Update _torchscript_model_adapter.py

[esgomezm]

converting the input to the model to float to avoid numpy array errors from torch

@oeway

[FynnBe]

I don't think this should be hardcoded. Maybe a model expects a tensor to have an integer datatype (e.g. a mask or index input).
If a model needs float input then it should specify that for that tensor here (inputs[i].data = 'float32).
(Internally the ensure_dtype processing step is then used to cast if necessary after pre- or before postprocessing.

[FynnBe]

should not be hardcoded IMO

[oeway]

I agree it should not be directly converted to float32, however, it should not leave it as is either since pytorch does not support all numpy types, it seems all the unsigned types cause error, so we can either do float32 right now, or explicitly write a mapping for unsupported types.

[FynnBe]

But data.type is a required field. So there is always a known, fixed data type at this point in the code.
(field data_type for v0_4 models).
Could you please point me to the example where you encountered the issue?