def _make_empbayes_fit(self, empbayes_grouping='order'):
if (self._empbayes_fit is None) or (empbayes_grouping !=
self.empbayes_grouping):
self.empbayes_grouping = empbayes_grouping
self._counter = {'iters': 0, 'evals': 0}
z0 = gv.BufferDict()
for group in self._empbayes_groupings():
z0[group] = 1.0
# Might need to change minargs default values for empbayes_fit to converge:
# tol=1e-8, svdcut=1e-12, debug=False, maxit=1000, add_svdnoise=False, add_priornoise=False
# Note: maxit != maxfev. See https://github.com/scipy/scipy/issues/3334
# For Nelder-Mead algorithm, maxfev < maxit < 3 maxfev?
# For debugging. Same as 'callback':
# https://github.com/scipy/scipy/blob/c0dc7fccc53d8a8569cde5d55673fca284bca191/scipy/optimize/optimize.py#L651
def analyzer(arg):
self._counter['evals'] += 1
print("\nEvals: ", self._counter['evals'], arg, "\n")
print(type(arg[0]))
return None
fit, z = lsqfit.empbayes_fit(z0,
fitargs=self._make_fitargs,
maxit=200,
analyzer=None)
print(z)
self._empbayes_fit = fit
return self._empbayes_fit
def main():
sys_stdout = sys.stdout
# version 1 - relative errors
sys.stdout = tee.tee(sys_stdout, open("eg7a.out", "w"))
# fit data and prior
x = np.array([1., 2., 3., 4.])
y = np.array([3.4422, 1.2929, 0.4798, 0.1725])
prior = gv.gvar(['10(1)', '1.0(1)'])
# fit function
def fcn(x, p):
return p[0] * gv.exp(-p[1] * x)
# find optimal dy
def fitargs(z):
dy = y * z
newy = gv.gvar(y, dy)
return dict(data=(x, newy), fcn=fcn, prior=prior)
fit, z = lsqfit.empbayes_fit(0.001, fitargs)
print fit.format(True)
if MAKE_PLOT:
ratio = fit.y / fcn(x, fit.pmean)
plt.errorbar(x=fit.x, y=gv.mean(ratio), yerr=gv.sdev(ratio), c='b')
plt.plot([0.5, 4.5], [1.0, 1.0], c='r')
# version 2 - additive errors
sys.stdout = tee.tee(sys_stdout, open("eg7b.out", "w"))
def fitargs(z):
dy = np.ones_like(y) * z
newy = gv.gvar(y, dy)
return dict(data=(x, newy), fcn=fcn, prior=prior)
fit, z = lsqfit.empbayes_fit(0.001, fitargs)
print fit.format(True)
if MAKE_PLOT:
ratio = fit.y / fcn(x, fit.pmean)
plt.errorbar(x=fit.x + 0.1,
y=gv.mean(ratio),
yerr=gv.sdev(ratio),
c='g')
plt.show()
def main():
gv.ranseed([2009,2010,2011,2012]) # initialize random numbers (opt.)
x,y = make_data() # make fit data
p0 = None # make larger fits go faster (opt.)
for nexp in range(3,8):
print('************************************* nexp =',nexp)
prior = make_prior(nexp)
fit = lsqfit.nonlinear_fit(data=(x,y),fcn=f,prior=prior,p0=p0,svdcut=SVDCUT)
print(fit) # print the fit results
E = fit.p['E'] # best-fit parameters
a = fit.p['a']
print('E1/E0 =',(E[1]/E[0]).fmt(),' E2/E0 =',(E[2]/E[0]).fmt())
print('a1/a0 =',(a[1]/a[0]).fmt(),' a2/a0 =',(a[2]/a[0]).fmt())
print()
if fit.chi2/fit.dof<1.:
p0 = fit.pmean # starting point for next fit (opt.)
