Incorrect output dimension on ios about mobilepydnet

[FilippoAleotti]

Hi,
In the IROS2018 model the network predicts 2 disparities (ie, a left-aligned and a "virtual" right-aligned maps). If you are using this checkpoint, you have to slice the tensor

[l0stpenguin]

Yes i am using that checkpoint. I am not sure about how exactly i should proceed with slicing. I found only 1 output model/L0/ResizeBilinear and used that to freeze the graph. Are you suggesting that i should be modifying the output before freezing the graph?

[FilippoAleotti]

Yes, something like: tf.image.resize_images(self.disp2[:,:,:,0], size)

[l0stpenguin]

I am not experienced in neural networks, so i am quite confused. I don't manipulate any outputs nor do i run any inference in python. I just used the script in this post to export the pb file:

https://blog.metaflow.fr/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc

Since it's has only 1 output with 2 channels, i don't know how i am supposed to slice after freezing the graph.

[FilippoAleotti]

You have two options:

• edit the graph: by adding the lines I wrote before, you’re changing the graph. Doing so, you have to freeze this new graph
• keep the graph as it is: if you know that the predicted disparity has two channels, where just the first one is meaningful at testing time, you can slice the output in your app (that is, you keep just the first channel)