ValueError: axes don't match array

[CarlosNacher]

Hi, help please! :(

I try to create a simple model like

import segmentation_models_3D as sm
import os
sm.set_framework('tf.keras')
os.environ["KERAS_BACKEND"] = "tensorflow"
model1 = sm.Unet(backbone_name="resnet50", input_shape=(32, 96, 96, 1), encoder_weights="imagenet")

but it says ValueError: axes don't match array. I tried #8 to make os.environ["KERAS_BACKEND"] = "tensorflow" as above shows and nothing!

The tst_keras.py file works for me, I don't understand nothing! (I'm trying to run the code at the beggining in Visual Studio Code, .ipynb file).

Please :(

[CarlosNacher]

Update:

The problem is the number of channels. If I try input_shape=(32, 96, 96, 1) instead of input_shape=(32, 96, 96, 3) it works. Can't use Unet for single channel volumes?

[ZFTurbo]

As I remember, there is no imagenet weights for single channel input. It can be fixed in code, but currently not supported

[ZFTurbo]

Simplest solution is to triplicate input for model)

[CarlosNacher]

Simplest solution is to triplicate input for model)

@ZFTurbo

This is a possibility, but it limits my batch size to 3 times smaller. If there is no alternative, it is an option!

In the code of convert_imagenet_weights_to_3D_models.py code, isn't there an option to do this for a channel?

Thank you very much for your quick reply!

[ZFTurbo]

1. If you increase input from 1 to 3, the required memory for model almost won't change.
2. Yes, convert_imagenet_weights_to_3D_models.py - this code can be used to generate required weigths from 2D variant for 1 channel only.

[CarlosNacher]

@ZFTurbo

1. If you increase input from 1 to 3, the required memory for model almost won't change.

If it changes, it is in fact three times as much. If a data occupies 0.5GB, tripling the channels will occupy 1.5GB. But for the time being, this approximation can work for me.

2. Yes, convert_imagenet_weights_to_3D_models.py - this code can be used to generate required weigths from 2D variant for 1 channel only.

Well, I'll try to do it in the future when I have more time!

[Nexer8]

@ZFTurbo, could you please explain what should be modified in convert_imagenet_weights_to_3D_models.py, so that it is possible to generate weights from 2D variant for 1 channel only?

If I set shape_size_3D = (64, 64, 64, 1), I get: ValueError: Layer bn_data weight shape (1,) is not compatible with provided weight shape (3,).

In case I also set shape_size_2D = (224, 224, 1), I get: ValueError: Cannot assign value to variable ' bn_data/beta:0': Shape mismatch.The variable shape (1,), and the assigned value shape (3,) are incompatible.