if DO_ERRORBUDGET:
outputs = OrderedDict([
('E1/E0', E[1]/E[0]), ('E2/E0', E[2]/E[0]),
('a1/a0', a[1]/a[0]), ('a2/a0', a[2]/a[0])
])
inputs = OrderedDict([
('E', fit.prior['E']), ('a', fit.prior['a']),
('y', y), ('svd', fit.svdcorrection)
])
print(fit.fmt_values(outputs))
print(fit.fmt_errorbudget(outputs,inputs))
if DO_EMPBAYES:
def fitargs(z,nexp=nexp,prior=prior,f=f,data=(x,y),p0=p0):
z = gv.exp(z)
prior['a'] = [gv.gvar(0.5,0.5*z[0]) for i in range(nexp)]
return dict(prior=prior,data=data,fcn=f,p0=p0)
##
z0 = [0.0]
fit,z = lsqfit.empbayes_fit(z0,fitargs,tol=1e-3)
print(fit) # print the optimized fit results
E = fit.p['E'] # best-fit parameters
a = fit.p['a']
print('E1/E0 =',(E[1]/E[0]).fmt(),' E2/E0 =',(E[2]/E[0]).fmt())
print('a1/a0 =',(a[1]/a[0]).fmt(),' a2/a0 =',(a[2]/a[0]).fmt())
print("prior['a'] =",fit.prior['a'][0].fmt())
print()
if DO_PLOT:
import pylab as pp
from gvar import mean,sdev
fity = f(x,fit.pmean)
ratio = y/fity
pp.xlim(0,21)
pp.xlabel('x')
pp.ylabel('y/f(x,p)')
pp.errorbar(x=x,y=mean(ratio),yerr=sdev(ratio),fmt='ob')
pp.plot([0.0,21.0],[1.0,1.0])
pp.show()
def main():
gv.ranseed([2009,2010,2011,2012]) # initialize random numbers (opt.)
x,y = make_data() # make fit data
p0 = None # make larger fits go faster (opt.)
for nexp in range(3,5):
print '************************************* nexp =',nexp
prior = make_prior(nexp)
fit = lsqfit.nonlinear_fit(data=(x,y),fcn=f,prior=prior,p0=p0)
print fit # print the fit results
E = fit.p['E'] # best-fit parameters
a = fit.p['a']
print 'E1/E0 =',E[1]/E[0],' E2/E0 =',E[2]/E[0]
print 'a1/a0 =',a[1]/a[0],' a2/a0 =',a[2]/a[0]
print
if fit.chi2/fit.dof<1.:
p0 = fit.pmean # starting point for next fit (opt.)
sys_stdout = sys.stdout
if DO_ERRORBUDGET:
lines = [
"E = fit.p['E']",
"a = fit.p['a']",
"print(E[1] / E[0])",
"print((E[1] / E[0]).partialsdev(fit.prior['E']))",
"print((E[1] / E[0]).partialsdev(fit.prior['a']))",
"print((E[1] / E[0]).partialsdev(y))"
]
sys.stdout = tee.tee(sys_stdout, open("eg4c.out","w"))
for line in lines:
print ">>>", line
if line[:5] == "print":
print(eval(line[5:]))
# print E[1]/E[0]
# print (E[1]/E[0]).partialsdev(fit.prior['E'])
# print (E[1]/E[0]).partialsdev(fit.prior['a'])
# print (E[1]/E[0]).partialsdev(y)
outputs = {'E1/E0':E[1]/E[0], 'E2/E0':E[2]/E[0],
'a1/a0':a[1]/a[0], 'a2/a0':a[2]/a[0]}
inputs = {'E':fit.prior['E'],'a':fit.prior['a'],'y':y}
sys.stdout = tee.tee(sys_stdout, open("eg4b.out","w"))
print fit.fmt_values(outputs)
print fit.fmt_errorbudget(outputs,inputs)
sys.stdout = sys_stdout
if DO_SIMULATIONS:
# fit simulations
sys.stdout = tee.tee(sys_stdout, open("eg4d.out","w"))
for sfit in fit.simulated_fit_iter(3):
print '************************************* simulation'
print(sfit)
sE = sfit.p['E'] # best-fit parameters
sa = sfit.p['a']
E = sfit.pexact['E']
a = sfit.pexact['a']
