def TrainRFRegression(df1):
# generate the equation to use for our design matrices
eqn = build_eqn(df1, 'regressand', ['any_regressand', 'X25', 'X26', 'X29'])
# build our design matrices
y, X = dmatrices(eqn, data=df1, return_type='dataframe')
# employ clustering to reduce our dimensionality
X_reduction = FeatureAgglomeration(n_clusters=50).fit(X, pd.np.ravel(y))
reduced_X = X_reduction.transform(X)
# define our regressor
mod = RandomForestRegressor(n_estimators=50)
# fit our data
res = mod.fit(reduced_X, pd.np.ravel(y))
# evaluate our fit
yp = pd.DataFrame({'predicted': res.predict(reduced_X)})
yp = yp['predicted']
yt = y['regressand']
r2 = metrics.r2_score(yt, yp)
rmse = metrics.mean_absolute_error(yt, yp)
# save our model, including scalers and feature agglomerator
with open('RFR_trained_model.pickle', 'wb') as output:
pickle.dump(res, output, pickle.HIGHEST_PROTOCOL)
return r2, rmse
def test_feature_agglomoration(self, n, data, corpus, docvecs):
print("Using feature agglomoration to reduce the matrix' dimensionality...")
affinities = ["euclidean", "l1", "l2", "manhattan", "cosine", "precomputed"]
linkages = ["ward", "complete", "average"]
agglos = []
for linkage in linkages:
if linkage is "ward":
agglos.append(FeatureAgglomeration(n_clusters=n, affinity="euclidean", linkage=linkage))
else:
for affinity in affinities:
agglos.append(FeatureAgglomeration(n_clusters=n, affinity=affinity, linkage=linkage))
for agglo in agglos:
print(agglo.get_params)
reduced_vectors = agglo.fit_transform(data)
clusters_kmeans = self.cluster_kmeans(docvecs, reduced_vectors=reduced_vectors, feature_names=corpus)
labels_db = self.cluster_dbscan(reduced_vectors)
#labels_hdb = self.cluster_hdbscan(reduced_vectors)
clusters_db = self.get_clusters(corpus, labels_db)
#clusters_hdb = self.get_clusters(corpus, labels_hdb)
#agglo = FeatureAgglomeration(n_clusters=n, affinity="euclidean", linkage="ward")
#return agglo.fit_transform(data)
return
class Regressor(BaseEstimator):
def __init__(self):
self.clf = Pipeline([
("RF", RandomForestRegressor(n_estimators=200, max_depth=15,
n_jobs=N_JOBS))])
self.scaler = StandardScaler()
self.agglo = FeatureAgglomeration(n_clusters=500)
def fit(self, X, y):
y = y.ravel()
n_samples, n_lags, n_lats, n_lons = X.shape
self.scaler.fit(X[:, -1].reshape(n_samples, -1))
X = X.reshape(n_lags * n_samples, -1)
connectivity = grid_to_graph(n_lats, n_lons)
self.agglo.connectivity = connectivity
X = self.scaler.transform(X)
X = self.agglo.fit_transform(X)
X = X.reshape(n_samples, -1)
self.clf.fit(X, y)
def predict(self, X):
n_samples, n_lags, n_lats, n_lons = X.shape
X = X.reshape(n_lags * n_samples, -1)
X = self.scaler.transform(X)
X = self.agglo.transform(X)
X = X.reshape(n_samples, -1)
return self.clf.predict(X)
def FeatureSelection(df,numeric_cols , corrCoefThres=0.9):
numeric_cols = numeric_cols
numdf = df[numeric_cols]
r_in_x = numdf.corr()
r_in_x = abs(r_in_x)
distance_in_x = 1 / r_in_x
for i in range(r_in_x.shape[0]):
distance_in_x.iloc[i, i] = 10 ^ 10
cpdist = distance_in_x.copy()
cpdist = cpdist.fillna(cpdist.max().max())
#df.isna().sum()
from scipy.spatial.distance import correlation
from sklearn.cluster import FeatureAgglomeration
corrcoefmin = corrCoefThres
fa = FeatureAgglomeration(n_clusters=None,affinity="precomputed",compute_full_tree=True, linkage="average" ,distance_threshold=1/corrcoefmin)
fa.fit(cpdist)
numdf.shape[1]
fa.n_clusters_
fadf = pd.DataFrame({"feature":numdf.columns.values , "label":fa.labels_})
selectedFeatures = fadf.groupby("label").head(1)["feature"].values
return selectedFeatures
def test_feature_agglomeration():
n_clusters = 1
X = np.array([0, 0, 1]).reshape(1, 3) # (n_samples, n_features)
agglo_mean = FeatureAgglomeration(n_clusters=n_clusters,
pooling_func=np.mean)
agglo_median = FeatureAgglomeration(n_clusters=n_clusters,
pooling_func=np.median)
assert_no_warnings(agglo_mean.fit, X)
assert_no_warnings(agglo_median.fit, X)
assert np.size(np.unique(agglo_mean.labels_)) == n_clusters
assert np.size(np.unique(agglo_median.labels_)) == n_clusters
assert np.size(agglo_mean.labels_) == X.shape[1]
assert np.size(agglo_median.labels_) == X.shape[1]
# Test transform
Xt_mean = agglo_mean.transform(X)
Xt_median = agglo_median.transform(X)
assert Xt_mean.shape[1] == n_clusters
assert Xt_median.shape[1] == n_clusters
assert Xt_mean == np.array([1 / 3.])
assert Xt_median == np.array([0.])
