def _test_generic(self, estimator_class):
loo_scorer = validation.LinearSmootherLeaveOneOutScorer()
loo_scorer_alt = _LinearSmootherLeaveOneOutScorerAlternative()
x = np.linspace(-2, 2, 5)
fd = skfda.FDataGrid(x**2, x)
estimator = estimator_class()
grid = validation.SmoothingParameterSearch(estimator, [2, 3],
scoring=loo_scorer)
grid.fit(fd)
score = np.array(grid.cv_results_['mean_test_score'])
grid_alt = validation.SmoothingParameterSearch(estimator, [2, 3],
scoring=loo_scorer_alt)
grid_alt.fit(fd)
score_alt = np.array(grid_alt.cv_results_['mean_test_score'])
np.testing.assert_array_almost_equal(score, score_alt)
def smooth_datagrid(fd):
'''
Smoothes the values in the datagrid object
Args:
fd (datagrid): Includes data_matrix, x vector for contionous data
representation.
Returns:
knn_fd (datagrid): datagrid values smoothed with an kN Kernel.
'''
# smooth the values with kNeighbors kernel
param_values = np.linspace(start=2, stop=25, num=24)
knn = val.SmoothingParameterSearch(ks.KNeighborsSmoother(), param_values)
knn.fit(fd)
knn_fd = knn.transform(fd)
return knn_fd
# As an example, we will smooth the first 300 curves only. In the following
# plot, the first five curves are shown.
dataset = skfda.datasets.fetch_phoneme()
fd = dataset['data'][:300]
fd[0:5].plot()
##############################################################################
# Here we show the general cross validation scores for different values of the
# parameters given to the different smoothing methods.
param_values_knn = np.arange(1, 24, 2)
param_values_others = param_values_knn / 32
# Local linear regression kernel smoothing.
llr = val.SmoothingParameterSearch(ks.LocalLinearRegressionSmoother(),
param_values_others)
llr.fit(fd)
llr_fd = llr.transform(fd)
# Nadaraya-Watson kernel smoothing.
nw = val.SmoothingParameterSearch(ks.NadarayaWatsonSmoother(),
param_values_others)
nw.fit(fd)
nw_fd = nw.transform(fd)
# K-nearest neighbours kernel smoothing.
knn = val.SmoothingParameterSearch(ks.KNeighborsSmoother(), param_values_knn)
knn.fit(fd)
knn_fd = knn.transform(fd)
fig = plt.figure()
estimator = kde.fit(y)
estimator = np.reshape(estimator[0], df.shape[0])
plt.scatter(x, y)
plt.scatter(x, estimator, c='r')
plt.show()
# Using SKFDA
df_grid=skfda.FDataGrid(df)
bandwidth = np.arange(0.1, 5, 0.2)
llr = val.SmoothingParameterSearch(
ks.LocalLinearRegressionSmoother(),
bandwidth)
fit = llr.fit(df_grid)
llr_df = llr.transform(df_grid)
plt.scatter(x, y)
plt.xlabel('x')
plt.ylabel('y')
plt.scatter(x, llr_df, c='r')
plt.show()
# just K-fold cross-validation
"""
rn = range(1,26)
kf3 = sk_ms.KFold(3, shuffle=False)
for train_index, test_index in kf3.split(rn):
dataset = skfda.datasets.fetch_phoneme()
fd = dataset['data'][:300]
fd[0:5].plot()
plt.show()
n_neighbors = np.arange(1, 24)
scale_factor = ((fd.domain_range[0][1] - fd.domain_range[0][0]) /
len(fd.grid_points[0]))
bandwidth = n_neighbors * scale_factor
# K-nearest neighbours kernel smoothing.
knn = val.SmoothingParameterSearch(
ks.KNeighborsSmoother(),
n_neighbors,
)
knn.fit(fd)
knn_fd = knn.transform(fd)
# Local linear regression kernel smoothing.
llr = val.SmoothingParameterSearch(
ks.LocalLinearRegressionSmoother(),
bandwidth,
)
llr.fit(fd)
llr_fd = llr.transform(fd)
# Nadaraya-Watson kernel smoothing.
nw = val.SmoothingParameterSearch(
ks.NadarayaWatsonSmoother(),
def test_other(self, method='kernel'):
import skfda
import skfda.preprocessing.smoothing.kernel_smoothers as ks
import skfda.preprocessing.smoothing.validation as val
self.restore_model(self.resume_iters)
self.G.eval()
self.set_requires_grad(self.G, False)
self.iteration = self.test_len - self.num_clips + 1
self.model_save_dir = os.path.join(self.checkpoint_dir, self.model_dir)
graph_embedding = []
node_abnormal = []
graph_ab_list = []
with torch.no_grad():
idx = 0
self.subject = 0
while idx < self.iteration:
# =================================================================================== #
# 1. Preprocess input data(Unfinished) #
# =================================================================================== #
node_feature, edge, sensor, abnormal, graph_abnormal, idx = self.load_data_test(
idx)
node_abnormal.append(abnormal)
graph_ab_list.append(graph_abnormal.view(self.batch_size))
# =================================================================================== #
# 2. Train the Model #
# =================================================================================== #
_, _, _, E = self.G(node_feature, edge, sensor)
# if idx == 0:
# graph_embedding.append(E)
# else:
graph_embedding.append(torch.unsqueeze(E[-1], 0))
del E
del node_feature, edge
torch.cuda.empty_cache()
idx += 1
print("Finish NN Part!")
graph_embedding = torch.cat(graph_embedding)
graph_embedding = graph_embedding.view(graph_embedding.shape[0],
-1).cpu()
residual = []
for idx in range(len(self.subject_test)):
start = int(self.subject_test[idx, 1]) - idx * (self.num_clips - 1)
end = int(self.subject_test[idx,
2]) - (idx + 1) * (self.num_clips - 1)
if method == 'kernel':
fd = skfda.representation.grid.FDataGrid(
graph_embedding[start:end + 1])
n_neighbors = np.arange(1, 24)
scale_factor = (
(fd.domain_range[0][1] - fd.domain_range[0][0]) /
len(fd.grid_points[0]))
bandwidth = n_neighbors * scale_factor
nw = val.SmoothingParameterSearch(ks.NadarayaWatsonSmoother(),
bandwidth)
nw.fit(fd)
nw_fd = nw.transform(fd)
r = self.loss_function(torch.tensor(nw_fd.data_matrix),
torch.tensor(fd.data_matrix),
graph=False,
device=self.device)
residual.append(r)
residual = torch.cat(residual).view(graph_embedding.shape[0],
self.num_sensor_dev, -1)
residual = torch.mean(residual, dim=-1)
node_abnormal = torch.cat(node_abnormal)
print(residual.shape)
print(node_abnormal.shape)
predict_result(residual, node_abnormal, "sensor")
if graph_abnormal is not None:
graph_predict = torch.mean(residual, dim=-1)
graph_ab_list = torch.tensor(graph_ab_list)
predict_result(graph_predict, graph_ab_list, 'graph')
