I trained a model with the input shape of (224, 224, 3) and I'm trying to change it to (300, 300, 3). For instance:
resnet50 = tf.keras.models.load_model(path_to_model)
model = tf.keras.models.Model([Input(shape=(300, 300, 3))], [resnet50.output])
# or
resnet50.inputs[0].set_shape([None, 300, 300, 3])
doesn't work.
I saw that the pretained model allows for different input shapes but adjusts the hole network architecture, for example, the size of the convolutional channels. I was wondering if I needed to do something similar or if for a trained model it is impossibel to change the input shape.
This would only work for convolutional layers as they do not care about input_shape because they are just sliding filters. However, if your model is trained on RGB images then also new_input shape should have 3 as channels.
Example:
first_model = VGG16(weights = None, input_shape=(224,224,3), include_top=False)
first_model.summary()
>> input_6 (InputLayer) [(None, 224, 224, 3)] 0
And second model:
new_input = tf.keras.Input((300,300,3))
x = first_model.layers[1](new_input) # First conv. layer
for new_layer in first_model.layers[2:]:
x = new_layer(x) # loop through layers using Functional API
second_model = tf.keras.Model(inputs=new_input, outputs=x)
second_model.summary()
>>
Layer (type) Output Shape Param #
=================================================================
input_9 (InputLayer) [(None, 300, 300, 3)] 0