"""Creating and running simulation scenarios for TLOmodel.
Scenarios are used to specify, configure and run a single or set of TLOmodel
simulations. A scenario is created by subclassing ``BaseScenario`` and specifying the
scenario options therein. You can override parameters of the simulation modules in
various ways in the scenario. See the ``BaseScenario`` class for more information.
The subclass of ``BaseScenario`` is then used to create *draws*, which can be considered
a fully-specified configuration of the scenario, or a parameter draw.
Each draw is *run* one or more times - run is a single execution of the simulation. Each
run of a draw has a different seed but is otherwise identical. Each run has its own
simulation seed, introducing randomness into each simulation. A collection of runs for a
given draw describes the random variation in the simulation.
A simple example of a subclass of ``BaseScenario``::
class MyTestScenario(BaseScenario):
def __init__(self):
super().__init__(
seed = 12,
start_date = Date(2010, 1, 1),
end_date = Date(2011, 1, 1),
initial_population_size = 200,
number_of_draws = 2,
runs_per_draw = 2,
)
def log_configuration(self):
return {
'filename': 'my_test_scenario',
'directory': './outputs',
'custom_levels': {'*': logging.INFO}
}
def modules(self):
return [
demography.Demography(resourcefilepath=self.resources),
enhanced_lifestyle.Lifestyle(resourcefilepath=self.resources),
]
def draw_parameters(self, draw_number, rng):
return {
'Lifestyle': {
'init_p_urban': rng.randint(10, 20) / 100.0,
}
}
In summary:
* A *scenario* specifies the configuration of the simulation. The simulation start and
end dates, initial population size, logging setup and registered modules. Optionally,
you can also override parameters of modules.
* A *draw* is a realisation of a scenario configuration. A scenario can have one or more
draws. Draws are uninteresting unless you are overriding parameters. If you do not
override any model parameters, you would only have one draw.
* A *run* is the result of running the simulation using a specific configuration. Each
draw would run one or more times. Each run for the same draw will have an identical
configuration except the simulation seed.
"""
import abc
import datetime
import json
import pickle
from itertools import product
from pathlib import Path, PurePosixPath
from typing import Optional
import numpy as np
from tlo import Date, Simulation, logging
from tlo.analysis.utils import parse_log_file
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
MAX_INT = 2**31 - 1
[docs]
class BaseScenario(abc.ABC):
"""An abstract base class for creating scenarios.
A scenario is a configuration of a simulation. Users should subclass this class and
must implement the following methods:
* ``__init__`` - to set scenario attributes,
* ``log_configuration`` - to configure filename, directory and logging levels for
simulation output,
* ``modules`` - to list disease, intervention and health system modules for the
simulation.
Users may also optionally implement:
* ``draw_parameters`` - override parameters for draws from the scenario.
"""
def __init__(
self,
seed: Optional[int] = None,
start_date: Optional[Date] = None,
end_date: Optional[Date] = None,
initial_population_size: Optional[int] = None,
number_of_draws: int = 1,
runs_per_draw: int = 1,
resources_path: Path = Path("./resources"),
):
"""
:param seed: The top-level seed to use for generating per run simulation seeds
for random number generators.
:param start_date: Date to start simulation at.
:param end_date: Date to end simulation at.
:param initial_population_size: Number of individuals to initialise population
with.
:param number_of_draws: Number of draws (distinct parameter sets) over which to
run simulation.
:param runs_per_draw: Number of independent model runs to perform per draw.
:param resources_path: Path to the directory containing resource files.
"""
self.seed = seed
self.start_date = start_date
self.end_date = end_date
self.pop_size = initial_population_size
self.number_of_draws = number_of_draws
self.runs_per_draw = runs_per_draw
self.resources = resources_path
self.rng = None
self.scenario_path = None
[docs]
@abc.abstractmethod
def log_configuration(self, **kwargs):
"""Implementation must return a dictionary configuring logging.
Example::
return {
'filename': 'test_scenario',
'directory': './outputs',
'custom_levels': {
'*': logging.WARNING,
'tlo.methods.demography': logging.INFO
}
}
"""
[docs]
@abc.abstractmethod
def modules(self):
"""Implementation must return a list of instances of TLOmodel modules to
register in the simulation.