print 'E1/E0 =', sE[1] / sE[0], ' E2/E0 =', sE[2] / sE[0]
print 'a1/a0 =', sa[1] / sa[0], ' a2/a0 =', sa[2] / sa[0]
print '\nSimulated Fit Values - Exact Values:'
print 'E1/E0:', (sE[1] / sE[0]) - (E[1] / E[0]),\
' E2/E0:', (sE[2] / sE[0]) - (E[2] / E[0])
print 'a1/a0:', (sa[1] / sa[0]) - (a[1] / a[0]),\
' a2/a0:', (sa[2] / sa[0]) - (a[2] / a[0])
# compute chi**2 comparing fit results to exact results
sim_results = [sE[0], sE[1], sa[0], sa[1]]
exact_results = [E[0], E[1], a[0], a[1]]
chi2 = gv.chi2(sim_results, exact_results, svdcut=1e-8)
print '\nParameter chi2/dof [dof] = %.2f' % (chi2/chi2.dof), '[%d]' % chi2.dof, ' Q = %.1f' % chi2.Q
print
sys.stdout = sys_stdout
if DO_EMPBAYES:
def fitargs(z,nexp=nexp,prior=prior,f=f,data=(x,y),p0=p0):
z = gv.exp(z)
prior['a'] = [gv.gvar(0.5,0.5*z[0]) for i in range(nexp)]
return dict(prior=prior,data=data,fcn=f,p0=p0)
##
z0 = [0.0]
fit,z = lsqfit.empbayes_fit(z0,fitargs,tol=1e-3)
sys.stdout = tee.tee(sys_stdout, open("eg4a.out","w"))
print fit # print the optimized fit results
E = fit.p['E'] # best-fit parameters
a = fit.p['a']
print 'E1/E0 =',E[1]/E[0],' E2/E0 =',E[2]/E[0]
print 'a1/a0 =',a[1]/a[0],' a2/a0 =',a[2]/a[0]
# print "prior['a'] =",fit.prior['a'][0]
sys.stdout = sys_stdout
print
if DO_PLOT:
import pylab as pp
from gvar import mean,sdev
fity = f(x,fit.pmean)
ratio = y/fity
pp.xlim(0,21)
pp.xlabel('x')
pp.ylabel('y/f(x,p)')
pp.errorbar(x=x,y=mean(ratio),yerr=sdev(ratio),fmt='ob')
pp.plot([0.0,21.0],[1.0,1.0])
pp.show()
data=data,
fcn=fcn,
prior=prior,
p0=p0,
)
residuals = linalg.solve_triangular(hyperchol, hp.buf - hypermean)
plausibility = -1 / 2 * (residuals @ residuals)
return args, plausibility
print('\ntrue hyperparameters:')
print(truehp)
print('\nfit:')
z0 = gvar.sample(hyperprior)
fit, fithp = lsqfit.empbayes_fit(z0, fitargs)
print(fit.format(maxline=True, pstyle='v'))
print(fit.format(maxline=-1))
gp = makegp(fithp)
prior = makeprior(gp)
pred = fcn(fit.p)
fitgrid = gp.predfromfit(dict(datagrid=fit.p['datagrid'], **constraints),
'plotgrid')
print('\ncheck constraints in fit:')
check_constraints(fitgrid)
def allkeys(d):
for k in d:
def main():
gd.ranseed([2009,2010,2011,2012]) # initialize random numbers (opt.)
x,y = make_data() # make fit data
p0 = None # make larger fits go faster (opt.)
for nexp in range(2,8):
if nexp == 2:
sys_stdout = sys.stdout
sys.stdout = tee.tee(sys_stdout, open("eg4GBF.out","w"))
print '************************************* nexp =',nexp
prior = make_prior(nexp)
fit = lsqfit.nonlinear_fit(data=(x,y),fcn=f,prior=prior,p0=p0)
print fit # print the fit results
# E = fit.p['E'] # best-fit parameters
# a = fit.p['a']
# print 'E1/E0 =',E[1]/E[0],' E2/E0 =',E[2]/E[0]
# print 'a1/a0 =',a[1]/a[0],' a2/a0 =',a[2]/a[0]
print
if nexp == 3:
sys.stdout = sys_stdout
if fit.chi2/fit.dof<1.:
p0 = fit.pmean # starting point for next fit (opt.)
