def train(self, sparseX, **kargv):
print 'Info: Training KNN'
if 'algorithm' not in kargv:
self.model = NearestNeighbors(kargv['topK'], metric=kargv['metric'])
else:
self.model = NearestNeighbors(kargv['topK'], metric=kargv['metric'], algorithm=kargv['algorithm'])
self.model.fit(sparseX)
def image_retrieval(constant_overwrites):
img_shape, attr, x_train, x_test = load_faces_dataset()
constants = merge_dict(get_constants(), constant_overwrites)
constants['img_shape'] = img_shape
encoder_filename = constants['encoder_filename']
decoder_filename = constants['decoder_filename']
reset_tf_session()
autoencoder, encoder, decoder = model_builder(network_builder, constants)
if os.path.exists(encoder_filename) and not constants['retrain']:
encoder.load_weights(encoder_filename)
else:
data = {'X_train': x_train, 'X_test': x_test}
train(autoencoder, data, constants)
encoder.save_weights(encoder_filename)
decoder.save_weights(decoder_filename)
images = x_train
codes = encoder.predict(images)
assert len(codes) == len(images)
nei_clf = NearestNeighbors(metric="euclidean")
nei_clf.fit(codes)
# Cherry-picked examples:
# smiles
show_similar(x_test[247], nei_clf, encoder, images)
# ethnicity
show_similar(x_test[56], nei_clf, encoder, images)
# glasses
show_similar(x_test[63], nei_clf, encoder, images)
class FriendTrainer(Trainer):
'''
@summary: 计算与用户相似的用户，为基于用户的协同过滤算法做准备
'''
def __init__(self, num=5, provider=None):
super().__init__(num, provider)
self.num = num
self.provider = provider if provider else UserInterestProvider()
self.model = None
def config(self, num=None, category=None):
self.model = None
self.num = num if num else self.num
self.provider = ProviderFactory.getProvider(
featureCategory=category) if category else self.provider
def train(self, userId):
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.num + 1).fit(
self.provider.provideAll())
distance, neighborList = self.model.kneighbors(
[self.provider.provide(userId)])
if distance[0][2] == 0:
return []
similarity = self.distanceToSimilarity(distance[0][1:])
res = []
for i in range(self.num):
res.append((neighborList[i], similarity[i]))
return res
def trainAll(self):
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.num + 1,
algorithm='auto').fit(
self.provider.provideAll())
res = []
distances, friends = self.model.kneighbors(self.provider.provideAll())
for count in range(len(friends)):
friend = []
if distances[count][2] == 0:
res.append(friend)
continue
similarity = self.distanceToSimilarity(distances[count])[1:]
neighborList = friends[count][1:]
for i in range(self.num):
friend.append((neighborList[i], similarity[i]))
res.append(friend)
print("User " + str(count) + " finded!")
if self.isUpdate():
# DBUtil.dumpFriends(res)
CacheUtil.dumpUserFriends(res)
DBUtil.dumpFriends(res)
return res
def clear(self):
del self.model
def distanceToSimilarity(self, distance):
return list(map(lambda x: 1 - x, distance))
def fit(self, X, y, sample_weight=None):
""" Prepare different things for fast computation of metrics """
X, y, sample_weight = check_xyw(X, y, sample_weight=sample_weight)
self._label_mask = numpy.array(y == self.uniform_label)
assert sum(self._label_mask) > 0, 'No events of uniform class!'
# weights of events
self._masked_weight = sample_weight[self._label_mask]
X_part = numpy.array(take_features(
X, self.uniform_features))[self._label_mask, :]
# computing knn indices
neighbours = NearestNeighbors(n_neighbors=self.n_neighbours,
algorithm='kd_tree').fit(X_part)
_, self._groups_indices = neighbours.kneighbors(X_part)
self._group_matrix = ut.group_indices_to_groups_matrix(
self._groups_indices, n_events=len(X_part))
# self._group_weights = ut.compute_group_weights_by_indices(self._groups_indices,
# sample_weight=self._masked_weight)
self._group_weights = ut.compute_group_weights(
self._group_matrix, sample_weight=self._masked_weight)
# self._divided_weights = ut.compute_divided_weight_by_indices(self._groups_indices,
# sample_weight=self._masked_weight)
self._divided_weights = ut.compute_divided_weight(
self._group_matrix, sample_weight=self._masked_weight)
return self
def _get_kernel(self, X, y=None):
if self.kernel == "rbf":
if y is None:
return rbf_kernel(X, X, gamma=self.gamma)
else:
return rbf_kernel(X, y, gamma=self.gamma)
elif self.kernel == "knn":
if self.nn_fit is None:
t0 = time()
self.nn_fit = NearestNeighbors(self.n_neighbors,
n_jobs=self.n_jobs).fit(X)
print("NearestNeighbors fit time cost:", time() - t0)
if y is None:
t0 = time()
result = self.nn_fit.kneighbors_graph(self.nn_fit._fit_X,
self.n_neighbors,
mode='connectivity')
print("construct kNN graph time cost:", time() - t0)
return result
else:
return self.nn_fit.kneighbors(y, return_distance=False)
elif callable(self.kernel):
if y is None:
return self.kernel(X, X)
else:
return self.kernel(X, y)
else:
raise ValueError("%s is not a valid kernel. Only rbf and knn"
" or an explicit function "
" are supported at this time." % self.kernel)
def fit(self, X, y):
t = time() # get labels for test data
# build the graph result is the affinity matrix
if self.kernel is 'dbscan' or self.kernel is None:
affinity_matrix = self.dbscan(X, self.eps, self.minPts)
# it is possible to use other kernels -> as parameter
elif self.kernel is 'rbf':
affinity_matrix = rbf_kernel(X, X, gamma=self.gamma)
elif self.kernel is 'knn':
affinity_matrix = NearestNeighbors(self.naighbors).fit(X).kneighbors_graph(X, self.naighbors).toarray()
else:
raise
print( "praph(%s) time %2.3fms"%(self.kernel, (time() - t) *1000))
if affinity_matrix.max() == 0 :
print("no affinity matrix found")
return y
degree_martix = np.diag(affinity_matrix.sum(axis=0))
affinity_matrix = np.matrix(affinity_matrix)
try:
inserve_degree_matrix = np.linalg.inv(degree_martix)
except np.linalg.linalg.LinAlgError as err:
if 'Singular matrix' in err.args:
# use a pseudo inverse if it's not possible to make a normal of the degree matrix
inserve_degree_matrix = np.linalg.pinv(degree_martix)
else:
raise
matrix = inserve_degree_matrix * affinity_matrix
# split labels in different vectors to calculate the propagation for the separate label
labels = np.unique(y)
labels = [x for x in labels if x != self.unlabeledValue]
# init the yn1 and y0
y0 = [[1 if (x == l) else 0 for x in y] for l in labels]
yn1 = y0
# function to set the probability to 1 if it was labeled in the source
toOrgLabels = np.vectorize(lambda x, y : 1 if y == 1 else x , otypes=[np.int0])
# function to set the index's of the source labeled
toOrgLabelsIndex = np.vectorize(lambda x, y, z : z if y == 1 else x , otypes=[np.int0])
lastLabels = np.argmax(y0, axis=0)
while True:
yn1 = yn1 * matrix
#first matrix to labels
ynLablesIndex = np.argmax(yn1, axis=0)
# row-normalize
yn1 /= yn1.max()
yn1 = toOrgLabels(yn1, y0)
for x in y0:
ynLablesIndex = toOrgLabelsIndex(ynLablesIndex, x, y0.index(x))
#second original labels to result
if np.array_equiv(ynLablesIndex, lastLabels):
break
lastLabels = ynLablesIndex
# result is the index of the labels -> cast index to the given labels
toLabeles = np.vectorize(lambda x : labels[x])
return np.array(toLabeles(lastLabels))[0]
def train(self, sparseX, **kargv):
print 'Info: Training KNN'
if 'algorithm' not in kargv:
self.model = NearestNeighbors(kargv['topK'],
metric=kargv['metric'])
else:
self.model = NearestNeighbors(kargv['topK'],
metric=kargv['metric'],
algorithm=kargv['algorithm'])
self.model.fit(sparseX)
class NearestNeighboor(object):
def __init__(self):
self.model = None
#preprocessing, featuring, training, predicting, parsing, evaluation
@staticmethod
def sparseX(candPartPath, qUserPath, candToId, qPackageToId, matShape):
row = []
col = []
val = []
for file_ in os.listdir(qUserPath):
f = gzip.open(qUserPath + os.sep + file_, mode='rb')
for line in f:
spl = line.strip().split('|')
username = spl[0]
if username not in candToId:
continue
for i in range(1, len(spl)):
sp = spl[i].split(':')
if sp[0] in qPackageToId:
row.append(candToId[username])
col.append(qPackageToId[sp[0]])
val.append(int(sp[1]))
f.close()
f = gzip.open(candPartPath, mode='rb')
for line in f:
spl = line.strip().split('|')
username = spl[0]
for i in range(1, len(spl)):
sp = spl[i].split(':')
row.append(candToId[username])
col.append(qPackageToId[sp[0]])
val.append(int(sp[1]))
f.close()
sparseX = csc_matrix((val, (row, col)), shape=matShape)
return sparseX
def train(self, sparseX, **kargv):
print 'Info: Training KNN'
if 'algorithm' not in kargv:
self.model = NearestNeighbors(kargv['topK'],
metric=kargv['metric'])
else:
self.model = NearestNeighbors(kargv['topK'],
metric=kargv['metric'],
algorithm=kargv['algorithm'])
self.model.fit(sparseX)
def predict(self, sparsePX):
pred = self.model.kneighbors(sparsePX, return_distance=True)
dist = pred[0]
idx = pred[1]
return (dist, idx)
def _entropy(data):
if len(data) == 1:
return 0.0
euler_const = 0.5772156649
nn = NearestNeighbors(n_neighbors=2).fit(data)
entropy = 0.0
for x in data:
nearest_neighbor_distance = nn.kneighbors(x)[0][0].max()
entropy += math.log(len(data) * nearest_neighbor_distance)
entropy = entropy / len(data) + math.log(2) + euler_const
return entropy
def preprocess_neighbors(self, rebuild=False, save=True):
neighbors_model_path = os.path.join(self.selected_dir,
"neighbors_model" + ".pkl")
neighbors_path = os.path.join(self.selected_dir, "neighbors" + ".npy")
neighbors_weight_path = os.path.join(self.selected_dir,
"neighbors_weight" + ".npy")
test_neighbors_path = os.path.join(self.selected_dir,
"test_neighbors" + ".npy")
test_neighbors_weight_path = os.path.join(
self.selected_dir, "test_neighbors_weight" + ".npy")
if os.path.exists(neighbors_model_path) and \
os.path.exists(neighbors_path) and \
os.path.exists(test_neighbors_path) and rebuild == False:
print("neighbors and neighbor_weight exist!!!")
