Avoid TensorFlow empty inputs in central interface

[valsdav]

Empty Tensors passed to the TensorFlow session may cause crashes on some architectures while returning silently empty results on others. See #44333

An explicit check on input tensors in the TensorFlow interface can prevent the issue once for all here:

• do not run the inference if input.NumElements() == 0
• Print a warning to the framework

A PR will follow, any thoughts about this?

[valsdav]

assign ml

[cmsbuild]

New categories assigned: ml

@valsdav,@wpmccormack you have been requested to review this Pull request/Issue and eventually sign? Thanks

[cmsbuild]

cms-bot internal usage

[cmsbuild]

A new Issue was created by @valsdav.

@smuzaffar, @Dr15Jones, @rappoccio, @sextonkennedy, @antoniovilela, @makortel can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[makortel]

• Print a warning to the framework

How necessary a warning is? Is the expectation that users would also avoid TF inference calls when input is empty, or that only the central TF interface does that?

I mean, an empty input (no objects) sounds completely valid input from physics side, so I'm wondering if the TF inference call could be just omitted silently in those cases?

[valsdav]

My expectation is that checks should be put in producers to avoid calling the inference for empty inputs, and that the guard in the central TF code is a more a protection.

Returning empty results would be correct for empty inputs, but if we do it silently we wouldn't notice problems (like in the DeepTau case) when the producer is having unexpected empty inputs and not having protections. What do you think about this?

I can do a quick survey of the producers using TF to check is protections are in place and add them also there.