Why is Keras complaining about incompatible input shape in this case?

I have trained a Keras-based autoencoder model with the following input layer:

depth = 1
width = height = 100
input_shape = (height, width, depth)
inputs = Input(shape=input_shape)
# rest of network definition ...

Width and height of my training images were 100 pixels in grayscale, thus with a depth of 1. Now I want to load my trained model in another script, load an image there, resize and send it to the Keras model:

size = 100
image = cv2.imread(args.image, cv2.IMREAD_GRAYSCALE)
image = cv2.resize(image, (size, size), interpolation=cv2.INTER_AREA)
image = image.astype("float32") / 255.0
image = np.expand_dims(image, axis=-1)
# at this point image.shape = (100, 100, 1)
recon = autoencoder.predict(image)

However, the call to autoencoder.predict(image) leads to the following error:

WARNING:tensorflow:Model was constructed with shape (None, 100, 100, 1) for input KerasTensor(type_spec=TensorSpec(shape=(None, 100, 100, 1), dtype=tf.float32, name='input_1'), name='input_1', description="created by layer 'input_1'"), but it was called on an input with incompatible shape (None, 100, 1, 1).

I don’t understand this, as the shape of the image when calling predict() is (100, 100, 1), which looks fine to me. Why is Keras complaining about an incompatible input shape of (None, 100, 1, 1)?

Answer

These simple lines of code generate the error

X = np.random.uniform(0,1, (100,100,1))

inp = Input((100,100,1))
out = Dense(1)(Flatten()(inp))
model = Model(inp, out)

model.predict(X)

This is because your Keras model expects data in this format (n_sample, 100, 100, 1)

A simple reshape when you predict a single image does the trick

model.predict(X.reshape(1,100,100,1))