Non-fixed tensor shapes

I have a model with input tensor shape like this: (unk_1, 256, 256, 3). Is there anything special in creating the input tensor for this model? Can I replace the parameter (or the unknown part) with any number (e.g. 1, 2, etc.)? What about the output tensor with param/unknown shapes?

It indicates that one of the input has a unknown dimension.
Depending on your model that usually means that it works for what ever values of that dimension.
Being unknown means we cannot infer things at import and thus need to discover and evaluate things at run-time which will have a performance implication.
But apart from that, we will figure out things as we execute the model so things will work and your output will take the shape they are supposed to given the input

1 Like