Is Image=>Image CNN possible with autokeras?

[andyDoucette]

Hi. :)

From what I read, autoKeras sounds amazingly powerful with a fairly simple API. I'm curious if it can be used to solve the problem I am tasked with. I'd like to make an auto-architected CNN that takes images of various sizes as input and outputs images of the same size as the input as output. This seems fairly straight-forward, as a CNN should be able to handle any image size (with a batchSize of 1, anyway). I'm struggling to find how to express this with autoKeras.

From what I see, the output needs to be either a Node or a Head, and I cannot find a Node or Head that represents a variable sized image. I could try and use a RegressionHead, but it seems that is a densely connected layer meant for regressing a specific number of outputs (for example, "count the number of peaches in this image"). What I'm trying to do is more like "Take an image of a peaches in a certain environment and convert it to an image of the same number of pears in the same environment". Just an analogy, but I hope you understand what I'm going for in the abstract. What are my options that still allow me to leverage AutoKera's amazing neural architecture search and hyperparameter optimization algorithms?