if DO_ERRORBUDGET:
print E[1]/E[0]
print (E[1]/E[0]).partialsdev(fit.prior['E'])
print (E[1]/E[0]).partialsdev(fit.prior['a'])
print (E[1]/E[0]).partialsdev(y)
outputs = {'E1/E0':E[1]/E[0], 'E2/E0':E[2]/E[0],
'a1/a0':a[1]/a[0], 'a2/a0':a[2]/a[0]}
inputs = {'E':fit.prior['E'],'a':fit.prior['a'],'y':y}
sys.stdout = tee.tee(sys_stdout, open("eg4GBFb.out","w"))
print fit.fmt_values(outputs)
print fit.fmt_errorbudget(outputs,inputs)
sys.stdout = sys_stdout
if DO_EMPBAYES:
def fitargs(z,nexp=nexp,prior=prior,f=f,data=(x,y),p0=p0):
z = gd.exp(z)
prior['a'] = [gd.gvar(0.5,0.5*z[0]) for i in range(nexp)]
return dict(prior=prior,data=data,fcn=f,p0=p0)
##
z0 = [0.0]
fit,z = lsqfit.empbayes_fit(z0,fitargs,tol=1e-3)
sys.stdout = tee.tee(sys_stdout, open("eg4GBFa.out","w"))
print fit # print the optimized fit results
E = fit.p['E'] # best-fit parameters
a = fit.p['a']
print 'E1/E0 =',E[1]/E[0],' E2/E0 =',E[2]/E[0]
print 'a1/a0 =',a[1]/a[0],' a2/a0 =',a[2]/a[0]
print "prior['a'] =",fit.prior['a'][0]
sys.stdout = sys_stdout
print
if DO_PLOT:
import pylab as pp
from gvar import mean,sdev
fity = f(x,fit.pmean)
ratio = y/fity
pp.xlim(0,21)
pp.xlabel('x')
pp.ylabel('y/f(x,p)')
pp.errorbar(x=x,y=mean(ratio),yerr=sdev(ratio),fmt='ob')
pp.plot([0.0,21.0],[1.0,1.0])
pp.show()
def main():
"""Executes the command line script
"""
if __name__ == "__main__":
parameter_file_path = 'data/foo.csv'
parameters = read_parameters('data/foo.csv')
data_file_path = 'data/HUP_ts.csv'
data = read_data(data_file_path)
error_file_path = 'data/data_errors.csv'
model = SEIRModel(fit_columns=["hospital_census", "vent_census"],
update_parameters=flexible_beta)
xval = True
k = 10
spline_power = 2
splines = np.linspace(0, data.shape[0] - 5, k).astype(int)
win = 40
pen = .002
beta_fun = 'flexible_beta'
pd.options.display.max_rows = 4000
pd.options.display.max_columns = 4000
else:
args = parse_args()
#
data_file_path = args.data_file
parameter_file_path = args.parameter_file
beta_fun = args.beta
spline_power = args.spline_power
xval = args.cross_validate if args.beta == "flexible_beta" else False
error_file_path = args.data_error_file
k = args.spline_dimension
if args.verbose:
for handler in LOGGER.handlers:
handler.setLevel(DEBUG)
LOGGER.debug("Received arguments:\n%s", args)
parameters = read_parameters(parameter_file_path)
LOGGER.debug("Read parameters:\n%s", parameters)
data = read_data(data_file_path)
LOGGER.debug("Read data:\n%s", data)
model = SEIRModel(
fit_columns=["hospital_census", "vent_census"],
update_parameters=flexible_beta if beta_fun == "flexible_beta" \
else logistic_social_policy,
)
# parse the splines
# TODO: note this will need to be generalized once we've got more features time-varying
if k > 0:
splines = np.arange(0, data.shape[0], int(data.shape[0] / k))
else:
splines = -99
assert args.beta != "flexible_beta", "You need to specify some splines with '-k <spline dimension> if you're using flexible beta"
## CROSS VALIDATION
if xval is True:
print("Doing rolling-window cross-validation")
assert error_file_path is not None, "Haven't yet implemented cross-validation for empirical bayes. Please supply a data error file (i.e.: `-y data/data_errors.csv`)"