# Test inverse transform
X_full_mean = agglo_mean.inverse_transform(Xt_mean)
X_full_median = agglo_median.inverse_transform(Xt_median)
assert np.unique(X_full_mean[0]).size == n_clusters
assert np.unique(X_full_median[0]).size == n_clusters
assert_array_almost_equal(agglo_mean.transform(X_full_mean), Xt_mean)
assert_array_almost_equal(agglo_median.transform(X_full_median), Xt_median)
def feature_agglomeration(voters_data, n, rounding=False):
featagg = FeatureAgglomeration(n_clusters=n)
featagg.fit(voters_data)
condensed = featagg.transform(voters_data)
feature_groups_map = dict(zip(voters_data.columns, featagg.labels_))
feature_groups_nos = []
for feature_group_key in feature_groups_map:
feature_groups_nos.append(feature_groups_map[feature_group_key])
feature_groups_nos
group_labels = []
for feature_group_no in set(feature_groups_nos):
group_label = ""
for feature_groups_key in feature_groups_map:
if feature_groups_map[feature_groups_key] == feature_group_no:
group_label = group_label + feature_groups_key + ", "
group_labels.append(group_label[0:-2])
group_labels
voters_agglomerated = pd.DataFrame(condensed,
columns=group_labels,
index=voters_data.index)
if rounding == True:
voters_agglomerated = voters_agglomerated.applymap(lambda x: round(x))
print("🔹→💠←🔹 {} features agglomerated into {} hybrid features.".format(
len(voters_data.columns), len(voters_agglomerated.columns)))
return voters_agglomerated
def cluster(self, cluster_images, cluster_db_path, diffnet_paht='model_checkpoints/', save_csv=False):
compressions = []
# Finding features
diffnet = DiffNet(self.db, db_path=self.db_path)
diffnet.restore(diffnet_paht)
print("Calculating features for", len(cluster_images), "images")
for img in cluster_images:
print("Finding features for:", img)
one_hot = diffnet.feedforward(img, cluster_db_path)
output = self.sess.run(self.compressed, feed_dict={self.Q: [one_hot], self.keep_prob: 1.})
compressions.append(output[0])
# Clustering
print("Performing clustering...")
compressions = np.array(compressions)
fa = FeatureAgglomeration(n_clusters=30)
X_clusters = fa.fit_transform(compressions)
print("Collecting data...")
csv_dict_arr = []
for i, img in enumerate(cluster_images):
csv_dict_arr.append({'1.img': img, '2.class': np.argmax(X_clusters[i]), '3.features': compressions[i]})
# Saving
if save_csv:
print("Saving data to csv...")
keys = load_label_list(csv_dict_arr[0])
with open('cluster_result.csv', 'w') as output_file:
dict_writer = csv.DictWriter(output_file, keys, delimiter=';')
dict_writer.writeheader()
dict_writer.writerows(csv_dict_arr)
return csv_dict_arr
def _feature_agglomeration_fit_method(data, n_parcels, connectivity, linkage):
"""Feature Agglomeration algorithm to fit on the data.
Parameters
----------
data : array_like, shape=(n_samples, n_voxels)
Masked subjects data
n_parcels : int
Number of parcels to parcellate.
connectivity : ndarray
Connectivity matrix
Defines for each feature the neighbouring features following a given
structure of the data.
linkage : str
which linkage criterion to use.
'ward' or 'linkage' or 'average'
Returns
-------
labels : ndarray
Labels to the data
"""
ward = FeatureAgglomeration(n_clusters=n_parcels, connectivity=connectivity,
linkage=linkage)
ward.fit(data)
return ward.labels_
def TestSGDRegression(df1):
# generate the equation to use for our design matrices
eqn = build_eqn(df1, 'regressand',
['any_regressand', 'X25', 'X26', 'X29', 'X13'])
# build our design matrices
X = dmatrix(eqn.replace('regressand ~ ', '0+'),
data=df1,
return_type='dataframe')
# load our model, including scalers and feature agglomerator
with open('SGD_trained_model.pickle', 'rb') as input:
res = pickle.load(input)
# employ clustering to reduce our dimensionality
X_reduction = FeatureAgglomeration(n_clusters=n_clusters).fit(X)
reduced_X = X_reduction.transform(X)
# standardize our data
X_scaler = StandardScaler().fit(reduced_X)
std_X = X_scaler.transform(reduced_X)
# predict the interest rates
yp = res.predict(std_X)
return yp
def feature_agglomeration(X, args={}):
"""
使用层次聚类对特征进行聚类，然后进行特征降维
"""
from sklearn.cluster import FeatureAgglomeration
fam = FeatureAgglomeration(**args)
fam.fit(X)
return fam
class FeatureAgglomerationDecomposer(Transformer):
type = 11
def __init__(self, n_clusters=2, affinity='euclidean', linkage='ward', pooling_func='mean',
random_state=1):
super().__init__("feature_agglomeration_decomposer")
self.input_type = [NUMERICAL, DISCRETE, CATEGORICAL]
self.compound_mode = 'only_new'
self.output_type = NUMERICAL
self.n_clusters = n_clusters
self.affinity = affinity
self.linkage = linkage
self.pooling_func = pooling_func
self.random_state = random_state
self.pooling_func_mapping = dict(mean=np.mean, median=np.median, max=np.max)
@ease_trans
def operate(self, input_datanode, target_fields=None):
from sklearn.cluster import FeatureAgglomeration
X, y = input_datanode.data
if self.model is None:
self.n_clusters = int(self.n_clusters)
n_clusters = min(self.n_clusters, X.shape[1])
if not callable(self.pooling_func):