neighbors = np.load(neighbors_path)
neighbors_weight = np.load(neighbors_weight_path)
test_neighbors = np.load(test_neighbors_path)
self.test_neighbors = test_neighbors
return neighbors, neighbors_weight, test_neighbors
print("neighbors and neighbor_weight do not exist, preprocessing!")
train_num = self.train_X.shape[0]
train_y = np.array(self.train_y)
test_num = self.test_X.shape[0]
max_neighbors = min(len(train_y), 200)
print("data shape: {}, labeled_num: {}".format(str(self.train_X.shape),
sum(train_y != -1)))
nn_fit = NearestNeighbors(7, n_jobs=-4).fit(self.train_X)
print("nn construction finished!")
neighbor_result = nn_fit.kneighbors_graph(
nn_fit._fit_X,
max_neighbors,
# 2,
mode="distance")
test_neighbors_result = nn_fit.kneighbors_graph(self.test_X,
max_neighbors,
mode="distance")
print("neighbor_result got!")
neighbors, neighbors_weight = csr_to_impact_matrix(
neighbor_result, train_num, max_neighbors)
test_neighbors, test_neighbors_weight = csr_to_impact_matrix(
test_neighbors_result, test_num, max_neighbors)
self.test_neighbors = test_neighbors
print("preprocessed neighbors got!")
# save neighbors information
if save:
pickle_save_data(neighbors_model_path, nn_fit)
np.save(neighbors_path, neighbors)
np.save(neighbors_weight_path, neighbors_weight)
np.save(test_neighbors_path, test_neighbors)
np.save(test_neighbors_weight_path, test_neighbors_weight)
return neighbors, neighbors_weight, test_neighbors
def fit(self, X, y, sample_weight=None):
""" Prepare different things for fast computation of metrics """
X, y, sample_weight = check_xyw(X, y, sample_weight=sample_weight)
self._mask = numpy.array(y == self.uniform_label)
assert sum(self._mask) > 0, 'No events of uniform class!'
self._masked_weight = sample_weight[self._mask]
X_part = numpy.array(take_features(X, self.uniform_features))[self._mask, :]
# computing knn indices
neighbours = NearestNeighbors(n_neighbors=self.n_neighbours, algorithm='kd_tree').fit(X_part)
_, self._groups_indices = neighbours.kneighbors(X_part)
self._group_weights = ut.compute_group_weights(self._groups_indices, sample_weight=self._masked_weight)
def distance_quality_matrix(X, y, n_neighbors=50):
"""On of the ways to measure the quality of knning: each element
shows how frequently events of class A are met in knn of labels of class B"""
labels = numpy.unique(y)
nn = NearestNeighbors(n_neighbors=n_neighbors)
nn.fit(X)
knn_indices = nn.kneighbors(X, n_neighbors=n_neighbors, return_distance=False)
confusion_matrix = numpy.zeros([len(labels), len(labels)], dtype=int)
for label1, labels2 in zip(y, numpy.take(y, knn_indices)):
for label2 in labels2:
confusion_matrix[label1, label2] += 1
return confusion_matrix
def train(self, userId):
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.num + 1).fit(
self.provider.provideAll())
distance, neighborList = self.model.kneighbors(
[self.provider.provide(userId)])
if distance[0][2] == 0:
return []
similarity = self.distanceToSimilarity(distance[0][1:])
res = []
for i in range(self.num):
res.append((neighborList[i], similarity[i]))
return res
def compute_knn_indices_of_signal(X, is_signal, n_neighbours=50):
"""For each event returns the knn closest signal(!) events. No matter of what class the event is.
:type X: numpy.array, shape = [n_samples, n_features] the distance is measured over these variables
:type is_signal: numpy.array, shape = [n_samples] with booleans
:rtype numpy.array, shape [len(dataframe), knn], each row contains indices of closest signal events
"""
assert len(X) == len(is_signal), "Different lengths"
signal_indices = numpy.where(is_signal)[0]
X_signal = numpy.array(X)[numpy.array(is_signal)]
neighbours = NearestNeighbors(n_neighbors=n_neighbours, algorithm='kd_tree').fit(X_signal)
_, knn_signal_indices = neighbours.kneighbors(X)
return numpy.take(signal_indices, knn_signal_indices)
class NearestNeighboor(object):
def __init__(self):
self.model = None
#preprocessing, featuring, training, predicting, parsing, evaluation
@staticmethod
def sparseX(candPartPath, qUserPath, candToId, qPackageToId, matShape):
row = []
col = []
val = []
for file_ in os.listdir(qUserPath):
f = gzip.open(qUserPath+os.sep+file_, mode='rb')
for line in f:
spl = line.strip().split('|')
username = spl[0]
if username not in candToId:
continue
for i in range(1, len(spl)):
sp = spl[i].split(':')
if sp[0] in qPackageToId:
row.append(candToId[username])
col.append(qPackageToId[sp[0]])
val.append(int(sp[1]))
f.close()
f = gzip.open(candPartPath, mode='rb')
for line in f:
spl = line.strip().split('|')
username = spl[0]
for i in range(1, len(spl)):
sp = spl[i].split(':')
row.append(candToId[username])
col.append(qPackageToId[sp[0]])
val.append(int(sp[1]))
f.close()
sparseX = csc_matrix((val, (row, col)), shape=matShape)
return sparseX
def train(self, sparseX, **kargv):
print 'Info: Training KNN'
if 'algorithm' not in kargv:
self.model = NearestNeighbors(kargv['topK'], metric=kargv['metric'])
else:
self.model = NearestNeighbors(kargv['topK'], metric=kargv['metric'], algorithm=kargv['algorithm'])
self.model.fit(sparseX)
def predict(self, sparsePX):
pred = self.model.kneighbors(sparsePX, return_distance = True)
dist = pred[0]
idx = pred[1]
return (dist, idx)
def compute_knn_indices_of_signal(X, is_signal, n_neighbours=50):
"""For each event returns the knn closest signal(!) events. No matter of what class the event is.
:type X: numpy.array, shape = [n_samples, n_features] the distance is measured over these variables
:type is_signal: numpy.array, shape = [n_samples] with booleans
:rtype numpy.array, shape [len(dataframe), knn], each row contains indices of closest signal events
"""
assert len(X) == len(is_signal), "Different lengths"
signal_indices = numpy.where(is_signal)[0]
X_signal = numpy.array(X)[numpy.array(is_signal)]
neighbours = NearestNeighbors(n_neighbors=n_neighbours,
algorithm='kd_tree').fit(X_signal)
_, knn_signal_indices = neighbours.kneighbors(X)
return numpy.take(signal_indices, knn_signal_indices)
def rvalue(X, Y, n_neighbors=10, theta=1):
neigh = NearestNeighbors(n_neighbors=n_neighbors).fit(X)
sum = 0
for i in range(len(X)):
_, [indices] = neigh.kneighbors([X[i]])
diff = [Y[index] for index in indices if Y[index] != Y[i]]
if len(diff) > theta:
sum += 1
return sum / len(X)
def computeSignalKnnIndices(uniform_variables, dataframe, is_signal, n_neighbors=50):
"""For each event returns the knn closest signal(!) events. No matter of what class the event is.
:type uniform_variables: list of names of variables, using which we want to compute the distance
:type dataframe: pandas.DataFrame, should contain these variables
:type is_signal: numpy.array, shape = [n_samples] with booleans
:rtype numpy.array, shape [len(dataframe), knn], each row contains indices of closest signal events
"""
assert len(dataframe) == len(is_signal), "Different lengths"
signal_indices = numpy.where(is_signal)[0]
for variable in uniform_variables:
assert variable in dataframe.columns, "Dataframe is missing %s column" % variable
uniforming_features_of_signal = numpy.array(dataframe.ix[is_signal, uniform_variables])
neighbours = NearestNeighbors(n_neighbors=n_neighbors, algorithm='kd_tree').fit(uniforming_features_of_signal)
_, knn_signal_indices = neighbours.kneighbors(dataframe[uniform_variables])
return numpy.take(signal_indices, knn_signal_indices)
def _get_kernel(self, X, y=None):
if self.kernel == "rbf":
if y is None:
return rbf_kernel(X, X, gamma=self.gamma)
else:
return rbf_kernel(X, y, gamma=self.gamma)
elif self.kernel == "knn":
if self.nn_fit is None:
self.nn_fit = NearestNeighbors(self.n_neighbors).fit(X)
if y is None:
