Keras fine-tuning InceptionV3 tensor dimension error

Question

I am trying to fine-tune a model in Keras:

    inception_model = InceptionV3(weights=None, include_top=False, input_shape=(150, 
150, 1))

    x = inception_model.output
    x = GlobalAveragePooling2D()(x)
    x = Dense(256, activation='relu', name='fc1')(x)
    x = Dropout(0.5)(x)
    predictions = Dense(10, activation='softmax', name='predictions')(x)
    classifier = Model(inception_model.input, predictions)


    ####training training training ... save weights


    classifier.load_weights("saved_weights.h5")
  
    classifier.layers.pop()
    classifier.layers.pop()
    classifier.layers.pop()
    classifier.layers.pop()
    ###enough poping to reach standard InceptionV3 

    x = classifier.output
    x = GlobalAveragePooling2D()(x)
    x = Dense(256, activation='relu', name='fc1')(x)
    x = Dropout(0.5)(x)
    predictions = Dense(10, activation='softmax', name='predictions')(x)
    classifier = Model(classifier.input, predictions)

But I get the error:

ValueError: Input 0 is incompatible with layer global_average_pooling2d_3: expected ndim=4, found ndim=2

Answer 1

You shouldn't use pop() method on models created using functional API (ie keras.models.Model ). Only Sequential models (ie keras.models.Sequential ) have a built-in pop() method (usage: model.pop() ). Instead, use index or the names of the layers to access a specific layer:

classifier.load_weights("saved_weights.h5")
x = classifier.layers[-5].output   # use index of the layer directly
x = GlobalAveragePooling2D()(x)