self.pooling_func = self.pooling_func_mapping[self.pooling_func]
self.model = FeatureAgglomeration(
n_clusters=n_clusters, affinity=self.affinity,
linkage=self.linkage, pooling_func=self.pooling_func)
self.model.fit(X)
X_new = self.model.transform(X)
return X_new
@staticmethod
def get_hyperparameter_search_space(dataset_properties=None, optimizer='smac'):
cs = ConfigurationSpace()
n_clusters = UniformIntegerHyperparameter("n_clusters", 2, 400, default_value=25)
affinity = CategoricalHyperparameter(
"affinity", ["euclidean", "manhattan", "cosine"], default_value="euclidean")
linkage = CategoricalHyperparameter(
"linkage", ["ward", "complete", "average"], default_value="ward")
pooling_func = CategoricalHyperparameter(
"pooling_func", ["mean", "median", "max"], default_value="mean")
cs.add_hyperparameters([n_clusters, affinity, linkage, pooling_func])
affinity_and_linkage = ForbiddenAndConjunction(
ForbiddenInClause(affinity, ["manhattan", "cosine"]),
ForbiddenEqualsClause(linkage, "ward"))
cs.add_forbidden_clause(affinity_and_linkage)
return cs
def untangle(X: Iterable,
y: Iterable,
n_clusters: int = None,
get_connectivity: bool = True,
compute_distances: bool = True,
kind: str = 'correlation',
agglo_kws: Union[dict, Bunch] = None) -> FeatureAgglomeration:
from nilearn.connectome import ConnectivityMeasure as CM
from sklearn.cluster import FeatureAgglomeration
from sklearn.covariance import LedoitWolf
from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import mutual_info_classif
agglo_defs = dict(affinity='euclidean',
compute_full_tree='auto',
linkage='ward',
pooling_func=np.mean,
distance_threshold=None,
compute_distances=compute_distances)
if get_connectivity is True:
connect_mat = CM(LedoitWolf(), kind=kind).fit_transform([X.values])[0]
else:
connect_mat = None
if n_clusters is None:
n_clusters = divmod(X.shape[1], 2)[0] - 1
if n_clusters == 0:
n_clusters = 1
if agglo_kws is None:
agglo_kws = {}
agglo_defs.update(agglo_kws)
agglo = FeatureAgglomeration(n_clusters=n_clusters,
connectivity=connect_mat,
**agglo_defs)
if not isinstance(y, pd.Series):
y = pd.Series(y)
if not isinstance(X, pd.DataFrame):
X = pd.DataFrame(X)
agglo.fit(X, y)
setattr(
agglo, 'cluster_indexes_',
pd.DataFrame(zip(agglo.labels_, agglo.feature_names_in_),
columns=['cluster',
'feature']).groupby('cluster').feature)
skb = SelectKBest(k=1, score_func=mutual_info_classif)
factor_leaders_ = [
skb.fit(X[itm[1]], y).get_feature_names_out()[0]
for itm in tuple(agglo.cluster_indexes_)
]
setattr(agglo, 'factor_leaders_', factor_leaders_)
return agglo
def featureagglomeration(data_train, data_test, label_train, label_test, args):
print('feature agglomeration')
FA = FeatureAgglomeration(n_clusters=10).fit(data_train)
transformation = FA.transform(data_test)
agglomeration = find_highest(transformation)
print('feature agglomeration done')
compare_class(agglomeration, label_test)
if args.create_mean:
create_images_from_rows('fa', mean_image(agglomeration, data_test))
def get_encoder(metas, train_data, target_output_dim):
tmpdir = metas['workspace']
model_path = os.path.join(tmpdir, 'feature_agglomeration.model')
model = FeatureAgglomeration(n_clusters=target_output_dim)
model.fit(train_data)
pickle.dump(model, open(model_path, 'wb'))
return FeatureAgglomerationEncoder(model_path=model_path)
def do_feature_agglomoration(self, data):
print("Using feature agglomoration to reduce the matrix' dimensionality...")
if self.k:
n = self.k
else:
n = 20
agglo = FeatureAgglomeration(n_clusters=n, affinity="cosine", linkage="complete")
return agglo.fit_transform(data)
def token_cluster(self, n_clusters=300):
from scipy import sparse
from sklearn.cluster import FeatureAgglomeration
FA = FeatureAgglomeration(n_clusters=3000)
self.bow_corpus = FA.fit_transform(self.bow_corpus)
self.bow_corpus = sparse.csr_matrix(self.bow_corpus)
def get_clusters(X: pd.DataFrame, n_clusters: int):
clt = FeatureAgglomeration(n_clusters=n_clusters)
clt.fit(X)
clusters = []
for i in range(n_clusters):
clusters.append(X.columns[clt.labels_ == i].tolist())
return clusters # type: list[str]
def dim_reduction_FA(data, distance_threshold=0.45):
"""
Params:
data: ndarry of shape (n_samples, n_features)
distance_threshold: Optimal threshold value for similarity measure
Returns: (reducedDimData, nReducedComponents)
"""
agglo = FeatureAgglomeration(n_clusters=None,distance_threshold=distance_threshold)
reducedDimData = agglo.fit_transform(data)
return reducedDimData, agglo.n_clusters_
def apply_feature_agglomeration(table, features, label, n_components):
from sklearn.cluster import FeatureAgglomeration
from paje import feature_file_processor
x, y = feature_file_processor.split_features_target(table, features, label)
fa = FeatureAgglomeration(n_clusters=n_components, linkage='ward')
pc = fa.fit_transform(x)
return feature_file_processor.generate_data_frame(pc, table[[label]])
def test_feature_agglomeration_feature_names_out():
"""Check `get_feature_names_out` for `FeatureAgglomeration`."""
X, _ = make_blobs(n_features=6, random_state=0)
agglo = FeatureAgglomeration(n_clusters=3)
agglo.fit(X)
n_clusters = agglo.n_clusters_
names_out = agglo.get_feature_names_out()
assert_array_equal([f"featureagglomeration{i}" for i in range(n_clusters)],
names_out)
def main():
# Parameters
data_directory = '../../data/generated-data-r-10-n-6-4/'
features_path = '../../data/features-generated-data-r-10-n-6-4'
booking_file = '../../data/booking.csv'
users_file = '../../data/user.csv'
rating_thresholds = []
true_objects_indexes = [0, 1, 2, 3, 4, 5]
false_objects_indexes = [6, 7, 8, 9]
file_names = os.listdir(data_directory)
img_ids_vector = [int(name.split('-')[0]) for name in file_names]
ratings_vector = [int(name.split('-')[-2]) for name in file_names]
name_vector = [data_directory + name for name in file_names]
images_indexes = [name.split('-')[3].split('.')[0] for name in file_names]
ratings_matrix, images_indexes_for_id, ids_indexes, users_matrix = load_data(
data_directory, booking_file, users_file, rating_thresholds)
features = get_features(features_path, name_vector)
fa = FeatureAgglomeration(n_clusters=50)
fa.fit(features)
features = fa.transform(features)
scores_auc = []
scores_rmse = []
for i in range(10):
cv_results_file = '../results/cv-generated-data-r-10-n-6-4-rf-fa-' + str(
i) + '.csv'
selection = ObjectSelection(show_selection_results=False,
selection_algorithm='rf')
selection.transform(ids=img_ids_vector,
features=features,
ratings=ratings_vector,
users_ratings=ratings_matrix,
users=users_matrix,
cv_results_file=cv_results_file,
images_indexes=images_indexes,
true_objects_indexes=true_objects_indexes,
false_objects_indexes=false_objects_indexes,
paths=name_vector,
z_score=False)
selection.evaluate(evaluation_metric='auc')
selection.evaluate(evaluation_metric='rmse')
print('\n\n-----\n\n')
score_auc, score_rmse = selection.evaluate(evaluation_metric='auc')
scores_auc.append(score_auc)
scores_rmse.append(score_rmse)
results_file = '../scores/generated-data-r-10-n-6-4-rf-fa-auc.csv'
save_scores(scores_auc, results_file)
results_file = '../scores/generated-data-r-10-n-6-4-rf-fa-rmse.csv'
save_scores(scores_rmse, results_file)
def getDR(dt_all, n_comp=12):
# cols
cols_encode_label = dt_all.filter(
regex="Encode_Label").columns.values.tolist()
cols_cat = dt_all.drop(
"ID", axis=1).select_dtypes(include=["object"]).columns.tolist()
# standardize
dt_all_norm = MinMaxScaler().fit_transform(
dt_all.drop(["y", "Fold"] + cols_cat + cols_encode_label, axis=1))
# tSVD
tsvd = TruncatedSVD(n_components=n_comp, random_state=420)
tsvd_results = tsvd.fit_transform(dt_all_norm)
# PCA
pca = PCA(n_components=n_comp, random_state=420)
pca_results = pca.fit_transform(dt_all_norm)
# ICA
ica = FastICA(n_components=n_comp, max_iter=5000, random_state=420)
ica_results = ica.fit_transform(dt_all_norm)
# GRP
grp = GaussianRandomProjection(n_components=n_comp,
eps=0.1,
random_state=420)
grp_results = grp.fit_transform(dt_all_norm)
# SRP
srp = SparseRandomProjection(n_components=n_comp,
dense_output=True,
random_state=420)
srp_results = srp.fit_transform(dt_all_norm)
# NMF
nmf = NMF(n_components=n_comp, init='nndsvdar', random_state=420)
nmf_results = nmf.fit_transform(dt_all_norm)
# F*G
f*g = FeatureAgglomeration(n_clusters=n_comp, linkage='ward')
fag_results = f*g.fit_transform(dt_all_norm)
# Append decomposition components to datasets
for i in range(1, n_comp + 1):
dt_all['DR_TSVD_' + str(i)] = tsvd_results[:, i - 1]
dt_all['DR_PCA_' + str(i)] = pca_results[:, i - 1]
dt_all['DR_ICA_' + str(i)] = ica_results[:, i - 1]
dt_all['DR_GRP_' + str(i)] = grp_results[:, i - 1]
dt_all['DR_SRP_' + str(i)] = srp_results[:, i - 1]
dt_all['DR_NMF_' + str(i)] = nmf_results[:, i - 1]
dt_all['DR_FAG_' + str(i)] = fag_results[:, i - 1]
return (dt_all)
def setup(self, keywords={}):
"""
Setup the algorithms
"""
for p in keywords.keys():
setattr(self, p, keywords[p])
if self.method == "agglomerative": self.obj = AgglomerativeClustering(n_clusters=self.n_clusters, linkage=self.linkage,
affinity=self.affinity)
if self.method == "feature": self.obj = FeatureAgglomeration(n_clusters=self.n_clusters, linkage=self.linkage,
affinity=self.affinity, distance_threshold=self.distance_threshold)
return
def variable_clustering(self, X_cat, woe_iv_df, n_clusters=15):
X_transformed = mt.BinWoe().transform_x_all(X_cat, woe_iv_df)
agglo = FeatureAgglomeration(n_clusters=n_clusters)
if len(X_transformed) > 20000:
X_agglo = X_transformed.sample(20000)
else:
X_agglo = X_transformed.copy()
agglo.fit(X_agglo)
vars_clusters = pd.DataFrame(data={'指标英文':X_transformed.columns.tolist(),
'cluster':list(agglo.labels_)})\
.sort_values('cluster')
return vars_clusters, X_transformed
def cont_feature_clusters_sklearn(self, n_clusters = 5):
""" This uses feature agglomeration from scikit learn and only works for continuous variables