# loop through windows, and in each one, forecast one week out.
penvec = 10**np.linspace(-10, 5, 16)
winstart = list(range(data.shape[0] - 14, (data.shape[0] - 7)))
tuples_for_starmap = [(p, w, parameter_file_path, splines, k,
data_file_path, error_file_path, k)
for p in penvec for w in winstart]
pool = mp.Pool(mp.cpu_count())
xval_results = pool.starmap(xval_wrapper, tuples_for_starmap)
pool.close()
xval_df = pd.DataFrame(xval_results)
# remove errors
errors = (xval_df.mse == -9999).sum()
# assert errors < xval_df.shape[0]*.2, "Lot's of errors when doing cross-validation. Breaking here rather than returning unreliable results."
xval_df = xval_df.loc[xval_df.mse > 0]
xval_df['rmse'] = xval_df.mse**.5
penframe = xval_df.groupby(['pen']).agg({
'rmse': ['mean', 'std']
},
as_index=False).reset_index()
penframe.columns = ['pen', 'mu', 'sig']
best_penalty = penframe.pen.loc[penframe.mu == min(
penframe.mu)].iloc[0]
print(
f"The best prior sd on the splines is {best_penalty}. Don't forget to look at the plot of cross-validation statistics (in the output directory) to make sure that there's nothing wacky going on."
)
parameters['pen_beta'] = gvar(0, best_penalty)
degen_flag = True
while degen_flag:
xx, pp = prepare_model_parameters(parameters=parameters,
data=data,
beta_fun=beta_fun,
splines=splines,
spline_power=spline_power)
LOGGER.debug("Parsed model meta pars:\n%s", xx)
LOGGER.debug("Parsed model priors:\n%s", pp)
model.fit_start_date = xx["day0"]
# If empirical bayes is selected to fit the data, this also returns the fit object
LOGGER.debug("Starting fit")
if args.data_error_file:
xx["error_infos"] = (read_csv(error_file_path).set_index("param")
["value"].to_dict())
LOGGER.debug("Using y_errs from file:\n%s", xx["error_infos"])
fit = nonlinear_fit(
data=(xx, get_yy(data, **xx["error_infos"])),
prior=pp,
fcn=model.fit_fcn,
# debug=args.verbose,
)
else:
LOGGER.debug("Employing empirical Bayes to infer y-errors")
# This fit varies the size of the y-errors of hosp_min and vent_min
# to optimize the description of the data (logGBF)
fit_kwargs = lambda error_infos: dict(
data=(xx, get_yy(data, hosp_rel=0, vent_rel=0, **error_infos)),
prior=pp,
fcn=model.fit_fcn,
debug=args.verbose,
)
fit, xx["error_infos"] = empbayes_fit(
{
"hosp_min": 10,
"vent_min": 1
}, fit_kwargs)
LOGGER.debug("Empbayes y_errs are:\n%s", xx["error_infos"])
# check for degeneracy
splinecoefvec = np.array([
fit.p['beta_splines'][i].mean
for i in range(len(fit.p['beta_splines']))
])
cv = np.std(splinecoefvec[1:]) / np.mean(splinecoefvec[1:])
BI_update = (fit.p['beta_intercept'] -
parameters['beta_intercept']).mean
coef_OM_range = np.ptp(np.log10(np.abs(splinecoefvec)))
if (cv < .1) | (BI_update**2 < .1) | (coef_OM_range > 2):
print(
'the best prior sd on the splines led to a degenerate fit. trimming it by one order of magnitude'
)
curr = np.log10(best_penalty)
assert curr > -5, "degenerate solutions all the way down. Something is broken."
