repo: https://github.com/lyazj/ad-part
architecture:
graph TD
library("ad-part library (libadhep)")
library --> definitions
library --> utilities
library --> Ntuple
library --> ParT
definitions --> addef
utilities --> adenv
utilities --> adfs
utilities --> admem
Ntuple --> adtype
Ntuple --> adbranch
Ntuple --> adnbranch
Ntuple --> adpdg
Ntuple --> adjet
ParT --> adtensor
ParT --> adrunner
graph TD
executable("ad-part executables")
executable --> test
executable --> core
test --> root-test
test --> delphes-test
test --> onnx-test
test --> test-more(...)
core --> pdg-gen
core --> flatten
core --> part
core --> part-gpu
core --> part-gen
core --> core-more(...)
adtensor architecture:
classDiagram
ADTensor
ADTensor <|-- ADPFTensor
ADTensor <|-- ADCFTensor
ADPFTensor <|-- ADPFPoints
ADPFTensor <|-- ADPFFeatures
ADPFTensor <|-- ADPFVectors
ADPFTensor <|-- ADPFMask
ADCFTensor <|-- ADSoftmax
class ADTensor {
Feature *data
int64_t size
int64_t *shape
int64_t shape_size
ADTensor(const int64_t *shape_in, int64_t shape_size_in)
tensor(const Ort::MemoryInfo &)
tensor(const Ort::MemoryInfo &, int64_t shape0)
get_name()*
}
class ADPFTensor {
ADPFTensor(int64_t nfeat)
get_name()*
set_value(int64_t n, const ADJet &)*
}
class ADCFTensor {
ADCFTensor()
get_name()*
get_value(int64_t n, ADParT &)*
}
class ADPFPoints {
ADPFPoints()
deta(int64_t n, int64_t p)
dphi(int64_t n, int64_t p)
get_name()*
set_value(int64_t n, const ADJet &)*
}
class ADPFFeatures {
ADPFFeatures()
log_pt(int64_t n, int64_t p)
log_e(int64_t n, int64_t p)
log_pt_rel(int64_t n, int64_t p)
log_e_rel(int64_t n, int64_t p)
delta_r(int64_t n, int64_t p)
charge(int64_t n, int64_t p)
is_charged_hadron(int64_t n, int64_t p)
is_neutral_hadron(int64_t n, int64_t p)
is_photon(int64_t n, int64_t p)
is_electron(int64_t n, int64_t p)
is_muon(int64_t n, int64_t p)
d0(int64_t n, int64_t p)
d0_err(int64_t n, int64_t p)
dz(int64_t n, int64_t p)
dz_err(int64_t n, int64_t p)
deta(int64_t n, int64_t p)
dphi(int64_t n, int64_t p)
get_name()*
set_value(int64_t n, const ADJet &)*
}
class ADPFVectors {
ADPFVectors()
px(int64_t n, int64_t p)
py(int64_t n, int64_t p)
pz(int64_t n, int64_t p)
e(int64_t n, int64_t p)
get_name()*
set_value(int64_t n, const ADJet &)*
}
class ADPFMask {
ADPFMask()
mask(int64_t n, int64_t p)
get_name()*
set_value(int64_t n, const ADJet &)*
}
class ADSoftmax {
ADSoftmax()
softmax(int64_t n, int64_t c)
get_name()*
get_value(int64_t n, ADParT &)*
}
design: split data preprocessing (flatten, normalize, etc.) and model running
- partially due to conveniences gained upon current third-party library interfaces
- enabling intermediate result checking, improving robustness and maintainability
- enabling a quick startup of the model, saving our time to tune or debug on it
flatten: from Delphes AT to flat Ntuple
main idea:
- In
ADJetconstructor, do RTTI for its constituents, routing to correspondingADParticleconstructor - Robustness: unrecognized objects are detected and warned
reference: https://github.com/delphes/delphes/blob/master/examples/Example3.C
ADJet(const ADPDGQuerier &, const Jet &);
ADParticle(const GenParticle &, const TLorentzVector &p4_jet, const ADPDGQuerier &);
ADParticle(const Track &, const TLorentzVector &p4_jet, const ADPDGQuerier &);
ADParticle(const Tower &, const TLorentzVector &p4_jet, const ADPDGQuerier &);
ADParticle(const Muon &, const TLorentzVector &p4_jet, const ADPDGQuerier &);part/part-gpu: from flat Ntuple to ParT scores
main idea:
- ONNX model exporting:
ParT_full.pt$\to$ part.onnx - run the model on GPU powered by ONNX Runtime