Eventually expand this to categorical variables using Cramer's V covariance matrix similar to
R tool using the iclust package """
#Import the library
from sklearn.cluster import FeatureAgglomeration
Cluster = FeatureAgglomeration(n_clusters=n_clusters)
Cluster.fit(self._dataset.iloc[:,self._cont_index_predictors])
df = pd.DataFrame({'Variable':self._dataset.columns[self._cont_index_predictors], 'Cluster':Cluster.labels_})
return df.sort_values(by='Cluster')
def fa_dim_red(x_train_scaled, dataset_name, features_num = 2):
z=0
losses = []
for k in range(1, x_train_scaled.shape[1]+1):
fa = FeatureAgglomeration(n_clusters=k)
fa_result = fa.fit_transform(x_train_scaled)
x_projected_fa = fa.inverse_transform(fa_result)
loss = ((x_train_scaled - x_projected_fa) ** 2).mean()
losses.append(loss)
np_feature_losses_percent = np.multiply(100, losses/np.sum(losses))
print('num of clustrs < 10% loss')
for i in range(len(np_feature_losses_percent)):
z=z+np_feature_losses_percent[i]
if z>90:
print(i+1)
break
print(np_feature_losses_percent)
plt.bar(list(range(1,len(np_feature_losses_percent)+1)),np_feature_losses_percent)
plt.title("FeatureAgglomeration Projection Losses % ("+str(dataset_name)+")")
plt.ylabel("Mean Squared Error (% of Total)")
plt.xlabel("Features")
plt.savefig((str(dataset_name))+' fa analysis.png')
plt.show()
fa = FeatureAgglomeration(n_clusters=features_num)
fa_result = fa.fit_transform(x_train_scaled, y_train)
print(fa_result.shape)
x_projected_fa = fa.inverse_transform(fa_result)
print(x_projected_ica.shape)
print(x_train_scaled.shape)
loss = ((x_train_scaled - x_projected_fa) ** 2).mean()
print('loss')
print(loss)
return fa_result,x_projected_fa
def _ward_fit_transform(all_subjects_data, fit_samples_indices,
connectivity, n_parcels, offset_labels):
"""Ward clustering algorithm on a subsample and apply to the whole dataset.
Computes a brain parcellation using Ward's clustering algorithm on some
images, then averages the signal within parcels in order to reduce the
dimension of the images of the whole dataset.
This function is used with Randomized Parcellation Based Inference, so we
need to save the labels to further perform the inverse transformation
operation. The function therefore needs an offset to be applied on the
labels so that they are unique across parcellations.
Parameters
----------
all_subjects_data : array_like, shape=(n_samples, n_voxels)
Masked subject images as an array.
fit_samples_indices : array-like,
Indices of the samples used to compute the parcellation.
connectivity : scipy.sparse.coo_matrix,
Graph representing the spatial structure of the images (i.e. connections
between voxels).
n_parcels : int,
Number of parcels for the parcellations.
offset_labels : int,
Offset for labels numbering.
The purpose is to have different labels in all the parcellations that
can be built by multiple calls to the current function.
Returns
-------
parcelled_data : numpy.ndarray, shape=(n_samples, n_parcels)
Average signal within each parcel for each subject.
labels : np.ndarray, shape=(n_voxels,)
Labels giving the correspondance between voxels and parcels.
"""
# fit part
data_fit = all_subjects_data[fit_samples_indices]
ward = FeatureAgglomeration(n_clusters=n_parcels,
connectivity=connectivity)
ward.fit(data_fit)
# transform part
labels = ward.labels_ + offset_labels # unique labels across parcellations
parcelled_data = ward.transform(all_subjects_data)
return parcelled_data, labels
def _ward_fit_transform(all_subjects_data, fit_samples_indices, connectivity,
n_parcels, offset_labels):
"""Ward clustering algorithm on a subsample and apply to the whole dataset.
Computes a brain parcellation using Ward's clustering algorithm on some
images, then averages the signal within parcels in order to reduce the
dimension of the images of the whole dataset.
This function is used with Randomized Parcellation Based Inference, so we
need to save the labels to further perform the inverse transformation
operation. The function therefore needs an offset to be applied on the
labels so that they are unique across parcellations.
Parameters
----------
all_subjects_data : array_like, shape=(n_samples, n_voxels)
Masked subject images as an array.
fit_samples_indices : array-like,
Indices of the samples used to compute the parcellation.
connectivity : scipy.sparse.coo_matrix,
Graph representing the spatial structure of the images (i.e. connections
between voxels).
n_parcels : int,
Number of parcels for the parcellations.
offset_labels : int,
Offset for labels numbering.
The purpose is to have different labels in all the parcellations that
can be built by multiple calls to the current function.
Returns
-------
parcelled_data : numpy.ndarray, shape=(n_samples, n_parcels)
Average signal within each parcel for each subject.
labels : np.ndarray, shape=(n_voxels,)
Labels giving the correspondance between voxels and parcels.
