Skip to content

Commit

Permalink
refactor inference model
Browse files Browse the repository at this point in the history
  • Loading branch information
@mafrahm
mafrahm committed Oct 16, 2023
1 parent 02f0657 commit 4201c92
Show file tree
Hide file tree
Showing 4 changed files with 223 additions and 166 deletions.
2 changes: 1 addition & 1 deletion hbw/config/defaults_and_groups.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ def default_ml_model(cls, container, task_params):

# the ml_model parameter is only used by `MLTraining` and `MLEvaluation`, therefore use some default
# NOTE: default_ml_model does not work for the MLTraining task
if cls.task_family in ("cf.MLTraining", "cf.MLEvaulation", "cf.MergeMLEvents", "cf.PrepareMLEvents"):
if cls.task_family in ("cf.MLTraining", "cf.MLEvaluation", "cf.MergeMLEvents", "cf.PrepareMLEvents"):
# TODO: we might want to distinguish between two default ML models (sl vs dl)
default_ml_model = "dense_default"

Expand Down
21 changes: 17 additions & 4 deletions hbw/inference/constants.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,10 +16,23 @@
}
signals = {*signals_ggHH, *signals_qqHH}

# collection of all datasets (only 2017 ATM)
e_datasets = {f"data_e_{i}" for i in ["b", "c", "d", "e", "f"]}
mu_datasets = {f"data_mu_{i}" for i in ["b", "c", "d", "e", "f"]}
datasets = {*e_datasets, *mu_datasets}
# mapping between lepton categories and datasets (only 2017 ATM)
data_datasets = {
"1e": {f"data_e_{i}" for i in ["b", "c", "d", "e", "f"]},
"1mu": {f"data_mu_{i}" for i in ["b", "c", "d", "e", "f"]},
"2e": {"data_e_b"}, # TODO: 2 lep datasets in cmsdb + config
"2mu": {"data_mu_b"}, # TODO
"emu": {"data_mu_b"}, # TODO
}
merged_datasets = set().union(*data_datasets.values())

# mapping between process names in the config and inference model
inference_procnames = {
# key: config process name, value: inference model process name
"foo": "bar",
# "st": "ST",
# "tt": "TT",
}

# mapping, which processes are used for which QCDScale (rate) uncertainty
processes_per_QCDScale = {
Expand Down
172 changes: 11 additions & 161 deletions hbw/inference/default.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,10 +4,13 @@
hbw inference model.
"""

from columnflow.inference import inference_model, ParameterType, ParameterTransformation
from columnflow.config_util import get_datasets_from_process
from columnflow.inference import InferenceModel, inference_model
import hbw.inference.constants as const # noqa

from hbw.inference.functions import (
add_inference_categories, add_inference_processes, add_inference_parameters,
)

#
# Defaults for all the Inference Model parameters
#
Expand Down Expand Up @@ -134,172 +137,17 @@
@inference_model(
**default_cls_dict,
)
def default(self):
def default(self: InferenceModel):
"""
This is the default Inference model.
Idea: first build an inclusive Inference Model with all Channels/Processes/Systematics,
then remove anything not listed in the attributes.
"""
year = self.config_inst.campaign.x.year # noqa; not used right now
ecm = self.config_inst.campaign.ecm

#
# categories
#

# TODO: use ML model inst if possible
ml_model_processes = [
"ggHH_kl_1_kt_1_sl_hbbhww",
"tt",
"st",
"v_lep",
# "w_lnu",
# "dy_lep",
]

# if process names need to be changed to fit some convention
inference_procnames = {
"foo": "bar",
# "st": "ST",
# "tt": "TT",
}

for proc in ml_model_processes:
for lep in ("e", "mu"):
cat_name = f"cat_1{lep}_{proc}"
if cat_name not in self.channels:
continue

cat_kwargs = {
"config_category": f"1{lep}__ml_{proc}",
# "config_variable": f"mlscore.{proc}_rebin",
"config_variable": f"mlscore.{proc}_manybins",
"mc_stats": self.mc_stats,
}
if self.skip_data:
cat_kwargs["data_from_processes"] = self.processes
else:
cat_kwargs["config_data_datasets"] = [f"data_e_{i}" for i in ["b", "c", "d", "e", "f"]]

self.add_category(cat_name, **cat_kwargs) # noqa

# add processes with corresponding datasets to all categories of the inference model
used_datasets = set()
for proc in self.processes:
if not self.config_inst.has_process(proc):
raise Exception(f"Process {proc} not included in the config {self.config_inst.name}")