Example::
return [
demography.Demography(resourcefilepath=self.resources),
enhanced_lifestyle.Lifestyle(resourcefilepath=self.resources),
healthsystem.HealthSystem(
resourcefilepath=self.resources,
disable=True,d
service_availability=['*']
),
...
]
"""
[docs]
def draw_parameters(self, draw_number, rng):
"""Implementation must return a dictionary of parameters to override for each draw.
The overridden parameters must be scalar (i.e. float, integer or string) as the following examples demonstrate.
The argument ``draw_number`` and a random number generator are available, if required.
* Change a parameter to a fixed value: ``{'Labour': {'average_age_at_pregnancy': 25}}``
* Sample a value from a distribution: ``{'Lifestyle': {'init_p_urban': rng.randint(10, 20) / 100.0}}``
* Set a value based on the draw number: ``{'Labour': {'average_age_at_pregnancy': [25, 30, 35][draw_number]}}``
Implementation of this method in a subclass is optional. If no parameters are to be overridden,
returns ``None``. If no parameters are overridden, only one draw of the scenario is required.
A full example for a scenario with 10 draws::
return {
'Lifestyle': {
'init_p_urban': rng.randint(10, 20) / 100.0,
},
'Labour': {
'average_age_at_pregnancy': -10 * rng.exponential(0.1),
'some_other_parameter': np.arange(0.1, 1.1, 0.1)[draw_number]
},
}
:param int draw_number: the specific draw number currently being executed by the simulation engine
:param numpy.random.RandomState rng: the scenario's random number generator for sampling from distributions
"""
return None
[docs]
def get_log_config(self, override_output_directory=None):
"""Returns the log configuration for the scenario, with some post_processing."""
log_config = self.log_configuration()
# If scenario doesn't have log filename specified, we used the scenario script name
if "filename" not in log_config or log_config["filename"] is None:
log_config["filename"] = Path(self.scenario_path).stem
if override_output_directory is not None:
log_config["directory"] = override_output_directory
# If directory is specified, we always write log files - so don't print to stdout
if "directory" in log_config and log_config["directory"] is not None:
log_config["suppress_stdout"] = True
return log_config
[docs]
def save_draws(self, return_config=False, **kwargs):
generator = DrawGenerator(self, self.number_of_draws, self.runs_per_draw)
output_path = self.scenario_path.parent / f"{self.scenario_path.stem}_draws.json"
# assert not os.path.exists(output_path), f'Cannot save run config to {output_path} - file already exists'
config = generator.get_run_config(self.scenario_path)
if len(kwargs) > 0:
for k, v in kwargs.items():
config[k] = v
if "commit" in config:
github_url = f"https://github.com/UCL/TLOmodel/tree/{config['commit']}"
config["github"] = github_url
if return_config:
return config
else:
generator.save_config(config, output_path)
return output_path
[docs]
class ScenarioLoader:
"""A utility class to load a scenario class from a file path"""
def __init__(self, scenario_path):
scenario_module = ScenarioLoader._load_scenario_script(scenario_path)
scenario_class = ScenarioLoader._get_scenario_class(scenario_module)
self.scenario = scenario_class()
self.scenario.scenario_path = scenario_path
@staticmethod
def _load_scenario_script(path):
import importlib.util
spec = importlib.util.spec_from_file_location(Path(path).stem, path)
foo = importlib.util.module_from_spec(spec)
spec.loader.exec_module(foo)
return foo
@staticmethod
def _get_scenario_class(scenario_module):
import inspect
classes = inspect.getmembers(scenario_module, inspect.isclass)
classes = [c for (n, c) in classes if BaseScenario == c.__base__]
assert len(classes) == 1, "Exactly one subclass of BaseScenario should be defined in the scenario script"
return classes[0]
[docs]
def get_scenario(self):
return self.scenario
[docs]
class DrawGenerator:
"""Creates and saves a JSON representation of draws from a scenario."""