# Nearest neighbors returns a directed matrix.
dir_graph = self.nn_fit.kneighbors_graph(self.nn_fit._fit_X,
self.n_neighbors,
mode='connectivity')
# Making the matrix symmetric
un_graph = dir_graph + dir_graph.T
# Since it is a connectivity matrix, all values should be
# either 0 or 1
un_graph[un_graph > 1.0] = 1.0
return un_graph
else:
return self.nn_fit.kneighbors(y, return_distance=False)
else:
raise ValueError("%s is not a valid kernel. Only rbf and knn"
" are supported at this time" % self.kernel)
def predictAll(self, userIndex):
if not self.default:
self.coldStart()
uf = self.getParam(userIndex)
if sum(uf) == 0:
return self.default
# data,transfer=self.afProvider.filterClicked()
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.maxNum,
algorithm='auto').fit(
self.afProvider.provideAll())
distance, candidates = self.model.kneighbors([uf])
res = []
for i in range(self.maxNum):
res.append((candidates[0][i], 1 - distance[0][i]))
return res
class SimPredictor(Predictor):
'''
@summary: 通过用户与新闻的主题向量的余弦相似度来得出用户评分
@addition:候选集选取使用聚类,将用户分到某一类,该类中所有的文章成为候选集
'''
def __init__(self, ufProvider=None, afProvider=None):
super().__init__()
self.model = None
self.maxNum = 10
self.default = None
def coldStart(self):
tmp = []
for i in range(Article.objects.count()):
tmp.append((i, len(CacheUtil.loadClickedForArticle(i))))
tmp.sort(key=lambda x: x[1], reverse=True)
total = sum(list(map(lambda x: x[1], tmp[0:self.maxNum])))
self.default = list(
map(lambda x: (x[0], x[1] / total), tmp[0:self.maxNum]))
return self.default
def config(self, maxNum=None, ufProvider=None, afProvider=None):
self.ufProvider = ufProvider if ufProvider else self.ufProvider
self.afProvider = afProvider if afProvider else self.afProvider
self.maxNum = maxNum if maxNum else self.maxNum
def predict(self, userId, articleId):
af = np.array(self.getFeature(articleId))
uf = np.array(self.getParam(userId))
if sum(uf) == 0:
if not self.default:
self.coldStart()
for pair in self.default:
if pair[0] == articleId:
return pair[1]
return 0
return np.dot(af,
uf) / (np.sqrt(np.dot(uf, uf)) * np.sqrt(np.dot(af, af)))
def predictAll(self, userIndex):
if not self.default:
self.coldStart()
uf = self.getParam(userIndex)
if sum(uf) == 0:
return self.default
# data,transfer=self.afProvider.filterClicked()
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.maxNum,
algorithm='auto').fit(
self.afProvider.provideAll())
distance, candidates = self.model.kneighbors([uf])
res = []
for i in range(self.maxNum):
res.append((candidates[0][i], 1 - distance[0][i]))
return res
def clear(self):
del self.model
def kmeanspp(X, k, seed):
# That we need to do this is a bug in _init_centroids
x_squared_norms = row_norms(X, squared=True)
# Use k-means++ to initialise the centroids
centroids = _init_centroids(X, k, 'k-means++', random_state=seed, x_squared_norms=x_squared_norms)
# OK, we should just short-circuit and get these from k-means++...
# quick and dirty solution
nns = NearestNeighbors()
nns.fit(X)
centroid_candidatess = nns.radius_neighbors(X=centroids, radius=0, return_distance=False)
# Account for "degenerated" solutions: serveral voxels at distance 0, each becoming a centroid
centroids = set()
for centroid_candidates in centroid_candidatess:
centroid_candidates = set(centroid_candidates) - centroids
if len(set(centroid_candidates) - centroids) == 0:
raise Exception('Cannot get an unambiguous set of centers;'
'theoretically this cannot happen, so check for bugs')
centroids.add(centroid_candidates.pop())
return np.array(sorted(centroids))
def computeSignalKnnIndices(uniform_variables,
dataframe,
is_signal,
n_neighbors=50):
"""For each event returns the knn closest signal(!) events. No matter of what class the event is.
:type uniform_variables: list of names of variables, using which we want to compute the distance
:type dataframe: pandas.DataFrame, should contain these variables
:type is_signal: numpy.array, shape = [n_samples] with booleans
:rtype numpy.array, shape [len(dataframe), knn], each row contains indices of closest signal events
"""
assert len(dataframe) == len(is_signal), "Different lengths"
signal_indices = numpy.where(is_signal)[0]
for variable in uniform_variables:
assert variable in dataframe.columns, "Dataframe is missing %s column" % variable
uniforming_features_of_signal = numpy.array(
dataframe.ix[is_signal, uniform_variables])
neighbours = NearestNeighbors(
n_neighbors=n_neighbors,
algorithm='kd_tree').fit(uniforming_features_of_signal)
_, knn_signal_indices = neighbours.kneighbors(dataframe[uniform_variables])
return numpy.take(signal_indices, knn_signal_indices)
def recommend():
#if request.method == 'POST':
f = request.files['file']
basepath = os.path.dirname(__file__)
file_path = os.path.join(basepath, 'uploads', secure_filename(f.filename))
f.save(file_path)
#custer for recommending
filelist.sort()
featurelist = []
for i, imagepath in enumerate(filelist):
print(" Status: %s / %s" % (i, len(filelist)), end="\r")
img = image.load_img(imagepath, target_size=(224, 224))
img_data = image.img_to_array(img)
img_data = np.expand_dims(img_data, axis=0)
img_data = preprocess_input(img_data)
features = np.array(model.predict(img_data))
featurelist.append(features.flatten())
nei_clf = NearestNeighbors(metric="euclidean")
nei_clf.fit(featurelist)
distances, neighbors = get_similar(file_path, n_neighbors=3)
return 'hello recommender'
def adaptive_evaluation_bkp(self):
train_X = self.data.get_train_X()
affinity_matrix = self.data.get_graph()
affinity_matrix.setdiag(0)
pred = self.pred_dist
test_X = self.data.get_test_X()
test_y = self.data.get_test_ground_truth()
# nn_fit = self.data.get_neighbors_model()
nn_fit = NearestNeighbors(n_jobs=-4).fit(train_X)
logger.info("nn construction finished!")
neighbor_result = nn_fit.kneighbors_graph(test_X,
100,
mode="distance")
logger.info("neighbor_result got!")
estimate_k = 5
s = 0
rest_idxs = self.data.get_rest_idxs()
# removed_idxs = self.remv
labels = []
for i in tqdm(range(test_X.shape[0])):
start = neighbor_result.indptr[i]
end = neighbor_result.indptr[i + 1]
j_in_this_row = neighbor_result.indices[start:end]
data_in_this_row = neighbor_result.data[start:end]
sorted_idx = data_in_this_row.argsort()
assert (len(sorted_idx) == 100)
j_in_this_row = j_in_this_row[sorted_idx]
estimated_idxs = j_in_this_row[:estimate_k]
estimated_idxs = np.array([i for i in estimated_idxs if i in rest_idxs])
adaptive_k = affinity_matrix[estimated_idxs, :].sum() / estimate_k
selected_idxs = j_in_this_row[:int(adaptive_k)]
p = pred[selected_idxs].sum(axis=0)
labels.append(p.argmax())
s += adaptive_k
# print(adaptive_k)
acc = accuracy_score(test_y, labels)
logger.info("exp accuracy: {}".format(acc))
print(s/test_X.shape[0])
def trainAll(self):
if not self.model:
self.model = NearestNeighbors(n_neighbors=self.num + 1,
algorithm='auto').fit(
self.provider.provideAll())
res = []
distances, friends = self.model.kneighbors(self.provider.provideAll())
for count in range(len(friends)):
friend = []
if distances[count][2] == 0:
res.append(friend)
continue
similarity = self.distanceToSimilarity(distances[count])[1:]
neighborList = friends[count][1:]
for i in range(self.num):
friend.append((neighborList[i], similarity[i]))
res.append(friend)
print("User " + str(count) + " finded!")
if self.isUpdate():
# DBUtil.dumpFriends(res)
CacheUtil.dumpUserFriends(res)
DBUtil.dumpFriends(res)
return res
def compute_parameters(self, trainX, trainY):
for variable in self.uniform_variables:
if variable not in trainX.columns:
raise ValueError("Dataframe is missing %s column" % variable)
if self.knn is None:
A = pairwise_distances(trainX[self.uniform_variables])
A = self.distance_dependence(A)
A *= (trainY[:, numpy.newaxis] == trainY[numpy.newaxis, :])
else:
is_signal = trainY > 0.5
# computing knn indices of same type
uniforming_features_of_signal = numpy.array(trainX.ix[is_signal, self.uniform_variables])
neighbours = NearestNeighbors(n_neighbors=self.knn, algorithm='kd_tree').fit(uniforming_features_of_signal)
signal_distances, knn_signal_indices = neighbours.kneighbors(uniforming_features_of_signal)
knn_signal_indices = numpy.where(is_signal)[0].take(knn_signal_indices)
uniforming_features_of_bg = numpy.array(trainX.ix[~is_signal, self.uniform_variables])
neighbours = NearestNeighbors(n_neighbors=self.knn, algorithm='kd_tree').fit(uniforming_features_of_bg)
bg_distances, knn_bg_indices = neighbours.kneighbors(uniforming_features_of_bg)
knn_bg_indices = numpy.where(~is_signal)[0].take(knn_bg_indices)
signal_distances = self.distance_dependence(signal_distances.flatten())
bg_distances = self.distance_dependence(bg_distances.flatten())
signal_ind_ptr = numpy.arange(0, sum(is_signal) * self.knn + 1, self.knn)
bg_ind_ptr = numpy.arange(0, sum(~is_signal) * self.knn + 1, self.knn)
signal_column_indices = knn_signal_indices.flatten()
bg_column_indices = knn_bg_indices.flatten()
A_sig = sparse.csr_matrix(sparse.csr_matrix((signal_distances, signal_column_indices, signal_ind_ptr),
shape=(sum(is_signal), len(trainX))))
A_bg = sparse.csr_matrix(sparse.csr_matrix((bg_distances, bg_column_indices, bg_ind_ptr),
shape=(sum(~is_signal), len(trainX))))
A = sparse.vstack((A_sig, A_bg), format='csr')
if self.row_normalize:
from sklearn.preprocessing import normalize
A = normalize(A, norm='l1', axis=1)
return A, numpy.ones(A.shape[0])
def upload():
if request.method == 'POST':
# Get the file from post request
f = request.files['file']
# Save the file to ./uploads
basepath = os.path.dirname(__file__)
file_path = os.path.join(basepath, 'uploads',
secure_filename(f.filename))
f.save(file_path)
# Make prediction
output_class = [
"batteries", "cloth", "e-waste", "glass", "light bulbs",
"metallic", "organic", "paper", "plastic"
]
preds = model_predict(file_path, model)
print(preds)
pred_class = output_class[np.argmax(preds)]
pred_class_percent = round(np.max(preds) * 100, 2)
result = 'It is ' + pred_class + ' waste' # Convert to string
pred_class = ' with ' + str(pred_class_percent) + '% confidence'
#k-nn for recommending
filelist.sort()
featurelist = []
for i, imagepath in enumerate(filelist):
print(" Status: %s / %s" % (i, len(filelist)), end="\r")
img = image.load_img(imagepath, target_size=(224, 224))
img_data = image.img_to_array(img)
img_data = np.expand_dims(img_data, axis=0)
img_data = preprocess_input(img_data)
features = np.array(model.predict(img_data))
featurelist.append(features.flatten())
nei_clf = NearestNeighbors(metric="euclidean")
nei_clf.fit(featurelist)
code = model_predict(file_path, model)
(distances, ), (idx, ) = nei_clf.kneighbors(code, n_neighbors=3)
