Grad CAM Map With Memory Saving Looks Odd

[ghost]

@yaroslavvb Have you tried to create grad cam or saliency maps after using memory_saving_gradients? Mine look very odd (columnar). I'm doing the Udacity program, and my keras-vis grad-cam plots look like this:

[yaroslavvb]

I haven't....I can't really tell if this looks odd or not, the gradient results should be the same with and without memory saving

[ghost]

@yaroslavvb I've noticed my model runs faster, with less memory, but with errors when I use the tensorflow implementation of the keras backend (from tensorflow.python.keras._impl.keras import backend as K, or tensorflow.keras.backend in later versions of tensorflow) instead of the backend with my keras install (from keras import backend). The only difference between my code and your keras test code (keras_test.py) is that I use the keras functional model API, e.g.

a = Input(shape=(32,))
b = Dense(32)(a)
model = Model(inputs=a, outputs=b)

instead of the Sequential model (e.g.

model = Sequential()
model.add(Dense(32, activation='relu', input_shape=(32,))
model.add(Dense(16, activation='relu'))
model.compile()

Should I change my code to using the Sequential model if trying to use memory_saving_gradients?

[ghost]

@yaroslavvb FYI, here is what the cam map looks like, having changed nothing but the keras backend from from tensorflow.python.keras._impl.keras import backend as K to from keras import backend as K: