def lasso_cv(y, x=None, cv=10, max_deg=3):
"""LASSO fit."""
if np.isnan(y).all():
y_pred = y
else:
y_pred = ap.lasso_cv(x, y, cv=cv, max_deg=max_deg, max_iter=1e3)
return y_pred
def std_series_filt(x, y, max_std=2, max_deg=3):
"""Filter entire vector if the detrended std > max_std."""
std = None
if not np.isnan(y).all():
i_notnan, = np.where(~np.isnan(y))
poly = lasso_cv(x[i_notnan], y[i_notnan], max_deg=max_deg)
std = np.nanstd(y[i_notnan] - poly) # std of detrended series
if std > max_std:
y[:] = np.nan
return y, std
def _peak_filt(x, y, n_std=3, max_deg=3):
"""Filter spikes in a vector.
See peak_filt()
"""
assert not np.isnan(y).all(), 'empty array'
y2 = y.copy()
i_notnan, = np.where(~np.isnan(y))
# detrend
poly = lasso_cv(x[i_notnan], y[i_notnan], max_deg=max_deg)
y2[i_notnan] = y[i_notnan] - poly
# filter
i_peaks, = np.where(np.abs(y2) > n_std * np.nanstd(y2))
y[i_peaks] = np.nan
return len(i_peaks)
s = df[k]
m, c = ap.linear_fit_robust(time, s.values, return_coef=True)
x, y = ap.linear_fit_robust(time, s.values, return_coef=False)
pol = np.polyfit(time, s.values, 2)
yy = np.polyval(pol, time)
#axs[i,j].fill_between(time, s.values+2*r, s.values-2*r, facecolor='0.5',
# edgecolor='w', alpha=0.2)
#axs[i,j].plot(time, zeros, ':', c='0.5', linewidth=0.5)
axs[i, j].plot(time, y, c='0.2', linewidth=0.75, zorder=2)
axs[i, j].plot(time, s.values, 's', c='0.5', markersize=4, zorder=1)
if 0:
axs[i, j].plot(time, yy, c='b', linewidth=1.5, zorder=2)
if 1:
axs[i, j].plot(time,
ap.lasso_cv(time, s, max_deg=3),
c='b',
linewidth=1.75,
zorder=4)
# set plots
#-------------------------------------------------------------
if i == 0:
ap.intitle('%s %.2f %s' % (k, m, UNITS),
ax=axs[i, j],
loc=8,
pad=-1,
borderalpha=0.7)
else:
ap.intitle('%s %.2f' % (k, m),
if 0: # subset
print 'subsetting...'
region = ap.dotson
data, _, _ = ap.get_subset(region, data, lon, lat)
nt, ny, nx = data.shape # i,j,k = t,y,x
#------------------------------------------------------
if 0: # plot alphas and MSEs (the prediction error curve)
N = 3
x = time
y = data[:,5,2] # 6,2 -> PIG
y = ap.referenced(y, to='mean')
y_pred, lasso = ap.lasso_cv(x, y, cv=10, max_deg=N, max_iter=1e3, return_model=True)
mse = lasso.mse_path_.mean(axis=1)
std = lasso.mse_path_.std(axis=1, ddof=1) / np.sqrt(10)
#plt.plot(np.log(lasso.alphas_), mse)
plt.errorbar(np.log(lasso.alphas_), mse, yerr=std)
plt.vlines(np.log(lasso.alpha_), ymin=mse.min(), ymax=mse.max(), color='r')
plt.xlabel('log(alpha)')
plt.ylabel('10-fold-average MSE')
plt.show()
exit()
#------------------------------------------------------
data = as_frame(data, time, lat, lon)
data = data.apply(ap.referenced, to='mean', raw=True)