#all images are loaded as np arrays
images = []
labels = []
j = 1
for i, image_path in enumerate(filelist):
images.append(load_data(image_path))
images = np.asarray(
images
) # all of the images are converted to np array of (1360,224,224,3)
print(distances, images[idx])
print(images[idx].shape)
final_result = result + pred_class
image_save = Image.fromarray(
(np.array(images[0]) * 255).astype(np.uint8))
#image_save = Image.fromarray(images[idx], "RGB")
image_save.save('out.jpg')
image_output = os.path.join(basepath, 'out.jpg')
immg = '<img src="' + image_output + '" style="height: 132px; width: 132px;">'
#return render_template('index.html', filename=image_output)
return final_result
return None
'MLPRegressor':MLPRegressor(), 'MaxAbsScaler':MaxAbsScaler(), 'MeanShift':MeanShift(), 'MinCovDet':MinCovDet(), 'MinMaxScaler':MinMaxScaler(), 'MiniBatchDictionaryLearning':MiniBatchDictionaryLearning(), 'MiniBatchKMeans':MiniBatchKMeans(), 'MiniBatchSparsePCA':MiniBatchSparsePCA(), 'MultiTaskElasticNet':MultiTaskElasticNet(), 'MultiTaskElasticNetCV':MultiTaskElasticNetCV(), 'MultiTaskLasso':MultiTaskLasso(), 'MultiTaskLassoCV':MultiTaskLassoCV(), 'MultinomialNB':MultinomialNB(), 'NMF':NMF(), 'NearestCentroid':NearestCentroid(), 'NearestNeighbors':NearestNeighbors(), 'Normalizer':Normalizer(), 'NuSVC':NuSVC(), 'NuSVR':NuSVR(), 'Nystroem':Nystroem(), 'OAS':OAS(), 'OneClassSVM':OneClassSVM(), 'OrthogonalMatchingPursuit':OrthogonalMatchingPursuit(), 'OrthogonalMatchingPursuitCV':OrthogonalMatchingPursuitCV(), 'PCA':PCA(), 'PLSCanonical':PLSCanonical(), 'PLSRegression':PLSRegression(), 'PLSSVD':PLSSVD(), 'PassiveAggressiveClassifier':PassiveAggressiveClassifier(), 'PassiveAggressiveRegressor':PassiveAggressiveRegressor(), 'Perceptron':Perceptron(),
# coding:utf-8
'''
Created on 2020年1月11日
@author: root
'''
from sklearn.neighbors.unsupervised import NearestNeighbors
import numpy as np
from com.msb.knn.KNNDateOnHand import *
datingDataMat, datingLabels = file2matrix('../../../data/datingTestSet2.txt')
normMat, ranges, minVals = autoNorm(datingDataMat)
nbrs = NearestNeighbors(n_neighbors=10).fit(normMat)
input_man = [[50000, 8, 9.5]]
S = (input_man - minVals) / ranges
distances, indices = nbrs.kneighbors(S)
# classCount K：类别名 V：这个类别中的样本出现的次数
classCount = {}
for i in range(10):
voteLabel = datingLabels[indices[0][i]]
classCount[voteLabel] = classCount.get(voteLabel, 0) + 1
sortedClassCount = sorted(classCount.items(), key=operator.itemgetter(1), reverse=True)
resultList = ['没感觉', '看起来还行', '极具魅力']
print(resultList[sortedClassCount[0][0] - 1])
<img src="https://github.com/hse-aml/intro-to-dl/blob/master/week4/images/similar_images.jpg?raw=1" style="width:60%">
To speed up retrieval process, one should use Locality Sensitive Hashing on top of encoded vectors. This [technique](https://erikbern.com/2015/07/04/benchmark-of-approximate-nearest-neighbor-libraries.html) can narrow down the potential nearest neighbours of our image in latent space (encoder code). We will caclulate nearest neighbours in brute force way for simplicity.
"""
# restore trained encoder weights
s = reset_tf_session()
encoder, decoder = build_deep_autoencoder(IMG_SHAPE, code_size=32)
encoder.load_weights("encoder.h5")
images = X_train
codes = ### YOUR CODE HERE: encode all images ###
assert len(codes) == len(images)
from sklearn.neighbors.unsupervised import NearestNeighbors
nei_clf = NearestNeighbors(metric="euclidean")
nei_clf.fit(codes)
def get_similar(image, n_neighbors=5):
assert image.ndim==3,"image must be [batch,height,width,3]"
code = encoder.predict(image[None])
(distances,),(idx,) = nei_clf.kneighbors(code,n_neighbors=n_neighbors)
return distances,images[idx]
def show_similar(image):
distances,neighbors = get_similar(image,n_neighbors=3)
class BaseLabelPropagation(BaseEstimator, ClassifierMixin, metaclass=ABCMeta):
"""Base class for label propagation module.
Parameters
----------
kernel : {'knn', 'rbf', callable}
String identifier for kernel function to use or the kernel function
itself. Only 'rbf' and 'knn' strings are valid inputs. The function
passed should take two inputs, each of shape [n_samples, n_features],
and return a [n_samples, n_samples] shaped weight matrix
gamma : float
Parameter for rbf kernel
n_neighbors : integer > 0
Parameter for knn kernel
alpha : float
Clamping factor
max_iter : integer
Change maximum number of iterations allowed
tol : float
Convergence tolerance: threshold to consider the system at steady
state
n_jobs : int or None, optional (default=None)
The number of parallel jobs to run.
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
for more details.
"""
def __init__(self,
kernel='rbf',
gamma=20,
n_neighbors=7,
alpha=1,
max_iter=30,
tol=1e-3,
n_jobs=None):
self.max_iter = max_iter
self.tol = tol
# kernel parameters
self.kernel = kernel
self.gamma = gamma
self.n_neighbors = n_neighbors
# clamping factor
self.alpha = alpha
self.n_jobs = n_jobs
self.graph_matrix = None
def _get_kernel(self, X, y=None):
if self.kernel == "rbf":
if y is None:
return rbf_kernel(X, X, gamma=self.gamma)
else:
return rbf_kernel(X, y, gamma=self.gamma)
elif self.kernel == "knn":
if self.nn_fit is None:
t0 = time()
self.nn_fit = NearestNeighbors(self.n_neighbors,
n_jobs=self.n_jobs).fit(X)
print("NearestNeighbors fit time cost:", time() - t0)
if y is None:
t0 = time()
result = self.nn_fit.kneighbors_graph(self.nn_fit._fit_X,
self.n_neighbors,
mode='connectivity')
print("construct kNN graph time cost:", time() - t0)
return result
else:
return self.nn_fit.kneighbors(y, return_distance=False)
elif callable(self.kernel):
if y is None:
return self.kernel(X, X)
else:
return self.kernel(X, y)
else:
raise ValueError("%s is not a valid kernel. Only rbf and knn"
" or an explicit function "
" are supported at this time." % self.kernel)
@abstractmethod
def _build_graph(self):
raise NotImplementedError("Graph construction must be implemented"
" to fit a label propagation model.")
def predict(self, X):
"""Performs inductive inference across the model.
Parameters
----------
X : array_like, shape = [n_samples, n_features]
Returns
-------
y : array_like, shape = [n_samples]
Predictions for input data
"""
probas = self.predict_proba(X)
return self.classes_[np.argmax(probas, axis=1)].ravel()
def predict_proba(self, X):
"""Predict probability for each possible outcome.
Compute the probability estimates for each single sample in X
and each possible outcome seen during training (categorical
distribution).
Parameters
----------
X : array_like, shape = [n_samples, n_features]
Returns
-------
probabilities : array, shape = [n_samples, n_classes]
Normalized probability distributions across
class labels
"""
check_is_fitted(self, 'X_')
X_2d = check_array(
X, accept_sparse=['csc', 'csr', 'coo', 'dok', 'bsr', 'lil', 'dia'])
weight_matrices = self._get_kernel(self.X_, X_2d)
if self.kernel == 'knn':
probabilities = np.array([
np.sum(self.label_distributions_[weight_matrix], axis=0)
for weight_matrix in weight_matrices
])
else:
weight_matrices = weight_matrices.T
probabilities = np.dot(weight_matrices, self.label_distributions_)
normalizer = np.atleast_2d(np.sum(probabilities, axis=1)).T
probabilities /= normalizer
return probabilities
def get_graph(self, X, y):
X, y = check_X_y(X, y)
self.X_ = X
check_classification_targets(y)
graph_matrix = self._build_graph()
return graph_matrix
def fit(self, X, y):
"""Fit a semi-supervised label propagation model based
All the input data is provided matrix X (labeled and unlabeled)
and corresponding label matrix y with a dedicated marker value for
unlabeled samples.
Parameters
----------
X : array-like, shape = [n_samples, n_features]
A {n_samples by n_samples} size matrix will be created from this
y : array_like, shape = [n_samples]
n_labeled_samples (unlabeled points are marked as -1)
All unlabeled samples will be transductively assigned labels
Returns
-------
self : returns an instance of self.
"""
t0 = time()
X, y = check_X_y(X, y)
self.X_ = X
check_classification_targets(y)
# actual graph construction (implementations should override this)
graph_matrix = self._build_graph()
t1 = time()
# label construction
# construct a categorical distribution for classification only
classes = np.unique(y)
classes = (classes[classes != -1])
self.classes_ = classes
n_samples, n_classes = len(y), len(classes)
alpha = self.alpha
if self._variant == 'spreading' and \
(alpha is None or alpha <= 0.0 or alpha >= 1.0):
raise ValueError('alpha=%s is invalid: it must be inside '
'the open interval (0, 1)' % alpha)
y = np.asarray(y)
unlabeled = y == -1
# initialize distributions
self.label_distributions_ = np.zeros((n_samples, n_classes))
for label in classes:
self.label_distributions_[y == label, classes == label] = 1
y_static = np.copy(self.label_distributions_)
if self._variant == 'propagation':
# LabelPropagation
y_static[unlabeled] = 0
else:
# LabelSpreading
y_static *= 1 - alpha
l_previous = np.zeros((self.X_.shape[0], n_classes))
unlabeled = unlabeled[:, np.newaxis]
if sparse.isspmatrix(graph_matrix):
graph_matrix = graph_matrix.tocsr()
for self.n_iter_ in range(self.max_iter):
if np.abs(self.label_distributions_ - l_previous).sum() < self.tol:
break
l_previous = self.label_distributions_
self.label_distributions_ = safe_sparse_dot(
graph_matrix, self.label_distributions_)
if self._variant == 'propagation':
normalizer = np.sum(self.label_distributions_,
axis=1)[:, np.newaxis]
self.label_distributions_ /= normalizer
self.label_distributions_ = np.where(unlabeled,
self.label_distributions_,
y_static)
else:
# clamp
self.label_distributions_ = np.multiply(
alpha, self.label_distributions_) + y_static
else:
warnings.warn('max_iter=%d was reached without convergence.' %
self.max_iter,
category=ConvergenceWarning)
self.n_iter_ += 1
normalizer = np.sum(self.label_distributions_, axis=1)[:, np.newaxis]
self.label_distributions_ /= normalizer
# set the transduction item
transduction = self.classes_[np.argmax(self.label_distributions_,
axis=1)]
self.transduction_ = transduction.ravel()
t2 = time()
print("building graph time cost: {}, spreading time cost: {}".format(
t1 - t0, t2 - t1))
return self
class DBPlot(BaseEstimator):
"""
Heuristic approach to estimate and visualize high-dimensional decision
boundaries for trained binary classifiers by using black-box optimization
to find regions in which the classifier is maximally uncertain (0.5 prediction
probability). The total number of keypoints representing the decision boundary
will depend on n_connecting_keypoints and n_interpolated_keypoints. Reduce
either or both to reduce runtime.
