def corr_func(x, y, **kwargs):
r = np.corrcoeff(x, y)[0][1]
ax = plt.gca()
ax.annotate("r = {:.2f}".format(r),
xy=(0.2, 0.8),
xycoords=ax.transAxes,
size=20)
def corr(data_samples: np.array, data_truth: np.array, agg=None, **kwargs):
"""Computes the empirical correlation betnween actuals and predictions
:param data: Predicted time series values (n_timesteps, n_timeseries)
:param data_truth: Ground truth time series values
:param agg: Aggregator function that creates forecast out of samples
"""
agg = np.median if not agg else agg
data = agg(data_samples, axis=0)
return np.round(np.corrcoeff(data, data_truth, rowvar=False), 3)
def corr(data_samples: np.array, data_truth: np.array, agg=None, **kwargs):
"""Returns the Pearson correlation coefficient betnween observed values
and aggregated predictions.
:param data_samples: Predicted time series values (n_timesteps, n_timeseries).
:param data_truth: Actual values observed.
:param agg: Property of the forecast distribution to use for evaluation.
"""
agg = np.median if not agg else agg
data = agg(data_samples, axis=0)
return np.round(np.corrcoeff(data, data_truth, rowvar=False), 3)
def distances(chi1,
chi2,
type_comp='auc',
taus=[1, 10, 25, 50],
normalize=True,
plot=False,
savefig=False,
filefig='plots/change_graph.png'):
'''
Compare two graphs based on their diffusion properties: assumes that the nodes are identified
INPUT:
======================================================
chi1, chi2 : two graphs of type either nx or pygsp
type_comp : the distances between distributions
that should be used (default: auc)
taus : the scales used for heat diffusion propoagation
plot, savefig,filefig: additional parameters (for plotting and saving plots)
OUTPUT:
======================================================
distance : distances between diffusion distribution at different scales
'''
n_nodes, dim_embed = heat_print1.shape
n_filters = len(taus)
level_size = dim_embed / n_filters
distances = np.zeros((n_filters, n))
for m in range(n_filters):
index_scale = range(m * level_size, (m + 1) * level_size)
for i in range(n_nodes):
if type_comp == "corr":
distances[m, i] = 1 - np.corrcoeff(chi1[i, index_scale],
chi2[i, index_scale])
elif type_comp == "auc":
distances[m, i] = abs(
compute_evolution_heat_diff(i,
m,
heat_print1,
heat_print2,
mode_diff=mode_diff))
elif type_comp == "emd":
### Required params:
### P,Q - Two histograms of size H
### D - The HxH matrix of the ground distance between bins of P and Q
H = 30
hist1, bins_arr = np.histogram(heat_print1[m], H)
#### Normalize histogram
w = [
bins_arr[i + 1] - bins_arr[i]
for i in range(len(bins_arr) - 1)
]
hist1 = hist1 * 1.0 / np.matrix(w).dot(hist1)
hist2, _ = np.histogram(heat_print2[m], bins_arr)
hist2 = hist2 * 1.0 / np.matrix(w).dot(hist2)
hist1 = np.reshape(np.matrix(hist1), [1, H])
hist2 = np.reshape(np.matrix(hist2), [1, H])
D = np.zeros((H, H))
for i in range(H):
for j in range(H):
D[i, j] = np.abs(bins_arr[i + 1] - bins_arr[j + 1])
distances[m, i] = emd(np.array(hist1.tolist()[0]),
np.array(hist2.tolist()[0]), D)
else:
print 'comparison type not implemented'
return np.nan
if plot == True:
plt.figure()
sb.heatmap(distances, cmap="hot")
if savefig == True:
plt.savefig(filefig)
agg_score = np.sum(distances)
return distances, agg_score
# COV(X, Y) = E[(X - E[X])(Y-E[X])] # r = sum() # def mean(x): return sum(x) / len(x) # def coff(x, y): # mx = mean(x) # my = mean(y) # vx = mean( ) # xx = map(lambda x: x - mx) # yy = map(lambda x: x - my) # numerator = sum(map(lambda xy: xy[0] * xy[1], zip(xx, yy)) if __name__ == "__main__": main() x1 = 15,12,8,8,7,7,7,6,5,3 x2 = 10,25,17,11,13,17,20,13,9,15 print(np.corrcoeff(x1, x2))