def run_joint_model(self, df, groups):
model = CurveModel(
df=df[df[self.col_group].isin(groups)].copy(),
**self.basic_model_dict
)
model.fit_params(**self.joint_model_fit_dict)
return model
def run_model(self, df, group):
"""Run each individual model.
"""
model = CurveModel(df=df[df[self.col_group] == group].copy(),
**self.basic_model_dict)
fit_dict = deepcopy(self.fit_dict)
fe_gprior = fit_dict['fe_gprior']
fe_gprior[1][1] *= self.prior_modifier(model.num_obs)
print(group)
print('\t update beta fe_gprior to', fe_gprior)
fit_dict.update({'fe_gprior': fe_gprior})
model.fit_params(**fit_dict)
return model
class BasicModel(ModelPipeline):
def __init__(self, fit_dict, basic_model_dict, **pipeline_kwargs):
"""
Generic class for a function to produce predictions from a model
with the following attributes.
Args:
**pipeline_kwargs: keyword arguments for the base class of ModelPipeline
predict_group: (str) which group to make predictions for
fit_dict: keyword arguments to CurveModel.fit_params()
basic_model_dict: additional keyword arguments to the CurveModel class
col_obs_se: (str) of observation standard error
col_covs: List[str] list of names of covariates to put on the parameters
param_names (list{str}):
Names of the parameters in the specific functional form.
link_fun (list{function}):
List of link functions for each parameter.
var_link_fun (list{function}):
List of link functions for the variables including fixed effects
and random effects.
"""
super().__init__(**pipeline_kwargs)
self.fit_dict = fit_dict
self.basic_model_dict = basic_model_dict
self.basic_model_dict.update({'col_obs_se': self.col_obs_se})
generator_kwargs = pipeline_kwargs
for arg in self.pop_cols:
generator_kwargs.pop(arg)
self.basic_model_dict.update(**generator_kwargs)
self.mod = None
self.setup_pipeline()
def refresh(self):
self.mod = None
def fit(self, df, group=None):
self.mod = CurveModel(df=df, **self.basic_model_dict)
self.mod.fit_params(**self.fit_dict)
def predict(self, times, predict_space, predict_group):
predictions = self.mod.predict(
t=times,
group_name=predict_group,
prediction_functional_form=predict_space)
return predictions
def run_model(self, df, group):
__doc__ = super().run_model.__doc__
model = CurveModel(df=df[df[self.col_group] == group].copy(),
**self.basic_model_dict)
fit_dict = deepcopy(self.fit_dict)
print(group)
fe_gprior = fit_dict['fe_gprior']
common_beta_bounds = fit_dict['fe_bounds'][1]
suggested_beta_bounds = self.init_parameters_estimations[
"fe_bounds_beta"].get(group, None)
if suggested_beta_bounds is not None:
print("\t Suggested beta bounds ", suggested_beta_bounds)
individual_beta_bounds = [
max(common_beta_bounds[0], suggested_beta_bounds[0] * 1.2),
common_beta_bounds[1]
]
fit_dict['fe_bounds'][1] = individual_beta_bounds
fe_gprior[1][0] = max(individual_beta_bounds[0], fe_gprior[1][0])
print('\t Update beta bounds to ', individual_beta_bounds)
else:
print('\t Use common beta bounds ', common_beta_bounds)
individual_beta_bounds = common_beta_bounds
fe_gprior[1][1] *= self.prior_modifier(model.num_obs)
print('\t Update beta fe_gprior to ', fe_gprior)
fit_dict.update({
'fe_gprior':
fe_gprior,
'fe_bounds': [
fit_dict['fe_bounds'][0], individual_beta_bounds,
fit_dict['fe_bounds'][2]
]
})
# print(fit_dict['fe_gprior'])
model.fit_params(**fit_dict)
if suggested_beta_bounds is not None:
fit_dict.update({
'fe_bounds': [
fit_dict['fe_bounds'][0], common_beta_bounds,
fit_dict['fe_bounds'][2]
]
})
return model
def run_model(self, group, **fit_kwargs):
"""Construct and run the model.