Parameters
----------
estimator : BaseEstimator instance, optional (default=`KNeighborsClassifier(n_neighbors=10)`).
Classifier for which the decision boundary should be plotted. Can be trained
or untrained (in which case the fit method will train it). Must have
probability estimates enabled (i.e. `estimator.predict_proba` must work).
Make sure it is possible for probability estimates to get close to 0.5
(more specifically, as close as specified by acceptance_threshold) - this usally
requires setting an even number of neighbors, estimators etc.
dimensionality_reduction : BaseEstimator instance, optional (default=`PCA(n_components=2)`).
Dimensionality reduction method to help plot the decision boundary in 2D. Can be trained
or untrained (in which case the fit method will train it). Must have n_components=2.
Must be able to project new points into the 2D space after fitting
(i.e. `dimensionality_reduction.transform` must work).
acceptance_threshold : float, optional (default=0.03)
Maximum allowed deviation from decision boundary (defined as the region
with 0.5 prediction probability) when accepting decision boundary keypoints
n_decision_boundary_keypoints : int, optional (default=60)
Total number of decision boundary keypoints added, including both connecting
and interpolated keypoints.
n_connecting_keypoints : int, optional (default=None)
Number of decision boundary keypoints estimated along lines connecting
instances from two different classes (each such line must cross the
decision boundary at least once). If None (default), it is set to 1/3
of n_decision_boundary_keypoints
n_interpolated_keypoints : int, optional (default=None)
Number of decision boundary keypoints interpolated between connecting
keypoints to increase keypoint density. If None (default), it is set to
2/3 of n_decision_boundary_keypoints
n_generated_testpoints_per_keypoint : int, optional (default=15)
Number of demo points generated around decision boundary keypoints, and
labeled according to the specified classifier, in order to enrich and
validate the decision boundary plot
linear_iteration_budget : int, optional (default=100)
Maximum number of iterations the optimizer is allowed to run for each
keypoint estimation while looking along linear trajectories
hypersphere_iteration_budget : int, optional (default=300)
Maximum number of iterations the optimizer is allowed to run for each
keypoint estimation while looking along hypersphere surfaces
verbose: bool, optional (default=True)
Verbose output
"""
def __init__(self, estimator=KNeighborsClassifier(n_neighbors=10), dimensionality_reduction=PCA(n_components=2), acceptance_threshold=0.03, n_decision_boundary_keypoints=60, n_connecting_keypoints=None, n_interpolated_keypoints=None, n_generated_testpoints_per_keypoint=15, linear_iteration_budget=100, hypersphere_iteration_budget=300, verbose=True):
if acceptance_threshold == 0:
raise Warning(
"A nonzero acceptance threshold is strongly recommended so the optimizer can finish in finite time")
if linear_iteration_budget < 2 or hypersphere_iteration_budget < 2:
raise Exception("Invalid iteration budget")
self.classifier = estimator
self.dimensionality_reduction = dimensionality_reduction
self.acceptance_threshold = acceptance_threshold
if n_decision_boundary_keypoints and n_connecting_keypoints and n_interpolated_keypoints and n_connecting_keypoints + n_interpolated_keypoints != n_decision_boundary_keypoints:
raise Exception(
"n_connecting_keypoints and n_interpolated_keypoints must sum to n_decision_boundary_keypoints (set them to None to use calculated suggestions)")
self.n_connecting_keypoints = n_connecting_keypoints if n_connecting_keypoints != None else n_decision_boundary_keypoints / 3
self.n_interpolated_keypoints = n_interpolated_keypoints if n_interpolated_keypoints != None else n_decision_boundary_keypoints * 2 / 3
self.linear_iteration_budget = linear_iteration_budget
self.n_generated_testpoints_per_keypoint = n_generated_testpoints_per_keypoint
self.hypersphere_iteration_budget = hypersphere_iteration_budget
self.verbose = verbose
self.decision_boundary_points = []
self.decision_boundary_points_2d = []
self.X_testpoints = []
self.y_testpoints = []
self.background = []
self.steps = 3
self.hypersphere_max_retry_budget = 20
self.penalties_enabled = True
self.random_gap_selection = False
def setclassifier(self, estimator=KNeighborsClassifier(n_neighbors=10)):
"""Assign classifier for which decision boundary should be plotted.
Parameters
----------
estimator : BaseEstimator instance, optional (default=KNeighborsClassifier(n_neighbors=10)).
Classifier for which the decision boundary should be plotted. Must have
probability estimates enabled (i.e. estimator.predict_proba must work).
Make sure it is possible for probability estimates to get close to 0.5
(more specifically, as close as specified by acceptance_threshold).
"""
self.classifier = estimator
def fit(self, X, y, training_indices=None):
"""Specify data to be plotted, and fit classifier only if required (the
specified clasifier is only trained if it has not been trained yet).
All the input data is provided in the matrix X, and corresponding
binary labels (values taking 0 or 1) in the vector y
Parameters
----------
X : array-like, shape = [n_samples, n_features]
A {n_samples by n_samples} size matrix containing data
y : array-like, shape = [n_samples]
Labels
training_indices : array-like or float, optional (default=None)
Indices on which the classifier has been trained / should be trained.
If float, it is converted to a random sample with the specified proportion
of the full dataset.
Returns
-------
self : returns an instance of self.
"""
if set(np.array(y, dtype=int).tolist()) != set([0, 1]):
raise Exception(
"Currently only implemented for binary classification. Make sure you pass in two classes (0 and 1)")
if training_indices == None:
train_idx = range(len(y))
elif type(training_indices) == float:
train_idx, test_idx = train_test_split(range(len(y)), test_size=0.5)
else:
train_idx = training_indices
self.X = X
self.y = y
self.train_idx = train_idx
#self.test_idx = np.setdiff1d(np.arange(len(y)), self.train_idx, assume_unique=False)
self.test_idx = list(set(range(len(y))).difference(set(self.train_idx)))
# fit classifier if necessary
try:
self.classifier.predict([X[0]])
except:
self.classifier.fit(X[train_idx, :], y[train_idx])
self.y_pred = self.classifier.predict(self.X)
# fit DR method if necessary
try:
self.dimensionality_reduction.transform([X[0]])
except:
self.dimensionality_reduction.fit(X, y)
try:
self.dimensionality_reduction.transform([X[0]])
except:
raise Exception(
"Please make sure your dimensionality reduction method has an exposed transform() method! If in doubt, use PCA or Isomap")
# transform data
self.X2d = self.dimensionality_reduction.transform(self.X)
self.mean_2d_dist = np.mean(pdist(self.X2d))
self.X2d_xmin, self.X2d_xmax = np.min(self.X2d[:, 0]), np.max(self.X2d[:, 0])
self.X2d_ymin, self.X2d_ymax = np.min(self.X2d[:, 1]), np.max(self.X2d[:, 1])
self.majorityclass = 0 if list(y).count(0) > list(y).count(1) else 1
self.minorityclass = 1 - self.majorityclass
minority_idx, majority_idx = np.where(y == self.minorityclass)[
0], np.where(y == self.majorityclass)[0]
self.Xminor, self.Xmajor = X[minority_idx], X[majority_idx]
self.Xminor2d, self.Xmajor2d = self.X2d[minority_idx], self.X2d[majority_idx]
# set up efficient nearest neighbor models for later use
self.nn_model_2d_majorityclass = NearestNeighbors(n_neighbors=2)
self.nn_model_2d_majorityclass.fit(self.X2d[majority_idx, :])
self.nn_model_2d_minorityclass = NearestNeighbors(n_neighbors=2)
self.nn_model_2d_minorityclass.fit(self.X2d[minority_idx, :])
# step 1. look for decision boundary points between corners of majority &
# minority class distribution
minority_corner_idx, majority_corner_idx = [], []
for extremum1 in [np.min, np.max]:
for extremum2 in [np.min, np.max]:
_, idx = self.nn_model_2d_minorityclass.kneighbors(
[[extremum1(self.Xminor2d[:, 0]), extremum2(self.Xminor2d[:, 1])]])
minority_corner_idx.append(idx[0][0])
_, idx = self.nn_model_2d_majorityclass.kneighbors(
[[extremum1(self.Xmajor2d[:, 0]), extremum2(self.Xmajor2d[:, 1])]])
majority_corner_idx.append(idx[0][0])
# optimize to find new db keypoints between corners
self._linear_decision_boundary_optimization(
minority_corner_idx, majority_corner_idx, all_combinations=True, step=1)
# step 2. look for decision boundary points on lines connecting randomly
# sampled points of majority & minority class
n_samples = int(self.n_connecting_keypoints)
from_idx = list(random.sample(list(np.arange(len(self.Xminor))), n_samples))
to_idx = list(random.sample(list(np.arange(len(self.Xmajor))), n_samples))
# optimize to find new db keypoints between minority and majority class
self._linear_decision_boundary_optimization(
from_idx, to_idx, all_combinations=False, step=2)
if len(self.decision_boundary_points_2d) < 2:
print("Failed to find initial decision boundary. Retrying... If this keeps happening, increasing the acceptance threshold might help. Also, make sure the classifier is able to find a point with 0.5 prediction probability (usually requires an even number of estimators/neighbors/etc).")
