Adjust input output sizes when any dim is None

[buddhapuneeth]

Whenever inputs and outputs have None as dim (except for batch size), after conversion they will have 0 as the dimension.
Batch size will be converted to 'N'.
input(None, None) -> input(N, 0)

With this fix even other dims will have variables like M1, M2.
input(None, None) -> input(N, M1)

[CLAassistant]

All committers have signed the CLA.

[jiafatom]

Thanks for your efforts. The concern here is the benefit of showing symbolic dim in inputs. If the model shows M1, M2 as input dimension, it does not propagate to the output size, so it looks like an isolated dim. Batch size dim is different, since the output dim also shows N, so we know the corresponding dim in outputs.

[buddhapuneeth]

@jiafatom currently if we don't make it M1, M2.. the inputs and outputs where dim = None will be converted to 0. I know the parameterized dimension won't be propagated over the layers. I mean in some cases input's M1 might be same as output's M2. But without this fix, on the inference side we are having issues as it is expecting dim as 0.
Simple example (embedded layer) to illustrate this:

without this fix:

with fix:

Please provide your feedback/suggestion on this.

[wenbingl]

Thanks a lot for your contribution!