Bug Issue

The doc of keras.applications.vgg16.preprocess_input() shows its description as below:

For the repro below, which is expected to work well, I found a bug that keras.applications.vgg16.preprocess_input can't deal correctly with x in 3D and data_format=='channels_first', using tf 2.20.0 and keras latest:

Repro

import tensorflow as tf
import numpy as np
import keras
data_format = 'channels_first'
image = np.random.randint(0, 256, size=(224, 224, 3), dtype=np.uint8)
image = tf.convert_to_tensor(image, dtype=tf.float32)
image = tf.transpose(image, perm=[2, 0, 1])  # Transpose to channels_first
preprocessed_image = keras.applications.vgg16.preprocess_input(image, data_format=data_format)
print(preprocessed_image.shape)

Output

InvalidArgumentError: {{function_node __wrapped__AddV2_device_/job:localhost/replica:0/task:0/device:CPU:0}} Incompatible shapes: [3,224,224] vs. [1,3,1] [Op:AddV2] name: 

The related codes is here:

It should deal with data_format=='channels_first' conditionally(3D v.s. 4D) just like here:

Thanks for noting!