return self.fit(X, y, training_indices)
# step 3. look for decision boundary points between already known db
# points that are too distant (search on connecting line first, then on
# surrounding hypersphere surfaces)
edges, gap_distances, gap_probability_scores = self._get_sorted_db_keypoint_distances() # find gaps
self.nn_model_decision_boundary_points = NearestNeighbors(n_neighbors=2)
self.nn_model_decision_boundary_points.fit(self.decision_boundary_points)
i = 0
retries = 0
while i < self.n_interpolated_keypoints:
if self.verbose:
print("Step 3/{}:{}/".format(self.steps, i, self.n_interpolated_keypoints))
if self.random_gap_selection:
# randomly sample from sorted DB keypoint gaps?
gap_idx = np.random.choice(len(gap_probability_scores),
1, p=gap_probability_scores)[0]
else:
# get largest gap
gap_idx = 0
from_point = self.decision_boundary_points[edges[gap_idx][0]]
to_point = self.decision_boundary_points[edges[gap_idx][1]]
# optimize to find new db keypoint along line connecting two db keypoints
# with large gap
db_point = self._find_decision_boundary_along_line(
from_point, to_point, penalize_tangent_distance=self.penalties_enabled)
if self.decision_boundary_distance(db_point) > self.acceptance_threshold:
if self.verbose:
print("No good solution along straight line - trying to find decision boundary on hypersphere surface around known decision boundary point")
# hypersphere radius half the distance between from and to db keypoints
R = euclidean(from_point, to_point) / 2.0
# search around either source or target keypoint, with 0.5 probability,
# hoping to find decision boundary in between
if random.random() > 0.5:
from_point = to_point
# optimize to find new db keypoint on hypersphere surphase around known keypoint
db_point = self._find_decision_boundary_on_hypersphere(from_point, R)
if self.decision_boundary_distance(db_point) <= self.acceptance_threshold:
db_point2d = self.dimensionality_reduction.transform([db_point])[0]
self.decision_boundary_points.append(db_point)
self.decision_boundary_points_2d.append(db_point2d)
i += 1
retries = 0
else:
retries += 1
if retries > self.hypersphere_max_retry_budget:
i += 1
dist = self.decision_boundary_distance(db_point)
msg = "Found point is too distant from decision boundary ({}), but retry budget exceeded ({})"
print(msg.format(dist, self.hypersphere_max_retry_budget))
elif self.verbose:
dist = self.decision_boundary_distance(db_point)
print("Found point is too distant from decision boundary ({}) retrying...".format(dist))
else:
db_point2d = self.dimensionality_reduction.transform([db_point])[0]
self.decision_boundary_points.append(db_point)
self.decision_boundary_points_2d.append(db_point2d)
i += 1
retries = 0
edges, gap_distances, gap_probability_scores = self._get_sorted_db_keypoint_distances() # reload gaps
self.decision_boundary_points = np.array(self.decision_boundary_points)
self.decision_boundary_points_2d = np.array(self.decision_boundary_points_2d)
if self.verbose:
print("Done fitting! Found {} decision boundary keypoints.".format(
len(self.decision_boundary_points)))
return self
def plot(self, plt=None, generate_testpoints=True, generate_background=True, tune_background_model=False, background_resolution=100, scatter_size_scale=1.0, legend=True):
"""Plots the dataset and the identified decision boundary in 2D.
(If you wish to create custom plots, get the data using generate_plot() and plot it manually)
Parameters
----------
plt : matplotlib.pyplot or axis object (default=matplotlib.pyplot)
Object to be plotted on
generate_testpoints : boolean, optional (default=True)
Whether to generate demo points around the estimated decision boundary
as a sanity check
generate_background : boolean, optional (default=True)
Whether to generate faint background plot (using prediction probabilities
of a fitted suppor vector machine, trained on generated demo points)
to aid visualization
tune_background_model : boolean, optional (default=False)
Whether to tune the parameters of the support vector machine generating
the background
background_resolution : int, optional (default=100)
Desired resolution (height and width) of background to be generated
scatter_size_scale : float, optional (default=1.0)
Scaling factor for scatter plot marker size
legend : boolean, optional (default=False)
Whether to display a legend
Returns
-------
plt : The matplotlib.pyplot or axis object which has been passed in, after
plotting the data and decision boundary on it. (plt.show() is NOT called
and will be required)
"""
if plt == None:
plt = mplt
if len(self.X_testpoints) == 0:
self.generate_plot(generate_testpoints=generate_testpoints, generate_background=generate_background,
tune_background_model=tune_background_model, background_resolution=background_resolution)
if generate_background and generate_testpoints:
try:
plt.imshow(np.flipud(self.background), extent=[
self.X2d_xmin, self.X2d_xmax, self.X2d_ymin, self.X2d_ymax], cmap="GnBu", alpha=0.33)
except (Exception, ex):
print("Failed to render image background")
# decision boundary
plt.scatter(self.decision_boundary_points_2d[:, 0], self.decision_boundary_points_2d[
:, 1], 600 * scatter_size_scale, c='c', marker='p')
# generated demo points
if generate_testpoints:
plt.scatter(self.X_testpoints_2d[:, 0], self.X_testpoints_2d[
:, 1], 20 * scatter_size_scale, c=['g' if i else 'b' for i in self.y_testpoints], alpha=0.6)
# training data
plt.scatter(self.X2d[self.train_idx, 0], self.X2d[self.train_idx, 1], 150 * scatter_size_scale,
facecolor=['g' if i else 'b' for i in self.y[self.train_idx]],
edgecolor=['g' if self.y_pred[self.train_idx[i]] == self.y[self.train_idx[i]] == 1
else ('b' if self.y_pred[self.train_idx[i]] == self.y[self.train_idx[i]] == 0 else 'r')
for i in range(len(self.train_idx))], linewidths=5 * scatter_size_scale)
# testing data
plt.scatter(self.X2d[self.test_idx, 0], self.X2d[self.test_idx, 1], 150 * scatter_size_scale,
facecolor=['g' if i else 'b' for i in self.y[self.test_idx]],
edgecolor=['g' if self.y_pred[self.test_idx[i]] == self.y[self.test_idx[i]] == 1
else ('b' if self.y_pred[self.test_idx[i]] == self.y[self.test_idx[i]] == 0 else 'r')
for i in range(len(self.test_idx))], linewidths=5 * scatter_size_scale, marker='s')
# label data points with their indices
for i in range(len(self.X2d)):
plt.text(self.X2d[i, 0] + (self.X2d_xmax - self.X2d_xmin) * 0.5e-2,
self.X2d[i, 1] + (self.X2d_ymax - self.X2d_ymin) * 0.5e-2, str(i), size=8)
if legend:
plt.legend(["Estimated decision boundary keypoints", "Generated demo data around decision boundary",
"Actual data (training set)", "Actual data (demo set)"], loc="lower right", prop={'size': 9})
# decision boundary keypoints, in case not visible in background
plt.scatter(self.decision_boundary_points_2d[:, 0], self.decision_boundary_points_2d[:, 1],
600 * scatter_size_scale, c='c', marker='p', alpha=0.1)
plt.scatter(self.decision_boundary_points_2d[:, 0], self.decision_boundary_points_2d[:, 1],
30 * scatter_size_scale, c='c', marker='p', edgecolor='c', alpha=0.8)
# minimum spanning tree through decision boundary keypoints
D = pdist(self.decision_boundary_points_2d)
edges = minimum_spanning_tree(squareform(D))
for e in edges:
plt.plot([self.decision_boundary_points_2d[e[0], 0], self.decision_boundary_points_2d[e[1], 0]],
[self.decision_boundary_points_2d[e[0], 1],
self.decision_boundary_points_2d[e[1], 1]],
'--c', linewidth=4 * scatter_size_scale)
plt.plot([self.decision_boundary_points_2d[e[0], 0], self.decision_boundary_points_2d[e[1], 0]],
[self.decision_boundary_points_2d[e[0], 1],
self.decision_boundary_points_2d[e[1], 1]],
'--k', linewidth=1)
if len(self.test_idx) == 0:
print("No demo performance calculated, as no testing data was specified")
else:
freq = itemfreq(self.y[self.test_idx]).astype(float)
imbalance = np.round(np.max((freq[0, 1], freq[1, 1])) / len(self.test_idx), 3)
acc_score = np.round(accuracy_score(
self.y[self.test_idx], self.y_pred[self.test_idx]), 3)
f1 = np.round(f1_score(self.y[self.test_idx], self.y_pred[self.test_idx]), 3)
plt.title("Test accuracy: " + str(acc_score) + ", F1 score: " +
str(f1) + ". Imbalance (max chance accuracy): " + str(imbalance))
if self.verbose:
print("Plot successfully generated! Don't forget to call the show() method to display it")
return plt
def generate_plot(self, generate_testpoints=True, generate_background=True, tune_background_model=False, background_resolution=100):
"""Generates and returns arrays for visualizing the dataset and the
identified decision boundary in 2D.
Parameters
----------
generate_testpoints : boolean, optional (default=True)
Whether to generate demo points around the estimated decision boundary
as a sanity check
generate_background : boolean, optional (default=True)
Whether to generate faint background plot (using prediction probabilities
of a fitted suppor vector machine, trained on generated demo points)
to aid visualization
tune_background_model : boolean, optional (default=False)
Whether to tune the parameters of the support vector machine generating
the background
background_resolution : int, optional (default=100)
Desired resolution (height and width) of background to be generated
Returns
-------
decision_boundary_points_2d : array
Array containing points in the dimensionality-reduced 2D space which
are very close to the true decision boundary
X_testpoints_2d : array
Array containing generated demo points in the dimensionality-reduced
2D space which surround the decision boundary and can be used for
visual feedback to estimate which area would be assigned which class
y_testpoints : array
Classifier predictions for each of the generated demo points
background: array
Generated background image showing prediction probabilities of the
classifier in each region (only returned if generate_background is set
to True!)