def __init__(self, scenario_class, number_of_draws, runs_per_draw):
self.scenario = scenario_class
assert self.scenario.seed is not None, "Must set a seed for the scenario. Add `self.seed = <integer>`"
self.scenario.rng = np.random.RandomState(seed=self.scenario.seed)
self.number_of_draws = number_of_draws
self.runs_per_draw = runs_per_draw
self.draws = self.setup_draws()
[docs]
def setup_draws(self):
assert self.scenario.number_of_draws > 0, "Number of draws must be greater than 0"
assert self.scenario.runs_per_draw > 0, "Number of samples/draw must be greater than 0"
if self.scenario.draw_parameters(1, self.scenario.rng) is None:
assert self.scenario.number_of_draws == 1, "Number of draws should equal one if no variable parameters"
return [self.get_draw(d) for d in range(0, self.scenario.number_of_draws)]
[docs]
def get_draw(self, draw_number):
return {
"draw_number": draw_number,
"parameters": self.scenario.draw_parameters(draw_number, self.scenario.rng),
}
[docs]
def get_run_config(self, scenario_path):
return {
"scenario_script_path": str(PurePosixPath(scenario_path)),
"scenario_seed": self.scenario.seed,
"runs_per_draw": self.runs_per_draw,
"draws": self.draws,
}
[docs]
def save_config(self, config, output_path):
with open(output_path, "w") as f:
f.write(json.dumps(config, indent=2))
[docs]
class SampleRunner:
"""Reads scenario draws from a JSON configuration and handles running of samples"""
def __init__(self, run_configuration_path):
with open(run_configuration_path, "r") as f:
self.run_config = json.load(f)
self.scenario = ScenarioLoader(self.run_config["scenario_script_path"]).get_scenario()
logger.info(key="message", data=f"Loaded scenario using {run_configuration_path}")
logger.info(key="message", data=f"Found {self.number_of_draws} draws; {self.runs_per_draw} runs/draw")
@property
def number_of_draws(self):
return len(self.run_config["draws"])
@property
def runs_per_draw(self):
return self.run_config["runs_per_draw"]
[docs]
def get_draw(self, draw_number):
total = self.number_of_draws
assert draw_number < total, f"Cannot get draw {draw_number}; only {total} defined."
return self.run_config["draws"][draw_number]
[docs]
def get_samples_for_draw(self, draw):
for sample_number in range(0, self.run_config["runs_per_draw"]):
yield self.get_sample(draw, sample_number)
[docs]
def get_sample(self, draw, sample_number):
assert sample_number < self.scenario.runs_per_draw, \
f"Cannot get sample {sample_number}; samples/draw={self.scenario.runs_per_draw}"
sample = draw.copy()
sample["sample_number"] = sample_number
# Instead of using the random number generator to create a seed for the simulation, we use an integer hash
# function to get an integer based on the sum of the scenario seed and sample_number. This means the
# seed can be created independently and out-of-order (i.e. instead of sampling a seed for each sample in order)
sample["simulation_seed"] = SampleRunner.low_bias_32(self.run_config["scenario_seed"] + sample_number)
return sample
[docs]
def run_sample_by_number(self, output_directory, draw_number, sample_number):
"""Runs a single sample from a draw, saving the output to the given directory"""
draw = self.get_draw(draw_number)
sample = self.get_sample(draw, sample_number)
log_config = self.scenario.get_log_config(output_directory)
logger.info(key="message", data=f"Running draw {sample['draw_number']}, sample {sample['sample_number']}")
sim = Simulation(
start_date=self.scenario.start_date,
seed=sample["simulation_seed"],
log_config=log_config
)
sim.register(*self.scenario.modules())
if sample["parameters"] is not None:
self.override_parameters(sim, sample["parameters"])
sim.make_initial_population(n=self.scenario.pop_size)
sim.simulate(end_date=self.scenario.end_date)
if sim.log_filepath is not None:
outputs = parse_log_file(sim.log_filepath)
for key, output in outputs.items():
if key.startswith("tlo."):
with open(Path(log_config["directory"]) / f"{key}.pickle", "wb") as f:
pickle.dump(output, f)
[docs]
def run(self):
"""Run all samples for the scenario. Used by `tlo scenario-run` to run the scenario locally"""
log_config = self.scenario.get_log_config()
root_dir = draw_dir = None
if log_config["directory"]: # i.e. write output files
timestamp = datetime.datetime.utcnow().strftime("%Y-%m-%dT%H%M%SZ")