best_penalty = 10**(curr - 1)
print(f"new best prior sd on the splines is {best_penalty}")
parameters['pen_beta'] = gvar(0, best_penalty)
else:
degen_flag = False
LOGGER.info("Fit result:\n%s", fit)
dump_results(args.output_dir,
fit=fit,
model=model,
extend_days=args.extend_days)
LOGGER.debug("Dumped results to:\n%s", args.output_dir)
# the Free Software Foundation, either version 3 of the License, or
# any later version (see <http://www.gnu.org/licenses/>).
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
from __future__ import print_function # makes this work for python2 and 3
import numpy as np
import gvar as gv
import lsqfit
x = np.array([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7])
y = np.array([
'0.133426(95)', '0.20525(15)', '0.27491(20)', '0.32521(25)',
'0.34223(28)', '0.32394(28)', '0.27857(27)'
])
def fcn(x, p):
return gv.exp(-p[0] - p[1] * x - p[2] * x**2 - p[3] * x**3)
def fitargs(z):
dp = z
prior = gv.gvar([gv.gvar(0, dp) for i in range(4)])
return dict(prior=prior, fcn=fcn, data=(x,y))
fit,z = lsqfit.empbayes_fit(1.0, fitargs)
print(fit.format(True))
y = np.array([
'0.133426(95)', '0.20525(15)', '0.27491(20)', '0.32521(25)', '0.34223(28)',
'0.32394(28)', '0.27857(27)'
])
def fcn(x, p):
return gv.exp(-p[0] - p[1] * x - p[2] * x**2 - p[3] * x**3)
def fitterargs(z):
prior = [gv.gvar(0, z) for i in range(4)]
return dict(prior=prior, fcn=fcn, data=(x, y))
fit, z = lsqfit.empbayes_fit(1., fitterargs)
sys.stdout = tee.tee(sys_stdout, open('eg4a.out', 'w'))
print fit.format(True)
sys.stdout = sys_stdout
prior = gv.gvar([
'2.5904 +- 2.6e-16',
'-6.53012 +- 6.5e-16',
'7.83211 +- 7.8e-16',
'-1.68813 +- 1.7e-16',
])
fit = lsqfit.nonlinear_fit(data=(x, y), prior=prior, fcn=fcn)
print fit
def bayes_xval(days_withheld=7, which_hospital="HUP"):
try:
parameters = read_parameters(
f"{datadir}{which_hospital}_parameters.csv")
data = read_data(f"{datadir}{which_hospital}_ts.csv")[:-days_withheld]
test_set = pd.read_csv(
f"{datadir}{which_hospital}_ts.csv")[-days_withheld:]
test_set.date = test_set.date.astype("datetime64[ns]")
model = SEIRModel(
fit_columns=["hospital_census", "vent_census"],
update_parameters=logisitic_social_policy,
)
xx, pp = prepare_model_parameters(parameters, data)
model.fit_start_date = xx["day0"]
fit_kwargs = lambda error_infos: dict(
data=(xx, get_yy(data, hosp_rel=0, vent_rel=0, **error_infos)),
prior=pp,
fcn=model.fit_fcn,
debug=True,
)
fit, xx["error_infos"] = empbayes_fit({
"hosp_min": 10,
"vent_min": 1
}, fit_kwargs)
# extend by 60 days
xx["dates"] = xx["dates"].union(
date_range(xx["dates"].max(), freq="D", periods=60))
prediction_df = model.propagate_uncertainties(xx, fit.p)
prediction_df.index = prediction_df.index.round("H")
# drop the index
prediction_df = prediction_df.reset_index()
prediction_df['hmu'] = prediction_df.hospital_census.apply(
lambda x: float(str(x).split("(")[0]))
prediction_df['hsig'] = prediction_df.hospital_census.apply(
lambda x: float(str(x).split("(")[1][:-1])