"""
if len(self.decision_boundary_points) == 0:
raise Exception("Please call the fit method first!")
if not generate_testpoints and generate_background:
print("Warning: cannot generate a background without testpoints")
if len(self.X_testpoints) == 0 and generate_testpoints:
if self.verbose:
print("Generating demo points around decision boundary...")
self._generate_testpoints()
if generate_background and generate_testpoints:
if tune_background_model:
params = {'C': np.power(10, np.linspace(0, 2, 2)),
'gamma': np.power(10, np.linspace(-2, 0, 2))}
grid = GridSearchCV(SVC(), params, n_jobs=-1 if os.name != 'nt' else 1)
grid.fit(np.vstack((self.X2d[self.train_idx], self.X_testpoints_2d)), np.hstack(
(self.y[self.train_idx], self.y_testpoints)))
bestparams = grid.best_params_
else:
bestparams = {'C': 1, 'gamma': 1}
self.background_model = SVC(probability=True, C=bestparams['C'], gamma=bestparams['gamma']).fit(np.vstack(
(self.X2d[self.train_idx], self.X_testpoints_2d)), np.hstack((self.y[self.train_idx], self.y_testpoints)))
xx, yy = np.meshgrid(np.linspace(self.X2d_xmin, self.X2d_xmax, background_resolution), np.linspace(
self.X2d_ymin, self.X2d_ymax, background_resolution))
Z = self.background_model.predict_proba(np.c_[xx.ravel(), yy.ravel()])[:, 0]
Z = Z.reshape((background_resolution, background_resolution))
self.background = Z
if generate_background and generate_testpoints:
return self.decision_boundary_points_2d, self.X_testpoints_2d, self.y_testpoints, Z
elif generate_testpoints:
return self.decision_boundary_points_2d, self.X_testpoints_2d, self.y_testpoints
else:
return self.decision_boundary_points_2d
def _generate_testpoints(self, tries=100):
"""Generate random demo points around decision boundary keypoints
"""
nn_model = NearestNeighbors(n_neighbors=3)
nn_model.fit(self.decision_boundary_points)
nn_model_2d = NearestNeighbors(n_neighbors=2)
nn_model_2d.fit(self.decision_boundary_points_2d)
#max_radius = 2*np.max([nn_model_2d.kneighbors([self.decision_boundary_points_2d[i]])[0][0][1] for i in range(len(self.decision_boundary_points_2d))])
self.X_testpoints = np.zeros((0, self.X.shape[1]))
self.y_testpoints = []
for i in range(len(self.decision_boundary_points)):
if self.verbose:
msg = "Generating testpoint for plotting {}/{}"
print(msg.format(i, len(self.decision_boundary_points)))
testpoints = np.zeros((0, self.X.shape[1]))
# generate Np points in Gaussian around decision_boundary_points[i] with
# radius depending on the distance to the next point
d, idx = nn_model.kneighbors([self.decision_boundary_points[i]])
radius = d[0][1] if d[0][1] != 0 else d[0][2]
if radius == 0:
radius = np.mean(pdist(self.decision_boundary_points_2d))
max_radius = radius * 2
radius /= 5.0
# add demo points, keeping some balance
max_imbalance = 5.0
y_testpoints = []
for j in range(self.n_generated_testpoints_per_keypoint - 2):
c_radius = radius
freq = itemfreq(y_testpoints).astype(float)
imbalanced = freq.shape[0] != 0
if freq.shape[0] == 2 and (freq[0, 1] / freq[1, 1] < 1.0 / max_imbalance or freq[0, 1] / freq[1, 1] > max_imbalance):
imbalanced = True
for try_i in range(tries):
testpoint = np.random.normal(self.decision_boundary_points[
i], radius, (1, self.X.shape[1]))
try:
testpoint2d = self.dimensionality_reduction.transform(testpoint)[0]
except: # DR can fail e.g. if NMF gets negative values
testpoint = []
continue
# demo point needs to be close to current key point
if euclidean(testpoint2d, self.decision_boundary_points_2d[i]) <= max_radius:
if not imbalanced: # needs to be not imbalanced
break
y_pred = self.classifier.predict(testpoint)[0]
# imbalanced but this would actually improve things
if freq.shape[0] == 2 and freq[y_pred, 1] < freq[1 - y_pred, 1]:
break
c_radius /= 2.0
if len(testpoint) != 0:
testpoints = np.vstack((testpoints, testpoint))
y_testpoints.append(self.classifier.predict(testpoint)[0])
self.X_testpoints = np.vstack((self.X_testpoints, testpoints))
self.y_testpoints = np.hstack((self.y_testpoints, y_testpoints))
self.X_testpoints_2d = self.dimensionality_reduction.transform(self.X_testpoints)
idx_within_bounds = np.where((self.X_testpoints_2d[:, 0] >= self.X2d_xmin) & (self.X_testpoints_2d[:, 0] <= self.X2d_xmax)
& (self.X_testpoints_2d[:, 1] >= self.X2d_ymin) & (self.X_testpoints_2d[:, 1] <= self.X2d_ymax))[0]
self.X_testpoints = self.X_testpoints[idx_within_bounds]
self.y_testpoints = self.y_testpoints[idx_within_bounds]
self.X_testpoints_2d = self.X_testpoints_2d[idx_within_bounds]
def decision_boundary_distance(self, x, grad=0):
"""Returns the distance of the given point from the decision boundary,
i.e. the distance from the region with maximal uncertainty (0.5
prediction probability)"""
return np.abs(0.5 - self.classifier.predict_proba([x])[0][1])
def get_decision_boundary_keypoints(self):
"""Returns the arrays of located decision boundary keypoints (both in the
original feature space, and in the dimensionality-reduced 2D space)
Returns
-------
decision_boundary_points : array
Array containing points in the original feature space which are very
close to the true decision boundary (closer than acceptance_threshold)
decision_boundary_points_2d : array
Array containing points in the dimensionality-reduced 2D space which
are very close to the true decision boundary
"""
if len(self.decision_boundary_points) == 0:
raise Exception("Please call the fit method first!")
return self.decision_boundary_points, self.decision_boundary_points_2d
def _get_sorted_db_keypoint_distances(self, N=None):
"""Use a minimum spanning tree heuristic to find the N largest gaps in the
line constituted by the current decision boundary keypoints.
"""
if N == None:
N = self.n_interpolated_keypoints
edges = minimum_spanning_tree(squareform(pdist(self.decision_boundary_points_2d)))
edged = np.array([euclidean(self.decision_boundary_points_2d[u],
self.decision_boundary_points_2d[v]) for u, v in edges])
gap_edge_idx = np.argsort(edged)[::-1][:N]
edges = edges[gap_edge_idx]
gap_distances = np.square(edged[gap_edge_idx])
gap_probability_scores = gap_distances / np.sum(gap_distances)
return edges, gap_distances, gap_probability_scores
def _linear_decision_boundary_optimization(self, from_idx, to_idx, all_combinations=True, retry_neighbor_if_failed=True, step=None, suppress_output=False):
"""Use global optimization to locate the decision boundary along lines
defined by instances from_idx and to_idx in the dataset (from_idx and to_idx
have to contain indices from distinct classes to guarantee the existence of
a decision boundary between them!)
"""
step_str = ("Step " + str(step) + "/" + str(self.steps) + ":") if step != None else ""
retries = 4 if retry_neighbor_if_failed else 1
for i in range(len(from_idx)):
n = len(to_idx) if all_combinations else 1
for j in range(n):
from_i = from_idx[i]
to_i = to_idx[j] if all_combinations else to_idx[i]
for k in range(retries):
if k == 0:
from_point = self.Xminor[from_i]
to_point = self.Xmajor[to_i]
else:
# first attempt failed, try nearest neighbors of source and destination
# point instead
_, idx = self.nn_model_2d_minorityclass.kneighbors([self.Xminor2d[from_i]])
from_point = self.Xminor[idx[0][k / 2]]
_, idx = self.nn_model_2d_minorityclass.kneighbors([self.Xmajor2d[to_i]])
to_point = self.Xmajor[idx[0][k % 2]]
if euclidean(from_point, to_point) == 0:
break # no decision boundary between equivalent points
db_point = self._find_decision_boundary_along_line(
from_point, to_point, penalize_tangent_distance=self.penalties_enabled, penalize_extremes=self.penalties_enabled)
if self.decision_boundary_distance(db_point) <= self.acceptance_threshold:
db_point2d = self.dimensionality_reduction.transform([db_point])[0]
if db_point2d[0] >= self.X2d_xmin and db_point2d[0] <= self.X2d_xmax and db_point2d[1] >= self.X2d_ymin and db_point2d[1] <= self.X2d_ymax:
self.decision_boundary_points.append(db_point)
self.decision_boundary_points_2d.append(db_point2d)
if self.verbose and not suppress_output:
# , ": New decision boundary keypoint found using linear optimization!"
print("{} {}/{}".format(step_str, i * n + j, len(from_idx) * n))
break
else:
if self.verbose and not suppress_output:
msg = "{} {}/{}: Rejected decision boundary keypoint (outside of plot area)"
print(msg.format(step_str, i * n + j, len(from_idx) * n))
def _find_decision_boundary_along_line(self, from_point, to_point, penalize_extremes=False, penalize_tangent_distance=False):
def objective(l, grad=0):
# interpolate between source and destionation; calculate distance from
# decision boundary
X = from_point + l[0] * (to_point - from_point)
error = self.decision_boundary_distance(X)
if penalize_tangent_distance:
# distance from tangent between class1 and class0 point in 2d space
x0, y0 = self.dimensionality_reduction.transform([X])[0]
x1, y1 = self.dimensionality_reduction.transform([from_point])[0]
x2, y2 = self.dimensionality_reduction.transform([to_point])[0]
error += 1e-12 * np.abs((y2 - y1) * x0 - (x2 - x1) * y0 +
x2 * y1 - y2 * x1) / np.sqrt((y2 - y1)**2 + (x2 - x1)**2)
if penalize_extremes:
error += 1e-8 * np.abs(0.5 - l[0])
return error
optimizer = self._get_optimizer()
optimizer.set_min_objective(objective)
cl = optimizer.optimize([random.random()])
db_point = from_point + cl[0] * (to_point - from_point)
return db_point
def _find_decision_boundary_on_hypersphere(self, centroid, R, penalize_known=False):
def objective(phi, grad=0):
# search on hypersphere surface in polar coordinates - map back to cartesian
cx = centroid + polar_to_cartesian(phi, R)
try:
cx2d = self.dimensionality_reduction.transform([cx])[0]
error = self.decision_boundary_distance(cx)
if penalize_known:
# slight penalty for being too close to already known decision boundary
# keypoints
db_distances = [euclidean(cx2d, self.decision_boundary_points_2d[k])
for k in range(len(self.decision_boundary_points_2d))]
error += 1e-8 * ((self.mean_2d_dist - np.min(db_distances)) /
self.mean_2d_dist)**2
return error
except (Exception, ex):
print("Error in objective function:", ex)
return np.infty
optimizer = self._get_optimizer(
D=self.X.shape[1] - 1, upper_bound=2 * np.pi, iteration_budget=self.hypersphere_iteration_budget)
optimizer.set_min_objective(objective)
db_phi = optimizer.optimize([rnd.random() * 2 * np.pi for k in range(self.X.shape[1] - 1)])
db_point = centroid + polar_to_cartesian(db_phi, R)
return db_point
def _get_optimizer(self, D=1, upper_bound=1, iteration_budget=None):
"""Utility function creating an NLOPT optimizer with default
parameters depending on this objects parameters
"""
if iteration_budget == None:
iteration_budget = self.linear_iteration_budget
opt = nlopt.opt(nlopt.GN_DIRECT_L_RAND, D)
# opt.set_stopval(self.acceptance_threshold/10.0)
opt.set_ftol_rel(1e-5)
opt.set_maxeval(iteration_budget)
opt.set_lower_bounds(0)
opt.set_upper_bounds(upper_bound)
return opt
class BaseLabelPropagation(six.with_metaclass(ABCMeta, BaseEstimator,
ClassifierMixin)):
"""Base class for label propagation module.
