T3 DNN Draft PR + Multi-class Architecture

[GNiendorf]

This PR introduces a T3 DNN with a multi-class classification architecture that outputs a softmax distribution for T3 candidates that are fake (0), real and prompt (1), and real and displaced (2). By placing an explicit cut on displaced t3 candidates and prompt t3 candidates, I can reject a lot of junk t3 objects at build-time without substantially decreasing the displaced track efficiency.

[GNiendorf]

/run all

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     45.5    394.2    185.7    150.2    146.2    549.1    123.0    234.3    152.1      3.5    1983.9    1389.3+/- 383.1     525.2   explicit[s=4] (target branch)
   avg     45.4    396.6    186.2    183.8     33.4    545.1    120.8     83.5    148.2      3.0    1745.9    1155.4+/- 264.2     468.2   explicit[s=4] (this PR)

[github-actions]

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

The full set of validation and comparison plots can be found here.

[GNiendorf]

Testing a new architecture type. Timing benefit is not as good, but displaced track efficiency loss diminishes.

/run all

[github-actions]

There was a problem while building and running in standalone mode. The logs can be found here.

[GNiendorf]

/run all

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     46.4    393.2    185.6    150.9    148.0    548.1    122.8    233.1    152.5      3.3    1983.9    1389.4+/- 387.3     527.3   explicit[s=4] (target branch)
   avg     46.1    396.3    187.0    181.3     52.4    507.5    123.1    135.3    145.4      3.6    1778.0    1224.4+/- 295.9     473.6   explicit[s=4] (this PR)

[github-actions]

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

The full set of validation and comparison plots can be found here.

[GNiendorf]

/run all

[GNiendorf]

This PR Timing on L40

Current Timing on L40

[GNiendorf]

Looks like there's some loss at high pT from updating the occupancies. @YonsiG Do you know what percentile the T3 occupancies were set at? Right now I have them all at 99.9%

Edit: Removed the occupancy update, will figure this out first in a separate PR.

[slava77]

@jchismar
you may want to test the T4 on top of this PR: perhaps the issue with fakes will be more manageable