Skip to content

Commit

Permalink
improving performance in hiv module (#830)
Browse files Browse the repository at this point in the history 
* "profiling for 1 month"

* "transfer scale and shape calculation out od the loop"

* "calculate scale and shape with linear model for all ages"

* "uses linear model to calculate mean months between aids and death depending on age"

* "modifies get_time_from_infection_to_aids to do vectorised calculations"

* "modifies  if statement in get_time_from_infection_to_aids to avoid tests failing but adds redundancy"

* "removes redundant comments"

* Update hiv.py

* Update hiv.py

* "updates formatting"

* Update hiv.py

* Update src/scripts/profiling/scale_run.py

Co-authored-by: Matt Graham <[email protected]>

* Update src/scripts/profiling/scale_run.py

Co-authored-by: Matt Graham <[email protected]>

* "restructures get_time_from_infection_to_aids into two seperate functions: 1) linear model parameter calculations, 2) sampling"

* "changes transformation to list"

* Avoid repeated dataframe access

* Add comment explaining shape param special case

* Resample only necessary subset + rename functions

---------

Co-authored-by: Matt Graham <[email protected]>
  • Loading branch information
@dimitrasal @matt-graham
dimitrasal and matt-graham authored Apr 19, 2023
1 parent a5912b3 commit 3c011e7
Showing 1 changed file with 83 additions and 94 deletions.
177 changes: 83 additions & 94 deletions src/tlo/methods/hiv.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -452,7 +452,9 @@ def pre_initialise_population(self):
"age_years",
conditions_are_mutually_exclusive=True,
conditions_are_exhaustive=True)
.when("<20", p["infection_to_death_weibull_scale_1519"])
.when("==0", p["mean_survival_for_infants_infected_prior_to_birth"])
.when(".between(1,4)", p["infection_to_death_infant_infection_after_birth_weibull_scale"])
.when(".between(5, 19)", p["infection_to_death_weibull_scale_1519"])
.when(".between(20, 24)", p["infection_to_death_weibull_scale_2024"])
.when(".between(25, 29)", p["infection_to_death_weibull_scale_2529"])
.when(".between(30, 34)", p["infection_to_death_weibull_scale_3034"])
Expand All @@ -467,7 +469,9 @@ def pre_initialise_population(self):
"age_years",
conditions_are_mutually_exclusive=True,
conditions_are_exhaustive=True)
.when("<20", p["infection_to_death_weibull_shape_1519"])
.when("==0", 1) # Weibull with shape=1 equivalent to exponential distribution
.when(".between(1,4)", p["infection_to_death_infant_infection_after_birth_weibull_shape"])
.when(".between(5, 19)", p["infection_to_death_weibull_shape_1519"])
.when(".between(20, 24)", p["infection_to_death_weibull_shape_2024"])
.when(".between(25, 29)", p["infection_to_death_weibull_shape_2529"])
.when(".between(30, 34)", p["infection_to_death_weibull_shape_3034"])
Expand All @@ -477,6 +481,16 @@ def pre_initialise_population(self):
.when(">= 50", p["infection_to_death_weibull_shape_4549"])
)

# -- Linear Model to give the mean months between aids and death depending on age
self.lm["offset_parameter_for_months_from_aids_to_death"] = LinearModel.multiplicative(
Predictor(
"age_years",
conditions_are_mutually_exclusive=True,
conditions_are_exhaustive=True)
.when("<5", p["mean_months_between_aids_and_death_infant"])
.when(">=5", p["mean_months_between_aids_and_death"])
)

# -- Linear Models for the Uptake of Services
# Linear model that give the increase in likelihood of seeking a 'Spontaneous' Test for HIV
# condition must be not on ART for test
Expand Down Expand Up @@ -840,23 +854,27 @@ def initialise_simulation(self, sim):
# AIDS Onset Event for those who are infected but not yet AIDS and have not ever started ART
# NB. This means that those on ART at the start of the simulation may not have an AIDS event --
# like it happened at some point in the past

for person_id in before_aids_idx:
# get days until develops aids, repeating sampling until a positive number is obtained.
days_until_aids = 0
while days_until_aids <= 0:
days_since_infection = (
self.sim.date - df.at[person_id, "hv_date_inf"]
).days
days_infection_to_aids = np.round(
(self.get_time_from_infection_to_aids(person_id)).months * 30.5
scale, shape, offset = self.get_time_from_infection_to_aids_distribution_parameters(before_aids_idx)
days_infection_to_aids = self.sample_time_from_infection_to_aids_given_parameters(scale, shape, offset)
days_since_infection = (self.sim.date - df.loc[before_aids_idx, "hv_date_inf"])
# If any days_since_infection >= days_infection_to_aids are negative resample
# these values until all are positive
days_until_aids_is_negative = days_since_infection >= days_infection_to_aids
while np.any(days_until_aids_is_negative):
days_infection_to_aids[days_until_aids_is_negative] = (
self.sample_time_from_infection_to_aids_given_parameters(
scale[days_until_aids_is_negative],
shape[days_until_aids_is_negative],
offset[days_until_aids_is_negative],
)
days_until_aids = days_infection_to_aids - days_since_infection

date_onset_aids = self.sim.date + pd.DateOffset(days=days_until_aids)
)
days_until_aids_is_negative = days_since_infection >= days_infection_to_aids
days_until_aids = days_infection_to_aids - days_since_infection
date_onset_aids = self.sim.date + pd.to_timedelta(days_until_aids, unit='D')
for person_id, date in zip(before_aids_idx, date_onset_aids):
sim.schedule_event(
HivAidsOnsetEvent(person_id=person_id, module=self, cause='AIDS_non_TB'),
date=date_onset_aids,
date=date,
)