"""
# fit part
data_fit = all_subjects_data[fit_samples_indices]
ward = FeatureAgglomeration(n_clusters=n_parcels,
connectivity=connectivity)
ward.fit(data_fit)
# transform part
labels = ward.labels_ + offset_labels # unique labels across parcellations
parcelled_data = ward.transform(all_subjects_data)
return parcelled_data, labels
def test_feature_agglomeration():
n_clusters = 1
X = np.array([0, 0, 1]).reshape(1, 3) # (n_samples, n_features)
agglo_mean = FeatureAgglomeration(n_clusters=n_clusters,
pooling_func=np.mean)
agglo_median = FeatureAgglomeration(n_clusters=n_clusters,
pooling_func=np.median)
assert_no_warnings(agglo_mean.fit, X)
assert_no_warnings(agglo_median.fit, X)
assert_true(np.size(np.unique(agglo_mean.labels_)) == n_clusters)
assert_true(np.size(np.unique(agglo_median.labels_)) == n_clusters)
assert_true(np.size(agglo_mean.labels_) == X.shape[1])
assert_true(np.size(agglo_median.labels_) == X.shape[1])
# Test transform
Xt_mean = agglo_mean.transform(X)
Xt_median = agglo_median.transform(X)
assert_true(Xt_mean.shape[1] == n_clusters)
assert_true(Xt_median.shape[1] == n_clusters)
assert_true(Xt_mean == np.array([1 / 3.]))
assert_true(Xt_median == np.array([0.]))
# Test inverse transform
X_full_mean = agglo_mean.inverse_transform(Xt_mean)
X_full_median = agglo_median.inverse_transform(Xt_median)
assert_true(np.unique(X_full_mean[0]).size == n_clusters)
assert_true(np.unique(X_full_median[0]).size == n_clusters)
assert_array_almost_equal(agglo_mean.transform(X_full_mean),
Xt_mean)
assert_array_almost_equal(agglo_median.transform(X_full_median),
Xt_median)
def cluster_sentences(sentences, nb_of_clusters=5):
tfidf_vectorizer = TfidfVectorizer(tokenizer=word_tokenizer,
stop_words=stopwords.words('english'),
max_df=0.9,
min_df=0.05,
lowercase=True)
#builds a tf-idf matrix for the sentences
tfidf_matrix_1 = tfidf_vectorizer.fit_transform(sentences)
tfidf_matrix = tfidf_matrix_1.todense()
kmeans = FeatureAgglomeration(n_clusters=nb_of_clusters)
kmeans.fit(tfidf_matrix)
clusters = collections.defaultdict(list)
for i, label in enumerate(kmeans.labels_):
clusters[label].append(i)
return dict(clusters)
def data_compression(fmri_masked, mask_img, mask_np, output_size):
"""
data : array_like
A matrix of shape (`V`, `N`) with `V` voxels `N` timepoints
The functional dataset that needs to be reduced
mask : a numpy array of the mask
output_size : integer
The number of elements that the data should be reduced to
"""
## Transform nifti files to a data matrix with the NiftiMasker
import time
from nilearn import input_data
datacompressiontime = time.time()
nifti_masker = input_data.NiftiMasker(mask_img=mask_img,
memory='nilearn_cache',
mask_strategy='background',
memory_level=1,
standardize=False)
ward = []
# Perform Ward clustering
from sklearn.feature_extraction import image
shape = mask_np.shape
connectivity = image.grid_to_graph(n_x=shape[0],
n_y=shape[1],
n_z=shape[2],
mask=mask_np)
#import pdb;pdb.set_trace()
from sklearn.cluster import FeatureAgglomeration
start = time.time()
ward = FeatureAgglomeration(n_clusters=output_size,
connectivity=connectivity,
linkage='ward')
ward.fit(fmri_masked)
#print("Ward agglomeration compressing voxels into clusters: %.2fs" % (time.time() - start))
labels = ward.labels_
#print ('Extracting reduced Dimension Data')
data_reduced = ward.transform(fmri_masked)
fmri_masked = []
#print('Data compression took ', (time.time()- datacompressiontime), ' seconds')
return {'data': data_reduced, 'labels': labels}
def test_linkage_misc():
# Misc tests on linkage
rng = np.random.RandomState(42)
X = rng.normal(size=(5, 5))
with pytest.raises(ValueError):
AgglomerativeClustering(linkage='foo').fit(X)
with pytest.raises(ValueError):
linkage_tree(X, linkage='foo')
with pytest.raises(ValueError):
linkage_tree(X, connectivity=np.ones((4, 4)))
# Smoke test FeatureAgglomeration
FeatureAgglomeration().fit(X)
# test hierarchical clustering on a precomputed distances matrix
dis = cosine_distances(X)
res = linkage_tree(dis, affinity="precomputed")
assert_array_equal(res[0], linkage_tree(X, affinity="cosine")[0])
# test hierarchical clustering on a precomputed distances matrix
res = linkage_tree(X, affinity=manhattan_distances)
assert_array_equal(res[0], linkage_tree(X, affinity="manhattan")[0])
def comput_coefs(self, X, y, size):
cv = KFold(2) # cross-validation generator for model selection
ridge = BayesianRidge()
cachedir = tempfile.mkdtemp()
mem = Memory(cachedir=cachedir, verbose=1)
# Ward agglomeration followed by BayesianRidge
connectivity = grid_to_graph(n_x=size, n_y=size)
ward = FeatureAgglomeration(n_clusters=10, connectivity=connectivity,
memory=mem)
clf = Pipeline([('ward', ward), ('ridge', ridge)])
# Select the optimal number of parcels with grid search
clf = GridSearchCV(clf, {'ward__n_clusters': [10, 20, 30]}, n_jobs=1, cv=cv)
clf.fit(X, y) # set the best parameters
coef_ = clf.best_estimator_.steps[-1][1].coef_
