def run_full_models_for_mcmc(burn_in: int, src_db_path: str, dest_db_path: str,
build_model, params: dict):
"""
Run the full baseline model and all scenarios for all accepted MCMC runs in src db.
"""
src_db = Database(src_db_path)
dest_db = Database(dest_db_path)
logger.info("Copying mcmc_run table to %s", dest_db_path)
mcmc_run_df = src_db.query("mcmc_run")
# Apply burn in and save to destination
burned_runs_str = ", ".join(mcmc_run_df[:burn_in].idx)
logger.info("Burned MCMC runs %s", burned_runs_str)
mcmc_run_df = mcmc_run_df[burn_in:]
dest_db.dump_df("mcmc_run", mcmc_run_df)
mcmc_runs = list(mcmc_run_df.T.to_dict().values())
for mcmc_run in mcmc_runs:
meta = {k: v for k, v in mcmc_run.items() if k in META_COLS}
if not meta["accept"]:
logger.info("Ignoring non-accepted MCMC run %s", meta["idx"])
continue
logger.info("Running full model for MCMC run %s", meta["idx"])
param_updates = {
k: v
for k, v in mcmc_run.items() if k not in META_COLS
}
run_idx = meta["idx"].split("_")[-1]
def update_func(ps: dict):
return update_params(ps, param_updates)
with Timer("Running model scenarios"):
num_scenarios = 1 + len(params["scenarios"].keys())
scenarios = []
for scenario_idx in range(num_scenarios):
scenario = Scenario(build_model, scenario_idx, params)
scenarios.append(scenario)
# Run the baseline scenario.
baseline_scenario = scenarios[0]
baseline_scenario.run(update_func=update_func)
baseline_model = baseline_scenario.model
# Run all the other scenarios
for scenario in scenarios[1:]:
scenario.run(base_model=baseline_model,
update_func=update_func)
with Timer("Saving model outputs to the database"):
models = [s.model for s in scenarios]
store_run_models(models, dest_db_path, run_idx=run_idx)
logger.info("Finished running full models for all accepted MCMC runs.")
def preprocess_demography(input_db: Database):
loc_df = read_location_df()
pop_df = read_population_df(loc_df)
birth_df = read_crude_birth_df(loc_df)
death_df = read_death_df(loc_df)
expect_df = read_life_expectancy_df(loc_df)
input_db.dump_df("countries", loc_df)
input_db.dump_df("population", pop_df)
input_db.dump_df("birth_rates", birth_df)
input_db.dump_df("deaths", death_df)
input_db.dump_df("life_expectancy", expect_df)
return loc_df
def preprocess_mobility(input_db: Database, country_df):
"""
Read Google Mobility data from CSV into input database
"""
mob_df = pd.read_csv(MOBILITY_CSV_PATH)
dhhs_cluster_mobility = reshape_to_clusters(mob_df)
# Drop all sub-region 2 data, too detailed.
major_region_mask = mob_df["sub_region_2"].isnull() & mob_df["metro_area"].isnull()
davao_mask = mob_df.metro_area == "Davao City Metropolitan Area"
mob_df = mob_df[major_region_mask | davao_mask].copy()
# These two regions are the same
mob_df.loc[(mob_df.sub_region_1 == "National Capital Region"), "sub_region_1"] = "Metro Manila"
mob_df.loc[(mob_df.metro_area == "Davao City Metropolitan Area"), "sub_region_1"] = "Davao City"
mob_df.loc[
(mob_df.sub_region_1 == "Federal Territory of Kuala Lumpur"), "sub_region_1"
] = "Kuala Lumpur"
mob_df = mob_df.append(dhhs_cluster_mobility)
# Drop all rows that have NA values in 1 or more mobility columns.
mob_cols = [c for c in mob_df.columns if c.endswith(MOBILITY_SUFFIX)]
mask = False
for c in mob_cols:
mask = mask | mob_df[c].isnull()
mob_df = mob_df[~mask].copy()
for c in mob_cols:
# Convert percent values to decimal: 1.0 being no change.
mob_df[c] = mob_df[c].apply(lambda x: 1 + x / 100)
# Drop unused columns, rename kept columns
cols_to_keep = [*mob_cols, "country_region", "sub_region_1", "date"]
cols_to_drop = [c for c in mob_df.columns if not c in cols_to_keep]
mob_df = mob_df.drop(columns=cols_to_drop)
mob_col_rename = {c: c.replace(MOBILITY_SUFFIX, "") for c in mob_cols}
mob_df.rename(columns={**mob_col_rename, "sub_region_1": "region"}, inplace=True)
# Convert countries to ISO3
countries = mob_df["country_region"].unique().tolist()
iso3s = {c: get_iso3(c, country_df) for c in countries}
iso3_series = mob_df["country_region"].apply(lambda c: iso3s[c])
mob_df.insert(0, "iso3", iso3_series)
mob_df = mob_df.drop(columns=["country_region"])
mob_df = mob_df.sort_values(["iso3", "region", "date"])
input_db.dump_df("mobility", mob_df)
def preprocess_social_mixing(input_db: Database, country_df):
for location in LOCATIONS:
for sheet_number, header_arg in SHEET_NUMBERS:
sheet_name = f"MUestimates_{location}_{sheet_number}.xlsx"
sheet_path = os.path.join(MIXING_DIRPATH, sheet_name)
xl = pd.ExcelFile(sheet_path)
sheet_names = xl.sheet_names
iso3s = [get_iso3(n, country_df) for n in sheet_names]
for idx, sheet_name in enumerate(sheet_names):
iso3 = iso3s[idx]
mix_df = pd.read_excel(xl,
header=header_arg,
sheet_name=sheet_name)
if sheet_number == "2":
renames = {n - 1: f"X{n}" for n in range(1, 17)}
mix_df.rename(columns=renames, inplace=True)
mix_df.insert(0, "location",
[location for _ in range(len(mix_df))])
mix_df.insert(0, "iso3", [iso3 for _ in range(len(mix_df))])
input_db.dump_df("social_mixing", mix_df)
def test_plot_uncertainty(tmp_path):
"""
Ensure uncertainty plotting code works.
"""
output_dir = tmp_path
powerbi_db_path = os.path.join(tmp_path, "powerbi.db")
targets = {
"incidence": {
"output_key": "incidence",
"title": "incidence",
"times": [],
"values": [],
"quantiles": [0.25, 0.5, 0.75],
},
"foo": {
"output_key": "foo",
"title": "foo",
"times": [],
"values": [],
"quantiles": [0.25, 0.5, 0.75],
},
}
funcs = [
lambda t: 2 * t + random.random(), lambda t: t**3 + random.random()
]
# Build data for plotting
do_df, mcmc_df, _ = build_synthetic_calibration(targets,
funcs,
chains=2,
runs=20,
times=20)
unc_df = calculate_mcmc_uncertainty(mcmc_df, do_df, targets)
# Create database for plotting
db = Database(powerbi_db_path)
db.dump_df("mcmc_run", mcmc_df)
db.dump_df("derived_outputs", do_df)
db.dump_df("uncertainty", unc_df)
# Create plots
plot_uncertainty(targets, powerbi_db_path, output_dir)
# Check plots
expected_foo_path = os.path.join(tmp_path, "foo", "uncertainty-foo-0.png")
expected_incidence_path = os.path.join(tmp_path, "incidence",
"uncertainty-incidence-0.png")
assert os.path.exists(expected_foo_path)
assert os.path.exists(expected_incidence_path)
def preprocess_our_world_in_data(input_db: Database):
df = pd.read_csv(OUR_WORLD_IN_DATA_CSV_PATH)
# Replace the one strange value for test numbers in Malaysia
df.loc[(df.iso_code == "MYS") & (df.new_tests > 1e5), "new_tests"] = np.nan
input_db.dump_df("owid", df)
def preprocess_covid_phl(input_db: Database):
df = pd.read_csv(COVID_PHL_CSV_PATH)
df = create_region_aggregates(df)
input_db.dump_df("covid_phl", df)
def preprocess_covid_au(input_db: Database):
df = pd.read_csv(COVID_AU_CSV_PATH)
input_db.dump_df("covid_au", df)
df = pd.read_csv(COVID_LGA_CSV_PATH)
df = reshape_to_clusters(df)