"""
model = CurveModel(self.df[self.df[self.col_group] == group].copy(),
col_t=self.col_t,
col_obs=self.col_obs,
col_covs=self.col_covs,
col_group=self.col_group,
param_names=self.param_names,
link_fun=self.link_fun,
var_link_fun=self.var_link_fun,
fun=self.fun,
col_obs_se=self.col_obs_se)
model.fit_params(**fit_kwargs)
return model
def test_curve_model(alpha_true, beta_true, p_true, n_data):
num_params = 3
params_true = np.array([alpha_true, beta_true, p_true])
independent_var = np.array(range(n_data)) * beta_true / (n_data - 1)
df = pd.DataFrame({
'independent_var':
independent_var,
'measurement_value':
generalized_logistic(independent_var, params_true),
'measurement_std':
n_data * [0.1],
'constant_one':
n_data * [1.0],
'data_group':
n_data * ['world'],
})
# Initialize a model
cm = CurveModel(df=df,
col_t='independent_var',
col_obs='measurement_value',
col_covs=num_params * [['constant_one']],
col_group='data_group',
param_names=['alpha', 'beta', 'p'],
link_fun=[exp_fun, identity_fun, exp_fun],
var_link_fun=[exp_fun, identity_fun, exp_fun],
fun=generalized_logistic,
col_obs_se='measurement_std')
inv_link_fun = [ln_fun, identity_fun, ln_fun]
fe_init = np.zeros(num_params)
for j in range(num_params):
fe_init[j] = inv_link_fun[j](params_true[j] / 3.0)
# Fit the parameters
cm.fit_params(fe_init=fe_init,
options={
'ftol': 1e-16,
'gtol': 1e-16,
'maxiter': 1000
})
params_estimate = cm.params
for i in range(num_params):
rel_error = params_estimate[i] / params_true[i] - 1.0
assert abs(rel_error) < 1e-6
def fit(self, df, group=None):
"""
Fits a loose, tight, beta, and p combinations model. If you pass in
update group it will override the initial parameters with new
initial parameters based on the df you pass.
Args:
df:
group: (str) passing in the group will update the initialization
dictionary (not replacing the old one) for this particular fit.
Returns:
"""
if group is not None:
init_dict = self.update_init_model(df=df, group=group)
else:
init_dict = deepcopy(self.init_dict)
for param in ['beta', 'p']:
if getattr(self, f'{param}_weight') == 0:
continue
for fit_type in ['loose', 'tight']:
model_arg_dict = deepcopy(
getattr(self, f'{param}_model_kwargs'))
fit_arg_dict = deepcopy(
getattr(self, f'{fit_type}_{param}_fit_dict'))
model = CurveModel(df=df, **model_arg_dict)
fe_init, re_init = compute_starting_params(
init_dict[param][fit_type])
fit_arg_dict.update(fe_init=fe_init, re_init=re_init)
model.fit_params(**fit_arg_dict)
setattr(self, f'{fit_type}_{param}_model', model)
# Set up the CurveModel
model = CurveModel(
df=df,
col_t='time',
col_obs='ln_death_rate',
col_group='group',
col_covs=[['intercept'], ['intercept'], ['intercept']],
param_names=['alpha', 'beta', 'p'],
link_fun=[lambda x: x, lambda x: x, lambda x: x],
var_link_fun=[lambda x: x, lambda x: x, lambda x: x],
fun=ln_gaussian_cdf
)
# Fit the model to estimate parameters
model.fit_params(fe_init=[0, 0, 1.],
fe_gprior=[[0, np.inf], [0, np.inf], [1., np.inf]],
options={'disp':True})
# Get predictions
y_pred = model.predict(
t=np.linspace(0,100,num=100),
group_name="Kerala"
)
ground_truth = df.ln_death_rate[df['group']=="Hubei"].reset_index(drop=True)
print(ground_truth)
print(np.exp(y_pred))
# Plot results
plt.plot(np.linspace(0,100,num=100), y_pred, '-')
plt.plot(ground_truth, '.')
