Handling image dimensions for RetinaNet

[shilpakancharla]

In order for RetinaNet to work, the image shape must be at least 512 by 512. Anything smaller than this, though divisible by 16, will not work. In the below Colab, I've tried out several different image shapes which have some dimension error regarding the image shape. I typically need to perform some kind of image augmentation in order to get the image shape to pass through the model.

It would be good to perhaps have an error message regarding this, telling users specifically how the image should be formatted.

Here are examples of the different kinds of errors I've run into, and the shapes I've tried to build the RetinaNet model with:

• [2, 960, 1280, 3] → unsuccessful build, there is a shape mismatch

• [2, 950, 1280, 3] → unsuccessful build, there seems to be some kind of error with the bounding box?

Here's a link to the Colab where I built this model: https://colab.sandbox.google.com/drive/1iXZJ9GYb4cineDhaFV534Mjmi29TLLCg#scrollTo=02cA_lsyQjmz

[LukeWood]

Thanks for the bug report @shilpakancharla ! I'll add better error messages to the RetinaNet to remedy this issue.