root_dir = Path(log_config["directory"]) / (Path(log_config["filename"]).stem + "-" + timestamp)
# loop over draws and samples
for draw in range(0, self.scenario.number_of_draws):
for sample in range(0, self.runs_per_draw):
if root_dir is not None:
draw_dir = root_dir / f"{draw}/{sample}"
draw_dir.mkdir(parents=True, exist_ok=True)
self.run_sample_by_number(draw_dir, draw, sample)
[docs]
@staticmethod
def override_parameters(sim, overridden_params):
for module_name, overrides in overridden_params.items():
if module_name in sim.modules:
module = sim.modules[module_name]
for param_name, param_val in overrides.items():
assert param_name in module.PARAMETERS, f"{module} does not have parameter '{param_name}'"
# assert np.isscalar(param_val),
# f"Parameter value '{param_val}' is not scalar type (float, int, str)"
old_value = module.parameters[param_name]
assert type(old_value) is type(param_val), f"Cannot override parameter '{param_name}' - wrong type"
module.parameters[param_name] = param_val
logger.info(
key="override_parameter",
data={
"module": module_name,
"name": param_name,
"old_value": old_value,
"new_value": module.parameters[param_name]
}
)
[docs]
@staticmethod
def low_bias_32(x):
"""A simple integer hash function with uniform distribution. Following description taken from
https://github.com/skeeto/hash-prospector
The integer hash function transforms an integer hash key into an integer hash result. For a hash function,
the distribution should be uniform. This implies when the hash result is used to calculate hash bucket
address, all buckets are equally likely to be picked. In addition, similar hash keys should be hashed to
very different hash results. Ideally, a single bit change in the hash key should influence all bits of the
hash result.
:param: x an integer
:returns: an integer
"""
x *= 0x7feb352d
x ^= x >> 15
x *= 0x846ca68b
x ^= x >> 16
return x % (2 ** 32)
def _nested_dictionary_from_flat(flat_dict):
"""
Helper function for transforming a flat dictionary mapping from 2-tuple keys to
values to a corresponding nested dictionary with outer dictionary keyed by the first
key in the tuple and inner dictionaries keyed by the second key in the tuple.
"""
outer_dict = {}
for (key_1, key_2), value in flat_dict.items():
inner_dict = outer_dict.setdefault(key_1, {})
inner_dict[key_2] = value
return outer_dict
[docs]
def make_cartesian_parameter_grid(module_parameter_values_dict):
"""Make a list of dictionaries corresponding to a grid across parameter space.
The parameters values in each parameter dictionary corresponds to an element in the
Cartesian product of iterables describing the values taken by a collection of
module specific parameters.
Intended for use in ``BaseScenario.draw_parameters`` to determine the set of
parameters for a scenario draw.
Example usage (in ``BaseScenario.draw_parameters``)::
return make_cartesian_parameter_grid(
{
"Mockitis": {
"numeric_parameter": np.linspace(0, 1.0, 5),
"categorical_parameter": ["category_a", "category_b"]
},
"HealthSystem": {
"cons_availability": ["default", "all"]
}
}
)[draw_number]
In practice it will be more performant to call ``make_cartesian_parameter_grid``
once in the scenario ``__init__`` method, store this as an attribute of the scenario
and reuse this in each call to ``draw_parameters``.
:param dict module_parameter_values_dict: A dictionary mapping from module names
(as strings) to dictionaries mapping from (string) parameter names associated
with the model to iterables of the values over which parameter should take in
the grid.
:returns: A list of dictionaries mapping from module names (as strings) to
dictionaries mapping from (string) parameter names associated with the model to
parameter values, with each dictionary in the list corresponding to a single
point in the Cartesian grid across the parameter space.
"""
flattened_parameter_values_dict = {
(module, parameter): values
for module, parameter_values_dict in module_parameter_values_dict.items()
for parameter, values in parameter_values_dict.items()
}
return [
_nested_dictionary_from_flat(
{
key: value
for key, value in zip(flattened_parameter_values_dict, parameter_values)
}
)
for parameter_values in product(*flattened_parameter_values_dict.values())
]