plt.show()
cm = CurveModel(
df=df,
col_t='independent_var',
col_obs='measurement_value',
col_covs=num_params*[['constant_one']],
col_group='data_group',
param_names=['alpha', 'beta', 'p'],
link_fun=[exp_fun, identity_fun, exp_fun],
var_link_fun=[exp_fun, identity_fun, exp_fun],
fun=generalized_logistic,
col_obs_se='measurement_std'
)
inv_link_fun = [ln_fun, identity_fun, ln_fun]
fe_init = np.zeros(num_params)
for j in range(num_params):
fe_init[j] = inv_link_fun[j](params_true[j] / 3.0)
cm.fit_params(
fe_init=fe_init,
options={
'ftol': 1e-16,
'gtol': 1e-16,
'maxiter': 1000
}
)
params_estimate = cm.params
print("True parameters:\n", params_true)
print("Estimated parameters:\n", params_estimate)
class GaussianCDF:
"""Fit a single Gaussian Atom to cumulative daily deaths"""
def __init__(self):
self.model = None
def fit(self, daily_deaths, social_distance=None):
daily_deaths = np.array(daily_deaths)
n_data = daily_deaths.shape[0]
# Prepare the data frame
df = pd.DataFrame()
df['death_rate'] = np.cumsum(daily_deaths)
df['time'] = np.arange(df['death_rate'].shape[0])
df['ln_death_rate'] = np.log(
df['death_rate'] + 1) # Add 1 to pad in case the #deaths are zero
df['group'] = ['all'] * n_data
df['cov_one'] = [1.0] * n_data
if social_distance is not None:
df['social_distance'] = social_distance
col_covs = [['cov_one'], ['cov_one', 'social_distance'],
['cov_one']]
num_fe = 4
fe_init = [-3, 100, 0, 10]
# col_covs = [['cov_one', 'social_distance'], ['cov_one', 'social_distance'], ['cov_one', 'social_distance']]
# num_fe = 6
# fe_init = [-3, 0, 100, 0, 10, 0]
else:
col_covs = [['cov_one'], ['cov_one'], ['cov_one']]
num_fe = 3
fe_init = [-3, 100, 1]
# Set up the CurveModel
self.model = CurveModel(
df=df,
col_t='time',
col_obs='ln_death_rate',
col_group='group',
col_covs=col_covs,
param_names=['alpha', 'beta', 'p'],
link_fun=[lambda x: np.exp(x), lambda x: x, lambda x: np.exp(x)],
var_link_fun=[lambda x: x] * num_fe,
fun=ln_gaussian_cdf)
# Fit the model to estimate parameters
self.model.fit_params(
fe_init=fe_init,
options={
'ftol': 1e-14,
'maxiter': 500
},
re_bounds=[[0, 0]] * num_fe # No random effects
)
def predict(self, t):
"""Get predictions for values in t"""
return self.model.predict(t=t, group_name='all')
def get_params(self):
return np.squeeze(self.model.params)
class BasicModelWithInit(BasicModel):
def __init__(self, smart_init_options=None, **kwargs):
if smart_init_options is None:
smart_init_options = {}
self.smart_init_options = smart_init_options
super().__init__(**kwargs)
if self.fit_dict['options']:
self.smart_init_options = {
**self.fit_dict['options'],
**self.smart_init_options
}
self.init_dict = None
self.mod = None
def run_init_model(self):
self.init_dict = self.get_init_dict(df=self.all_data,
groups=self.groups)
def update_init_model(self, df, group):
"""
Update the initial model with a re-fit model
from the specified group. Returns a new copy of the init dict
Args:
df: (pd.DataFrame) data used to update the init model
group: (str) the group to update
Returns:
"""
new_init_dict = deepcopy(self.init_dict)
new_init_dict.update(self.get_init_dict(df=df, groups=[group]))
return new_init_dict
def get_init_dict(self, df, groups):
"""
Run the init model for each location.
Args:
df: (pd.DataFrame) data frame to fit the model that will
be subset by group
groups: (str) groups to get in the dict
Returns:
(dict) dictionary of fixed effects keyed by group
"""
model = CurveModel(df=df, **self.basic_model_dict)
init_fit_dict = deepcopy(self.fit_dict)
init_fit_dict.update(options=self.smart_init_options)
init_dict = get_initial_params(groups=groups,
model=model,
fit_arg_dict=init_fit_dict)
return init_dict
def fit(self, df, group=None):
"""
Fits a loose, tight, beta, and p combinations model. If you pass in
update group it will override the initial parameters with new
initial parameters based on the df you pass.
Args:
df:
group: (str) passing in the group will update the initialization
dictionary (not replacing the old one) for this particular fit.
Returns:
"""
if group is not None:
init_dict = self.update_init_model(df=df, group=group)
else:
init_dict = deepcopy(self.init_dict)
fit_dict = deepcopy(self.fit_dict)
fe_init, re_init = compute_starting_params(init_dict)
fit_dict.update(fe_init=fe_init, re_init=re_init)
self.mod = CurveModel(df=df, **self.basic_model_dict)
self.mod.fit_params(**fit_dict)
def refresh(self):
self.mod = None