Problem when running 3D model on 2-channel stack

[lcferme]

Hello,

I have trained a 3D model on 2-channel images, but I have problems when I try to run it on a new dataset having similar shape zyxc, (50, 233, 257, 2).
Here is the code I am using:

import os
import sys
import numpy as np
from tqdm import tqdm
from glob import glob
from csbdeep.utils import Path, normalize
from stardist.models import StarDist3D
from tifffile import imread, imsave

Z = sorted(glob(pathname+"/*.tif"))
Z = list(map(imread,Z))

n_channel = 1 if Z[0].ndim == 3 else Z[0].shape[1]
axis_norm = (0,1,2)   # normalize channels independently
if n_channel >1:
      if Z[0].shape[:-1]!=n_channel:
        Z = [np.moveaxis(z, 1, -1) for z in Z]
        n_tiles = (n_tiles_Z, n_tiles_Y, n_tiles_X, n_channel)

model = StarDist3D(None, name=model_name, basedir=model_dir)

#Normalization
Z = [normalize(z, 1,99.8) for z in Z]


# run prediction
n_tiles_Z =  1
n_tiles_Y =  4
n_tiles_X =  4
n_tiles = (n_tiles_Z, n_tiles_Y, n_tiles_X)

for i in range(len(Z)):
    labels, polygons = model.predict_instances(Z[i], axes="ZYXC", n_tiles = n_tiles)

And here is where I get the error messages:

Traceback (most recent call last):
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 674, in _axes_tile_overlap
    self._tile_overlap
AttributeError: 'StarDist3D' object has no attribute '_tile_overlap'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "C_prediction_2C.py", line 136, in <module>
    labels, polygons = model.predict_instances(Z[i], axes="ZYXC", n_tiles = n_tiles)
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 419, in predict_instances
    prob, dist = self.predict(img, axes=axes, normalizer=normalizer, n_tiles=n_tiles, show_tile_progress=show_tile_progress, **predict_kwargs)
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 322, in predict
    axes_net_tile_overlaps = self._axes_tile_overlap(axes_net)
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 676, in _axes_tile_overlap
    self._tile_overlap = self._compute_receptive_field()
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 668, in _compute_receptive_field
    return [(m-np.min(i), np.max(i)-m) for (m,i) in zip(mid,ind)]
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\stardist\models\base.py", line 668, in <listcomp>
    return [(m-np.min(i), np.max(i)-m) for (m,i) in zip(mid,ind)]
  File "<__array_function__ internals>", line 6, in amin
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\numpy\core\fromnumeric.py", line 2793, in amin
    keepdims=keepdims, initial=initial, where=where)
  File "C:\Users\Lucrezia\stardist\stardist_venv\lib\site-packages\numpy\core\fromnumeric.py", line 90, in _wrapreduction
    return ufunc.reduce(obj, axis, dtype, out, **passkwargs)
ValueError: zero-size array to reduction operation minimum which has no identity

I am not sure what I am doing wrong... by chance do you have a clue of what I should modify/correct?

Thank you a lot!!
Lucrezia

[maweigert]

Hi @lcferme ,

Might related to #121 . Could you try the following:

1. Close all running notebooks and restart the prediction notebook
2. print the shape of the input

print(Z[0].shape)

2. Increase the number of tiles to e.g. (2,8,8) and try predicting

[lcferme]

Hi @maweigert !

I am not running a notebook, just a script from the terminal. Everything works as long as I have one channel, but then if I try with two I get that error.
In this specific case, the shape of the image is (50, 233, 257, 2). As you see, it is quite small and I actually did not get any error once I give n_tiles = None in model.predict_instances()
For bigger files, what is the correct way to set up the tiles? Should I specify the number of channels , like n_tiles = (2, 8, 8, 2) for ZYXC or I only define ZYX?

Thank you for your help!!

[uschmidt83]

n_channel = 1 if Z[0].ndim == 3 else Z[0].shape[1]
if n_channel >1:
      if Z[0].shape[:-1]!=n_channel:
        Z = [np.moveaxis(z, 1, -1) for z in Z]

Note that Z[0].shape[:-1]!=n_channel is always True, since Z[0].shape[:-1] is a tuple and n_channel is an integer.

For bigger files, what is the correct way to set up the tiles? Should I specify the number of channels , like n_tiles = (2, 8, 8, 2) for ZYXC or I only define ZYX?

n_tiles must have an entry for each of the axes, but channels cannot be tiled, i.e. the entry for channels must be 1. E.g.:

model.predict_instances(Z[i], axes="ZYXC", n_tiles=(2,8,8,1))

[uschmidt83]

This works for me:

from stardist.models import StarDist3D, Config3D

config = Config3D(n_channel_in=2)
model = StarDist3D(config, 'my_model', basedir=None)

x = np.zeros((12,133,157,2), dtype=np.float32)
out = model.predict_instances(x, axes='ZYXC', n_tiles=(1,4,4,1))

[lcferme]

@uschmidt83 thanks, I corrected the mistake with the tuple :)
But I still have problems running the prediction on bigger files, for instance for a file of shape (193, 1699, 1938, 2) I get the same error.

[uschmidt83]

Just tried it with this shape, and it works on my GPU (8GB RAM):

from stardist.models import StarDist3D, Config3D

config = Config3D(n_channel_in=2)
model = StarDist3D(config, 'my_model', basedir=None)

x = np.zeros((193, 1699, 1938, 2), dtype=np.float32)
out = model.predict_instances(x, axes='ZYXC', n_tiles=(2,16,16,1))

Does this not run for you?