input_db.dump_df("covid_dhhs_test", df)
def preprocess_social_mixing(input_db: Database, country_df):
for location in LOCATIONS:
for sheet_number, header_arg in SHEET_NUMBERS:
sheet_name = f"MUestimates_{location}_{sheet_number}.xlsx"
sheet_path = os.path.join(MIXING_DIRPATH, sheet_name)
xl = pd.ExcelFile(sheet_path)
sheet_names = xl.sheet_names
iso3s = [get_iso3(n, country_df) for n in sheet_names]
for idx, sheet_name in enumerate(sheet_names):
iso3 = iso3s[idx]
mix_df = pd.read_excel(xl,
header=header_arg,
sheet_name=sheet_name)
if sheet_number == "2":
renames = {n - 1: f"X{n}" for n in range(1, 17)}
mix_df.rename(columns=renames, inplace=True)
mix_df.insert(0, "location",
[location for _ in range(len(mix_df))])
mix_df.insert(0, "iso3", [iso3 for _ in range(len(mix_df))])
input_db.dump_df("social_mixing", mix_df)
# Next gen social mixing
original_mm = input_db.query("social_mixing")
df = pd.read_csv(
os.path.join(MIXING_DIRPATH, "synthetic_contacts_2020.csv"))
df = df[df.setting == "overall"]
df.drop(columns="setting", inplace=True)
df.replace(
{
"0 to 4": "00 to 04",
"5 to 9": "05 to 09",
"all": "all_locations",
"others": "other_locations",
},
inplace=True,
)
# The contactor is in j (columns) and the contactee is in i (rows)
df = df.pivot_table(
index=["iso3c", "location_contact", "age_cotactee"],
columns="age_contactor",
values="mean_number_of_contacts",
)
df = df.reset_index()
df.drop(columns="age_cotactee", inplace=True)
cols = list(df.columns[2:])
new_col = ["X" + str(x) for x in range(1, len(cols) + 1)]
replace_col = dict(zip(cols, new_col))
df.rename(columns=replace_col, inplace=True)
df.rename(columns={
"iso3c": "iso3",
"location_contact": "location"
},
inplace=True)
iso3_diff = set(original_mm.iso3).difference(df.iso3)
iso3_mask = original_mm.iso3.isin(iso3_diff)
df = df.append(original_mm[iso3_mask], ignore_index=True)
input_db.dump_df("social_mixing_2020", df)
def test_plot_post_calibration(tmp_path):
plot_dir = tmp_path
mcmc_dir_path = os.path.join(tmp_path, "mcmc")
os.makedirs(mcmc_dir_path)
targets = {
"incidence": {
"output_key": "incidence",
"title": "incidence",
"times": [],
"values": [],
"quantiles": [0.25, 0.5, 0.75],
},
"foo": {
"output_key": "foo",
"title": "foo",
"times": [],
"values": [],
"quantiles": [0.25, 0.5, 0.75],
},
}
# A dummy prior to pass postirior checks
priors = [{
"param_name": "contact_rate",
"distribution": "uniform",
"distri_params": [0.01, 0.03]
}]
funcs = [
lambda t: 2 * t + random.random(), lambda t: t**3 + random.random()
]
# Build data for plotting
do_df, mcmc_df, params_df = build_synthetic_calibration(targets,
funcs,
chains=2,
runs=20,
times=20)
chains = set(mcmc_df["chain"].tolist())
# Create databases for plotting
for chain in chains:
db_path = os.path.join(mcmc_dir_path, f"chain-{chain}.db")
db = Database(db_path)
db.dump_df("mcmc_run", mcmc_df[mcmc_df["chain"] == chain])
db.dump_df("mcmc_params", params_df[params_df["chain"] == chain])
db.dump_df("derived_outputs", do_df[do_df["chain"] == chain])
# Create plots
plot_post_calibration(targets, mcmc_dir_path, plot_dir, priors)
# Check plots - do a super basic check
expected_files = [
"burn-in.png",
"loglikelihood-traces.png",
"acceptance_ratio.png",
"params-traces",
"calibration-fit",
"params-vs-loglikelihood",
"posteriors",
]
for fname in expected_files:
p = os.path.join(plot_dir, fname)
assert os.path.exists(p)
if os.path.isdir(p):
assert len(os.listdir(p)) > 0
