Run-time Accuracy Debugging ("Debug Tensors")

[hgaspar]

It should be possible to enable run-time accuracy debugging, i.e. inspection of values of a tensor, for the purpose of detecting 0s, or NaNs, or any other user-specified condition, for example specified via user provided code (e.g. a lambda) that can execute on the tensor.

It should be possible to disable this functionality at runtime, and in that case, there should be no (or at least minimal) performance penalty, compared to a model compilation with this capability not enabled.

When the user specified condition is satisfied, e.g. a 0, or all zeros in a row, or a NaN is detected in a tensor, a typical flow could be to copy the data to a CPU buffer, and terminate executing the frame.