Shape mismatch in input gradient - up_sampling2d/MLCUpsample2D. about tensorflow_macos

[anna-tikhonova]

Thank you very much for reporting this issue. Could you provide a full reproducible test case that we could run locally?

[Andreasgejlm]

Certainly. The following code reproduces the up_sampling2d error when run in the Tensorflow_macos venv:

import tensorflow as tf
import numpy as np
from tensorflow.keras import Model, Input, layers, optimizers, utils, losses
tf.compat.v1.disable_eager_execution()

training_data = np.random.rand(32, 256, 256, 3)
y_true = np.random.rand(32, 256, 256, 2)

val_data = np.random.rand(32, 256, 256, 3)

batch_size = 32

train_in = Input((256, 256, 3), name='Input', batch_size=batch_size)
x = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(train_in)
x = layers.MaxPooling2D((2, 2))(x)
x = layers.Conv2D(256, (3, 3), activation='relu', padding='same')(x)
x = layers.UpSampling2D()(x)
x = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(x)


real_branch = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(x)
real_branch = layers.Conv2D(1, (3, 3), activation='relu', padding='same', name="real")(real_branch)
imag_branch = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(x)
imag_branch = layers.Conv2D(1, (3, 3), activation='relu', padding='same', name="imag")(imag_branch)
concat = layers.concatenate([real_branch, imag_branch], axis=3)

model = Model(train_in, concat)

model.summary()
model.compile(optimizer=optimizers.Adam(), loss=losses.mean_squared_error, metrics=['acc'])
model.fit(training_data, y_true, epochs=3, batch_size=batch_size, validation_data=(val_data, y_true))

[anna-tikhonova]

Thank you for providing the code. We will investigate this issue and get back to you.

[pooyadavoodi]

@Andreasgejlm I just verified that this issue is resolved in versions >= v0.1alpha1 of our release. Could you please confirm this.