ValueError on predicting after model is trained with analysis "patchwise-crop/grid" regime

[jumutc]

The pipeline looks like this:

# Create a pixel value normalization Subfunction for z-score scaling
sf_zscore = Normalization(mode="z-score")
sf_grayscale = Convert2Grayscale()

# Assemble Subfunction classes into a list
sf = [sf_grayscale, sf_zscore]
# Configure data augmentation part
aug = Data_Augmentation(cycles=2, scaling=True, rotations=True,
                        elastic_deform=True, mirror=True, brightness=True,
                        contrast=True, gamma=True, gaussian_noise=True)
aug.config_p_per_sample = 0.5
aug.config_contrast_range = (0.1, 10)
aug.config_brightness_range = (0.1, 10)
aug.config_gamma_range = (0.1, 10)
aug.config_gaussian_noise_range = (0.1, 1)
pp = Preprocessor(data_io, batch_size=14, subfunctions=sf,
                  prepare_subfunctions=True, prepare_batches=True,
                  data_aug=aug, analysis="patchwise-crop", patch_shape=(320, 320))

Original images are of much bigger size of course but when I do predict model.predict(sample_list[0:4]):

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-15-cae07e0f8070> in <module>
----> 1 model.predict(sample_list[0:4])

~\anaconda3\envs\bacteria_cfu\lib\site-packages\miscnn\neural_network\model.py in predict(self, sample_list, return_output, activation_output)
    165                 pred_batch = self.model.predict_on_batch(batch)
    166                 pred_list.append(pred_batch)
--> 167             pred_seg = np.concatenate(pred_list, axis=0)
    168             # Postprocess prediction
    169             pred_seg = self.preprocessor.postprocessing(sample, pred_seg,

<__array_function__ internals> in concatenate(*args, **kwargs)

ValueError: need at least one array to concatenate

I have checked both analysis regimes and the ValueError appears in both.

[muellerdo]

Hey @jumutc,

MIScnn does not automatically resizes the images if they are smaller than the desired patch shape.

Problem:

• Image A shape (400, 400)
• Image B shape (200, 400)
• Desired input shape (320, 320): A->(320,320), B->(200,320)
• Error because he can not combine them to a batch or image is not equal to given patch size

Side node:
If running patchwise-crop training, there could be no exceptions thrown due to the random cropping of batchgenerators will automatically pad an image if it is smaller than patch size. https://github.com/MIC-DKFZ/batchgenerators/blob/master/batchgenerators/augmentations/crop_and_pad_augmentations.py

Debugging:
Please try to output your image sizes (after preprocessing) and verify if they are all bigger or equal than your patch size.

Fast Solution:

• Try adding the resize or padding subfunction to ensure patch size.
• (I would recommend padding with minimum mode due to it does not destroy aspect ratios of images)

Cheers, Dominik

[jumutc]

@muellerdo all images are much larger than 320x320 and the sizes are >1k for each side. Can it be that because of different sizes of different images we have a problem in batchgenerators? I also noticed that in batches folder everything is generated for predictions for being used so something is wrong in data_generator.py most probably.

[jumutc]

I tried to debug this case and the problem indeed is in that class:

from miscnn.neural_network.data_generator import DataGenerator

dataGen = DataGenerator(sample_list[:1], model.preprocessor,
                        training=False, validation=False,
                        shuffle=False, iterations=None)
print(len(dataGen))
print(dataGen[0])

with output being:

0
[[[[-2.0725093  -2.0371544  -2.0371544 ]
   [-2.0017996  -1.9310896  -2.0017996 ]
   [-2.1432192  -2.0017996  -2.3907037 ]
   ...
   [-0.7643772  -0.870442   -1.1532813 ]
   [-0.9765067  -0.870442   -1.1179265 ]
   [-0.9411518  -1.0118617  -1.0825715 ]]]]