def fit_predict(self,
X,
period_frequencies=None,
full_series=False,
num_iter=20):
if not isinstance(X, TimeSeriesDataset):
X = np.asanyarray(X, dtype='d')
else:
X = X.np_like_firstn()
n = X.shape[0]
if period_frequencies is None:
period_frequencies = np.ones(n)
else:
period_frequencies = np.asanyarray(period_frequencies, dtype='d')
assert period_frequencies.shape[0] == n
P = np.zeros(shape=(n, TOTAL_PARAMS), dtype='d')
with np.errstate(over='raise'):
for i in xrange(X.shape[0]):
P[i] = fit_one(X[i], period_frequencies[i], num_iter)
if full_series:
Y = np.zeros((n, X.shape[1] + self.steps_ahead), dtype='d')
else:
Y = np.zeros((n, ), dtype='d')
for i in xrange(X.shape[0]):
if full_series:
Y[i] = spikem(P[i], X.shape[1] + self.steps_ahead)
else:
Y[i] = spikem(P[i], X.shape[1] + self.steps_ahead)[-1]
return Y
def fit_one(tseries, period, num_iter=20):
duration = tseries.shape[0]
init_params, bounds = get_initial_parameters(tseries, period)
curr_params = init_params.copy()
nonz = np.where(tseries > 0)[0]
if nonz.shape[0] > 0:
st = max(nonz[0] - 1, 0)
else:
st = 0
ed = tseries.argmax()
def mse_func(lm_params, param_name, all_parameters):
param_value = lm_params[param_name].value
parameter_idx = PARAMETERS[param_name]
predicted = spikem_wrapper(param_value, parameter_idx, all_parameters,
tseries.shape[0])
return np.array([np.sqrt(((predicted - tseries)**2).mean())])
for _ in range(num_iter):
for param_name in TO_FIT:
param_idx = PARAMETERS[param_name]
if param_name == 'nb':
best = 0
min_mse = float('inf')
copy_params = curr_params.copy()
for i in xrange(st, ed):
copy_params[param_idx] = i
predicted = spikem(copy_params, tseries.shape[0])
try:
mse = np.sqrt(((predicted - tseries)**2).mean())
except: #overflow
mse = float('inf')
if mse < min_mse:
best = i
curr_params[param_idx] = best
else:
bound = bounds[param_idx]
lm_params = to_lmstyle(param_name, curr_params, bounds)
try:
lmfit.minimize(mse_func, lm_params, \
args=(param_name, curr_params))
curr_params[param_idx] = lm_params[param_name].value
except: #overflow
pass
return curr_params