# Schedule the AIDS death events for those who have got AIDS already
Expand Down Expand Up @@ -1153,93 +1171,64 @@ def do_new_infection(self, person_id):
df.at[person_id, "hv_date_inf"] = self.sim.date

# Schedule AIDS onset events for this person
date_onset_aids = self.sim.date + self.get_time_from_infection_to_aids(
person_id=person_id
parameters = self.get_time_from_infection_to_aids_distribution_parameters(
[person_id]
)
date_onset_aids = (
self.sim.date
+ self.sample_time_from_infection_to_aids_given_parameters(*parameters)
).iloc[0]
self.sim.schedule_event(
event=HivAidsOnsetEvent(self, person_id, cause='AIDS_non_TB'), date=date_onset_aids
event=HivAidsOnsetEvent(self, person_id, cause='AIDS_non_TB'),
date=date_onset_aids,
)

def get_time_from_infection_to_aids(self, person_id):
"""Gives time between onset of infection and AIDS, returning a pd.DateOffset.
For those infected prior to, or at, birth: (this is a draw from an exponential distribution)
For those infected after birth but before reaching age 5.0 (this is drawn from a weibull distribution)
For adults: (this is a drawn from a weibull distribution (with scale depending on age);
* NB. It is further assumed that the time from aids to death is 18 months.
def sample_time_from_infection_to_aids_given_parameters(self, scale, shape, offset):
"""Generate time(s) between onset of infection and AIDS as Pandas time deltas.
The times are generated from translated Weibull distributions discretised to
an integer number of months.
:param scale: Scale parameters of Weibull distributions (unit: years).
:param shape: Shape parameters of Weibull distributions.
:param offset: Offset to (negatively) shift Weibull variable by (unit: months).
:return: Generated time deltas.
"""

df = self.sim.population.props
age = df.at[person_id, "age_exact_years"]
p = self.parameters
months_to_death = self.rng.weibull(shape) * scale * 12
months_to_aids = np.round(months_to_death - offset).clip(0).astype(int)

if age == 0.0:
# The person is infected prior to, or at, birth:
months_to_death = int(self.rng.exponential(
scale=p["mean_survival_for_infants_infected_prior_to_birth"]
)
* 12,
)
return pd.to_timedelta(months_to_aids * 30.5, unit='D')

months_to_aids = int(
max(
0.0,
np.round(
months_to_death
- self.parameters["mean_months_between_aids_and_death_infant"]
),
)
)
elif age < 5.0:
# The person is infected after birth but before age 5.0:
months_to_death = int(
max(
0.0,
self.rng.weibull(
p[
"infection_to_death_infant_infection_after_birth_weibull_shape"
]
)
* p["infection_to_death_infant_infection_after_birth_weibull_scale"]
* 12,
)
)
months_to_aids = int(
max(
0.0,
np.round(
months_to_death
- self.parameters["mean_months_between_aids_and_death_infant"]
),
)
)
else:
# The person is infected after age 5.0
# - get the shape parameters (unit: years)
scale = (
self.lm["scale_parameter_for_infection_to_death"].predict(
self.sim.population.props.loc[[person_id]]
).values[0]
)
# - get the scale parameter (unit: years)
shape = (
self.lm["shape_parameter_for_infection_to_death"].predict(
self.sim.population.props.loc[[person_id]]
).values[0]
)
# - draw from Weibull and convert to months
months_to_death = self.rng.weibull(shape) * scale * 12
# - compute months to aids, which is somewhat shorter than the months to death
months_to_aids = int(
max(
0.0,
np.round(
months_to_death
- self.parameters["mean_months_between_aids_and_death"]
),
)
)
def get_time_from_infection_to_aids_distribution_parameters(self, person_ids):
"""Compute per-person parameters of distribution of time from infection to aids.
Evaluates three linear models which output age specific scale, shape and offset
parameters for the (translated) Weibull distribution used to generate the time
from infection to aids for an individual.
For those infected prior to, or at, birth, a Weibull distribution with shape
parameter 1 (equivalent to an exponential distribution) is used.
return pd.DateOffset(months=months_to_aids)
For those infected after birth a Weibull distribution with both shape and
scale depending on age is used.
:param person_ids: Iterable of ID indices of individuals to get parameters for.
:return: Per-person parameters as a 3-tuple ``(scale, shape, offset)`` of
``pandas.Series`` objects.
"""
subpopulation = self.sim.population.props.loc[person_ids]
# get the scale parameters (unit: years)
scale = self.lm["scale_parameter_for_infection_to_death"].predict(subpopulation)
# get the shape parameter
shape = self.lm["shape_parameter_for_infection_to_death"].predict(subpopulation)
# get the mean months between aids and death (unit: months)
offset = self.lm["offset_parameter_for_months_from_aids_to_death"].predict(
subpopulation
)
return scale, shape, offset

def get_time_from_aids_to_death(self):
"""Gives time between onset of AIDS and death, returning a pd.DateOffset.
Expand Down

0 comments on commit 3c011e7

Please sign in to comment.