if "(" in str(x) else float(x))
prediction_df['vmu'] = prediction_df.vent_census.apply(
lambda x: float(str(x).split("(")[0]))
prediction_df['vsig'] = prediction_df.vent_census.apply(
lambda x: float(str(x).split("(")[1][:-1])
if "(" in str(x) else float(x))
# merge
mm = prediction_df.merge(test_set, how='left')
tomerge = copy.deepcopy(data)
tomerge.columns = ["obs_" + i for i in tomerge.columns]
tomerge.reset_index(inplace=True)
mm = mm.merge(tomerge, how='outer')
# compute simple msfe
hMSFE = np.mean((mm.hmu - mm.hosp)**2)
vMSFE = np.mean((mm.vmu - mm.vent)**2)
loss_approx = (hMSFE + vMSFE) / 2
# now run MCMC
params_raw = pd.read_csv(f"{datadir}{which_hospital}_parameters.csv")
df = do_chains(n_iters=9000,
params=params_raw,
obs=read_data(f"{datadir}{which_hospital}_ts.csv"),
best_penalty=None,
sample_obs=False,
holdout=days_withheld,
n_chains=1,
parallel=False)
df = df.loc[df.iter > 1000]
arrs = np.stack([df.arr.iloc[i] for i in range(df.shape[0])])
arrs_test = arrs[:, data.shape[0]:(data.shape[0] + days_withheld), :]
median_pred = np.median(arrs_test, axis=0)
loss_mcmc = (np.mean((median_pred[:,3] - test_set.hosp)**2) \
+ np.mean((median_pred[:,5] - test_set.vent)**2))/2 # output
# prediction quantiles: the proportion of times the prediction is greater than the mean
hq_m = [(arrs_test[:, day, 3] > test_set.hosp.iloc[day]).mean()
for day in range(days_withheld)]
vq_m = [(arrs_test[:, day, 5] > test_set.vent.iloc[day]).mean()
for day in range(days_withheld)]
mmtail = mm.tail(days_withheld)
hq_n = [(np.random.normal(mmtail.hmu.iloc[day], mmtail.hsig.iloc[day],
10000) > test_set.hosp.iloc[day]).mean()
for day in range(days_withheld)]
vq_n = [(np.random.normal(mmtail.vmu.iloc[day], mmtail.vsig.iloc[day],
10000) > test_set.vent.iloc[day]).mean()
for day in range(days_withheld)]
plotr = dict(hq_m=hq_m, vq_m=vq_m, hq_n=hq_n, vq_n=vq_n)
resh_n = mm.hmu[:len(data) +
len(test_set)] - np.array(data.hosp.tolist() +
test_set.hosp.tolist())
resh_m = np.median(
arrs[:, :len(data) + len(test_set), 3],
axis=0) - np.array(data.hosp.tolist() + test_set.hosp.tolist())
resv_n = mm.vmu[:len(data) +
len(test_set)] - np.array(data.vent.tolist() +
test_set.vent.tolist())
resv_m = np.median(
arrs[:, :len(data) + len(test_set), 5],
axis=0) - np.array(data.vent.tolist() + test_set.vent.tolist())
plotq = dict(
resh_n=mm.hmu[:len(data) + len(test_set)] -
np.array(data.hosp.tolist() + test_set.hosp.tolist()),
resh_m=np.median(arrs[:, :len(data) + len(test_set), 3], axis=0) -
np.array(data.hosp.tolist() + test_set.hosp.tolist()),
resv_n=mm.vmu[:len(data) + len(test_set)] -
np.array(data.vent.tolist() + test_set.vent.tolist()),
resv_m=np.median(arrs[:, :len(data) + len(test_set), 5], axis=0) -
np.array(data.vent.tolist() + test_set.vent.tolist()))
out = dict(which_hospital=which_hospital,
days_out=days_withheld,
loss_mcmc=loss_mcmc,
loss_approx=loss_approx,
plotq=plotq,
plotr=plotr)
return out
except Exception as e:
print(e)