# get datasets corresponding to this process
datasets = [
d.name for d in
get_datasets_from_process(self.config_inst, proc, strategy="inclusive")
]

# check that no dataset is used multiple times
if datasets_already_used := used_datasets.intersection(datasets):
raise Exception(f"{datasets_already_used} datasets are used for multiple processes")
used_datasets |= set(datasets)

self.add_process(
inference_procnames.get(proc, proc),
config_process=proc,
is_signal=("HH_" in proc),
config_mc_datasets=datasets,
)

#
# parameters
#

# groups
self.add_parameter_group("experiment")
self.add_parameter_group("theory")

# lumi
lumi = self.config_inst.x.luminosity
for unc_name in lumi.uncertainties:
if unc_name not in self.systematics:
continue

self.add_parameter(
unc_name,
type=ParameterType.rate_gauss,
effect=lumi.get(names=unc_name, direction=("down", "up"), factor=True),
transformations=[ParameterTransformation.symmetrize],
)

# add QCD scale (rate) uncertainties to inference model
# TODO: combine scale and mtop uncertainties for specific processes?
# TODO: some scale/pdf uncertainties should be rounded to 3 digits, others to 4 digits
# NOTE: it might be easier to just take the recommended uncertainty values from HH conventions at
# https://gitlab.cern.ch/hh/naming-conventions instead of taking the values from CMSDB
for k, procs in const.processes_per_QCDScale.items():
syst_name = f"QCDScale_{k}"
if syst_name not in self.systematics:
continue

for proc in procs:
if proc not in self.processes:
continue
process_inst = self.config_inst.get_process(proc)
if "scale" not in process_inst.xsecs[ecm]:
continue
self.add_parameter(
syst_name,
process=inference_procnames.get(proc, proc),
type=ParameterType.rate_gauss,
effect=tuple(map(
lambda f: round(f, 3),
process_inst.xsecs[ecm].get(names=("scale"), direction=("down", "up"), factor=True),
)),
)
self.add_parameter_to_group(syst_name, "theory")

# add PDF rate uncertainties to inference model
for k, procs in const.processes_per_pdf_rate.items():
syst_name = f"pdf_{k}"
if syst_name not in self.systematics:
continue

for proc in procs:
if proc not in self.processes:
continue
process_inst = self.config_inst.get_process(proc)
if "pdf" not in process_inst.xsecs[ecm]:
continue

self.add_parameter(
f"pdf_{k}",
process=inference_procnames.get(proc, proc),
type=ParameterType.rate_gauss,
effect=tuple(map(
lambda f: round(f, 3),
process_inst.xsecs[ecm].get(names=("pdf"), direction=("down", "up"), factor=True),
)),
)
self.add_parameter_to_group(syst_name, "theory")

for shape_uncertainty, shape_processes in const.processes_per_shape.items():
if shape_uncertainty not in self.systematics:
continue

# If "all" is included, takes all processes except for the ones specified (starting with !)
if "all" in shape_processes:
_remove_processes = {proc[:1] for proc in shape_processes if proc.startswith("!")}
shape_processes = set(self.processes) - _remove_processes

self.add_parameter(
shape_uncertainty,
process=shape_processes,
type=ParameterType.shape,
config_shift_source=const.source_per_shape[shape_uncertainty],
)

is_theory = "pdf" in shape_uncertainty or "murf" in shape_uncertainty
if is_theory:
self.add_parameter_to_group(syst_name, "theory")
else:
self.add_parameter_to_group(syst_name, "experiment")
add_inference_categories(self)
add_inference_processes(self)
add_inference_parameters(self)

#
# post-processing
Expand Down Expand Up @@ -336,5 +184,7 @@ def default(self):
"lumi_13TeV_2017",
]

cls_dict["ml_model_name"] = "dense_test"

# minimal model for quick test purposes
test = default.derive("test", cls_dict=cls_dict)
Loading

0 comments on commit 4201c92

Please sign in to comment.