coef_ = clf.best_estimator_.steps[0][1].inverse_transform(coef_)
coef_agglomeration_ = coef_.reshape(size, size)
# Anova univariate feature selection followed by BayesianRidge
f_regression = mem.cache(feature_selection.f_regression) # caching function
anova = feature_selection.SelectPercentile(f_regression)
clf = Pipeline([('anova', anova), ('ridge', ridge)])
# Select the optimal percentage of features with grid search
clf = GridSearchCV(clf, {'anova__percentile': [5, 10, 20]}, cv=cv)
clf.fit(X, y) # set the best parameters
coef_ = clf.best_estimator_.steps[-1][1].coef_
coef_ = clf.best_estimator_.steps[0][1].inverse_transform(coef_.reshape(1, -1))
coef_selection_ = coef_.reshape(size, size)
return dict(
coef_selection_=coef_selection_,
coef_agglomeration_=coef_agglomeration_,
cachedir=cachedir
)
def createPipe(embed, classif, nmca, aggregation, nsubs):
# Dimension Reduction
n_comp = 20 if nsubs > 70 else 15
if embed == "pca":
emb = ('pca', PCA(n_components=n_comp))
else:
emb = ('fa', FeatureAgglomeration(n_clusters=n_comp))
# Classifiers
neib = int(nmca * nsubs * 0.1) if aggregation == "mega" else int(nsubs *
0.1)
clfs = {
'svc':
('svc', SVC(class_weight="balanced", probability=True, max_iter=1e6)),
'knn': ('knn', KNeighborsClassifier(n_neighbors=neib)),
'rfc': ('rfc', RandomForestClassifier(class_weight="balanced")),
'ada': ('ada', AdaBoostClassifier()),
'lrc': ('lrc',
LogisticRegression(class_weight="balanced",
solver='liblinear',
max_iter=1e6))
}
pipe = Pipeline(steps=[emb, clfs[classif]])
return pipe
def test_ward_agglomeration():
"""
Check that we obtain the correct solution in a simplistic case
"""
rnd = np.random.RandomState(0)
mask = np.ones([10, 10], dtype=np.bool)
X = rnd.randn(50, 100)
connectivity = grid_to_graph(*mask.shape)
assert_warns(DeprecationWarning, WardAgglomeration)
with warnings.catch_warnings(record=True) as warning_list:
warnings.simplefilter("always", DeprecationWarning)
if hasattr(np, 'VisibleDeprecationWarning'):
# Let's not catch the numpy internal DeprecationWarnings
warnings.simplefilter('ignore', np.VisibleDeprecationWarning)
ward = WardAgglomeration(n_clusters=5, connectivity=connectivity)
ward.fit(X)
assert_equal(len(warning_list), 1)
agglo = FeatureAgglomeration(n_clusters=5, connectivity=connectivity)
agglo.fit(X)
assert_array_equal(agglo.labels_, ward.labels_)
assert_true(np.size(np.unique(agglo.labels_)) == 5)
X_red = agglo.transform(X)
assert_true(X_red.shape[1] == 5)
X_full = agglo.inverse_transform(X_red)
assert_true(np.unique(X_full[0]).size == 5)
assert_array_almost_equal(agglo.transform(X_full), X_red)
# Check that fitting with no samples raises a ValueError
assert_raises(ValueError, agglo.fit, X[:0])
def test_ward_agglomeration():
"""
Check that we obtain the correct solution in a simplistic case
"""
rnd = np.random.RandomState(0)
mask = np.ones([10, 10], dtype=np.bool)
X = rnd.randn(50, 100)
connectivity = grid_to_graph(*mask.shape)
assert_warns(DeprecationWarning, WardAgglomeration)
with ignore_warnings():
ward = WardAgglomeration(n_clusters=5, connectivity=connectivity)
ward.fit(X)
agglo = FeatureAgglomeration(n_clusters=5, connectivity=connectivity)
agglo.fit(X)
assert_array_equal(agglo.labels_, ward.labels_)
assert_true(np.size(np.unique(agglo.labels_)) == 5)
X_red = agglo.transform(X)
assert_true(X_red.shape[1] == 5)
X_full = agglo.inverse_transform(X_red)
assert_true(np.unique(X_full[0]).size == 5)
assert_array_almost_equal(agglo.transform(X_full), X_red)
# Check that fitting with no samples raises a ValueError
assert_raises(ValueError, agglo.fit, X[:0])
def select_train_predict(X, Y, Z, feature_list, selection_method, estimator_method, n_features, selection_args, estimator_args):
W = []
features = []
if selection_method != '2step_kbest':
n_features = min(n_features, len(feature_list))
if estimator_method == 'svm' and selection_method == 'rfe':
estimator_args['kernel'] = 'linear'
estimator = ESTIMATORS[estimator_method](**estimator_args)
if selection_method == 'cluster':
agglom = FeatureAgglomeration(n_clusters=n_features, affinity='cosine', linkage='average')
clusters = agglom.fit_predict(X).tolist()
sample = [clusters.index(i) for i in range(n_features)]
X = X[:,sample]
Z = Z[:,sample]
selection_method = None
if selection_method is None:
for i, y in enumerate(Y):
estimator.fit(X, y)
w = estimator.predict(Z)
W.append(w)
if (i+1) % (len(Y) / 10) == 0:
print '.',
if selection_method == 'rfe':
selector = RFE(estimator=estimator, n_features_to_select=n_features, **selection_args)
for i, y in enumerate(Y):
selector = selector.fit(X, y)
features.append(feature_list[selector.support_])
w = selector.predict(Z)
W.append(w)
if (i+1) % (len(Y) / 10) == 0:
print '.',
if selection_method == 'myrfe':
selector = MyRFE(estimator=estimator, n_features=n_features, **selection_args)
for i, y in enumerate(Y):
selector.fit(X, y)
features.append(feature_list[selector.support])
w = selector.predict(Z)
W.append(w)
