[BUG] nvtabular input datatypes are wrong

[bschifferer]

Describe the bug Although some columns are float32, NVTabular defines all input datatypes as int32. This causes error when deploying to Triton Inference Server

Run criteo example inlcuding the notebook 04-Triton-Inference-with-TF

[karlhigley]

Which fields aren't supposed to be int32 and what error does it cause?

[bschifferer]

fields starting with I are numerical features and should be float32.

I got errors, when I tried using perf_analyzer with custom data. I think I am not sure, how Triton handles the normal requests, as in the notebook, when we send a float32 to an features with int32

[EvenOldridge]

@bschifferer is this still broken?

[bschifferer]

I think it is fixed