
raw_data = RawData.from_arrays(x,x,axes = "YX")                        
X, Y, XY_axes = create_patches(raw_data, patch_size=(64,64), n_patches_per_image=10)                                                          

print(X.dtype)

dtype('float32')