if (i+1) % (len(Y) / 10) == 0:
print '.',
if selection_method == 'kbest':
selector = SelectKBest(f_regression, k=n_features, **selection_args)
for i, y in enumerate(Y):
X2 = selector.fit_transform(X, y)
Z2 = selector.transform(Z)
features.append(feature_list[selector.get_support()])
estimator.fit(X2, y)
w = estimator.predict(Z2)
W.append(w)
if (i+1) % (len(Y) / 10) == 0:
print '.',
print
return W, features
def __init__(self):
self.clf = Pipeline([
("RF", RandomForestRegressor(n_estimators=200, max_depth=15,
n_jobs=N_JOBS))])
self.scaler = StandardScaler()
self.agglo = FeatureAgglomeration(n_clusters=500)
# Transpose and scale parameters featListOriginal[[0,1,5,7,9,10],:] = featListOriginal[[0,1,5,7,9,10],:]*0.8 featListOriginal[[2,3,4],:] = featListOriginal[[2,3,4],:]*0.8*0.2 featListOriginal[6,:] = featListOriginal[6,:]*0.8*0.8 featListOriginal = NP.transpose(featListOriginal) ## STANDARDIZE FEATURES ################################################### # Don't standardize the centroids for k in range(2,numFeat): featList[k] = (featList[k] - NP.mean(featList[k]))/NP.sqrt(NP.var(featList[k])) # Transpose the feature list to use in clustering featList = NP.transpose(featList) feat_aggl = FeatureAgglomeration(2) feat_aggl.fit(featList[:,2:]) ## AGGLOMERATIVE CLUSTERING ############################################### aggl_all = AgglomerativeClustering(2) X_All = featList[:,2:] y2 = aggl_all.fit_predict(X_All) ## PCA ############################################################### pca_model = PCA(2) X_PCA = pca_model.fit_transform(X_All) print(pca_model.explained_variance_ratio_) ## SPLIT INTO numBINS ###############################################
def FeatureAgglomeration(self, clist, numClusters=2): FEATAGGL = FeatureAgglomeration(numClusters) FEATAGGL.fit(self.featList[:,clist]) self.featureTree = FEATAGGL.children_ self.featureLabels = FEATAGGL.labels_ self.featureCList = clist
from sklearn.naive_bayes import BernoulliNB
from sklearn.cluster import FeatureAgglomeration
from sklearn.metrics import roc_curve, auc
from sklearn.cross_validation import KFold
from sklearn import metrics
###############################################################################
# Data IO and generation
# import some data to play with
#file = "/home/kbhalla/Desktop/Data/day_samp_new.npy"
file = "/home/rmendoza/Documents/Data/day_samp_new_0604.npy"
with open(file, "r") as file_in:
matrix = smio.load_sparse_csr(file_in)
X = matrix[:,:-1]
FA = FeatureAgglomeration(n_clusters=250)
print np.shape(X)
y = matrix[:,-1]
X = FA.fit_transform(X,y)
n_samples, n_features = X.shape
k = int(0.8*n_samples)
#random_state = np.random.RandomState(0)
#X = np.c_[X, random_state.randn(n_samples, 2*n_features)]
X_test, y_test = X[k:,:], y[k:]
X, y = X[:k, :], y[:k]
sm = SMOTE(ratio=0.95)
X,y = sm.fit_sample(X, y)
print np.shape(X)
start = time.time()
##################################################################
# Then we use FeatureAgglomeration from scikit-learn. Indeed, the voxels
# are the features of the data matrix.
#
# In addition, we use caching. As a result, the clustering doesn't have
# to be recomputed later.
# Computing the ward for the first time, this is long...
from sklearn.cluster import FeatureAgglomeration
# If you have scikit-learn older than 0.14, you need to import
# WardAgglomeration instead of FeatureAgglomeration
import time
start = time.time()
ward = FeatureAgglomeration(n_clusters=1000, connectivity=connectivity,
linkage='ward', memory='nilearn_cache')
ward.fit(fmri_masked)
print("Ward agglomeration 1000 clusters: %.2fs" % (time.time() - start))
# Compute the ward with more clusters, should be faster as we are using
# the caching mechanism
start = time.time()
ward = FeatureAgglomeration(n_clusters=2000, connectivity=connectivity,
linkage='ward', memory='nilearn_cache')
ward.fit(fmri_masked)
print("Ward agglomeration 2000 clusters: %.2fs" % (time.time() - start))
##################################################################
# Visualize results
# ------------------
#
def find_cluster(childrens, start, height, sample_size):
res = start
for i in range(height - 1):
res = find_feature(childrens, res + sample_size)
cluster = rec_cluster(childrens, childrens[res], sample_size)
cluster.sort()
return cluster
def find_feature_cluster(children, feature, height, sample_size):
return find_cluster(children, find_feature(children, feature), height, sample_size)
BENCH_DATA = genfromtxt('Sequential_Application_SATUNSAT_track_wo_names.csv', delimiter=',')
#BENCH_DATA = BENCH_DATA.transpose()
print(BENCH_DATA.shape)
#print(np.isnan(BENCH_DATA).any())
ward = FeatureAgglomeration(linkage='average')
#print(ward.fit_predict(BENCH_DATA))
ward.fit(BENCH_DATA)
#print(ward.children_)
#print(find_feature_cluster(ward.children_, 0, 2, 300))
plt.title('SAT Feature_Agglomeration')
plot_dendrogram(ward, leaf_font_size = 12)
#plt.savefig('SAT_Feature_Agglomeration.png')
plt.show()