Parameters
----------
kernel : {'knn', 'rbf'}
String identifier for kernel function to use.
Only 'rbf' and 'knn' kernels are currently supported..
gamma : float
Parameter for rbf kernel
alpha : float
Clamping factor
max_iter : float
Change maximum number of iterations allowed
tol : float
Convergence tolerance: threshold to consider the system at steady
state
n_neighbors : integer > 0
Parameter for knn kernel
"""
def __init__(self, kernel='rbf', gamma=20, n_neighbors=7,
alpha=1, max_iter=30, tol=1e-3):
self.max_iter = max_iter
self.tol = tol
# kernel parameters
self.kernel = kernel
self.gamma = gamma
self.n_neighbors = n_neighbors
# clamping factor
self.alpha = alpha
def _get_kernel(self, X, y=None):
if self.kernel == "rbf":
if y is None:
return rbf_kernel(X, X, gamma=self.gamma)
else:
return rbf_kernel(X, y, gamma=self.gamma)
elif self.kernel == "knn":
if self.nn_fit is None:
self.nn_fit = NearestNeighbors(self.n_neighbors).fit(X)
if y is None:
# Nearest neighbors returns a directed matrix.
dir_graph = self.nn_fit.kneighbors_graph(self.nn_fit._fit_X,
self.n_neighbors,
mode='connectivity')
# Making the matrix symmetric
un_graph = dir_graph + dir_graph.T
# Since it is a connectivity matrix, all values should be
# either 0 or 1
un_graph[un_graph > 1.0] = 1.0
return un_graph
else:
return self.nn_fit.kneighbors(y, return_distance=False)
else:
raise ValueError("%s is not a valid kernel. Only rbf and knn"
" are supported at this time" % self.kernel)
@abstractmethod
def _build_graph(self):
raise NotImplementedError("Graph construction must be implemented"
" to fit a label propagation model.")
def predict(self, X):
"""Performs inductive inference across the model.
Parameters
----------
X : array_like, shape = [n_samples, n_features]
Returns
-------
y : array_like, shape = [n_samples]
Predictions for input data
"""
probas = self.predict_proba(X)
return self.classes_[np.argmax(probas, axis=1)].ravel()
def predict_proba(self, X):
"""Predict probability for each possible outcome.
Compute the probability estimates for each single sample in X
and each possible outcome seen during training (categorical
distribution).
Parameters
----------
X : array_like, shape = [n_samples, n_features]
Returns
-------
probabilities : array, shape = [n_samples, n_classes]
Normalized probability distributions across
class labels
"""
check_is_fitted(self, 'X_')
X_2d = check_array(X, accept_sparse = ['csc', 'csr', 'coo', 'dok',
'bsr', 'lil', 'dia'])
weight_matrices = self._get_kernel(self.X_, X_2d)
if self.kernel == 'knn':
probabilities = []
for weight_matrix in weight_matrices:
ine = np.sum(self.label_distributions_[weight_matrix], axis=0)
probabilities.append(ine)
probabilities = np.array(probabilities)
else:
weight_matrices = weight_matrices.T
probabilities = np.dot(weight_matrices, self.label_distributions_)
normalizer = np.atleast_2d(np.sum(probabilities, axis=1)).T
probabilities /= normalizer
return probabilities
def fit(self, X, y):
"""Fit a semi-supervised label propagation model based
All the input data is provided matrix X (labeled and unlabeled)
and corresponding label matrix y with a dedicated marker value for
unlabeled samples.
Parameters
----------
X : array-like, shape = [n_samples, n_features]
A {n_samples by n_samples} size matrix will be created from this
y : array_like, shape = [n_samples]
n_labeled_samples (unlabeled points are marked as -1)
All unlabeled samples will be transductively assigned labels
Returns
-------
self : returns an instance of self.
"""
X, y = check_X_y(X, y)
self.X_ = X
check_classification_targets(y)
# actual graph construction (implementations should override this)
graph_matrix = self._build_graph()
# label construction
# construct a categorical distribution for classification only
classes = np.unique(y)
classes = (classes[classes != -1])
self.classes_ = classes
n_samples, n_classes = len(y), len(classes)
y = np.asarray(y)
unlabeled = y == -1
clamp_weights = np.ones((n_samples, 1))
clamp_weights[~unlabeled, 0] = 1 - self.alpha
# initialize distributions
self.label_distributions_ = np.zeros((n_samples, n_classes))
for label in classes:
self.label_distributions_[y == label, classes == label] = 1
y_static = np.copy(self.label_distributions_)
if self.alpha > 0.:
y_static *= self.alpha
y_static[unlabeled] = 0
l_previous = np.zeros((self.X_.shape[0], n_classes))
remaining_iter = self.max_iter
if sparse.isspmatrix(graph_matrix):
graph_matrix = graph_matrix.tocsr()
while (_not_converged(self.label_distributions_, l_previous, self.tol)
and remaining_iter > 1):
l_previous = self.label_distributions_
self.label_distributions_ = safe_sparse_dot(
graph_matrix, self.label_distributions_)
# clamp
self.label_distributions_ = np.multiply(
clamp_weights, self.label_distributions_) + y_static
remaining_iter -= 1
normalizer = np.sum(self.label_distributions_, axis=1)[:, np.newaxis]
self.label_distributions_ /= normalizer
if remaining_iter <= 1:
warnings.warn('max_iter was reached without convergence.',
category=ConvergenceWarning)
# set the transduction item
transduction = self.classes_[np.argmax(self.label_distributions_,
axis=1)]
self.transduction_ = transduction.ravel()
self.n_iter_ = self.max_iter - remaining_iter
return self
def _generate_testpoints(self, tries=100):
"""Generate random demo points around decision boundary keypoints
"""
nn_model = NearestNeighbors(n_neighbors=3)
nn_model.fit(self.decision_boundary_points)
nn_model_2d = NearestNeighbors(n_neighbors=2)
nn_model_2d.fit(self.decision_boundary_points_2d)
#max_radius = 2*np.max([nn_model_2d.kneighbors([self.decision_boundary_points_2d[i]])[0][0][1] for i in range(len(self.decision_boundary_points_2d))])
self.X_testpoints = np.zeros((0, self.X.shape[1]))
self.y_testpoints = []
for i in range(len(self.decision_boundary_points)):
if self.verbose:
msg = "Generating testpoint for plotting {}/{}"
print(msg.format(i, len(self.decision_boundary_points)))
testpoints = np.zeros((0, self.X.shape[1]))
# generate Np points in Gaussian around decision_boundary_points[i] with
# radius depending on the distance to the next point
d, idx = nn_model.kneighbors([self.decision_boundary_points[i]])
radius = d[0][1] if d[0][1] != 0 else d[0][2]
if radius == 0:
radius = np.mean(pdist(self.decision_boundary_points_2d))
max_radius = radius * 2
radius /= 5.0
# add demo points, keeping some balance
max_imbalance = 5.0
y_testpoints = []
for j in range(self.n_generated_testpoints_per_keypoint - 2):
c_radius = radius
freq = itemfreq(y_testpoints).astype(float)
imbalanced = freq.shape[0] != 0
if freq.shape[0] == 2 and (freq[0, 1] / freq[1, 1] < 1.0 / max_imbalance or freq[0, 1] / freq[1, 1] > max_imbalance):
imbalanced = True
for try_i in range(tries):
testpoint = np.random.normal(self.decision_boundary_points[
i], radius, (1, self.X.shape[1]))
try:
testpoint2d = self.dimensionality_reduction.transform(testpoint)[0]
except: # DR can fail e.g. if NMF gets negative values
testpoint = []
continue
# demo point needs to be close to current key point
if euclidean(testpoint2d, self.decision_boundary_points_2d[i]) <= max_radius:
if not imbalanced: # needs to be not imbalanced
break
y_pred = self.classifier.predict(testpoint)[0]
# imbalanced but this would actually improve things
if freq.shape[0] == 2 and freq[y_pred, 1] < freq[1 - y_pred, 1]:
break
c_radius /= 2.0
if len(testpoint) != 0:
testpoints = np.vstack((testpoints, testpoint))
y_testpoints.append(self.classifier.predict(testpoint)[0])
self.X_testpoints = np.vstack((self.X_testpoints, testpoints))
self.y_testpoints = np.hstack((self.y_testpoints, y_testpoints))
self.X_testpoints_2d = self.dimensionality_reduction.transform(self.X_testpoints)
idx_within_bounds = np.where((self.X_testpoints_2d[:, 0] >= self.X2d_xmin) & (self.X_testpoints_2d[:, 0] <= self.X2d_xmax)
& (self.X_testpoints_2d[:, 1] >= self.X2d_ymin) & (self.X_testpoints_2d[:, 1] <= self.X2d_ymax))[0]
self.X_testpoints = self.X_testpoints[idx_within_bounds]
self.y_testpoints = self.y_testpoints[idx_within_bounds]
self.X_testpoints_2d = self.X_testpoints_2d[idx_within_bounds]
def _preprocess_neighbors(self, rebuild=False, save=True):
neighbors_model_path = os.path.join(
self.selected_dir,
"neighbors_model-step" + str(self.model.step) + ".pkl")
neighbors_path = os.path.join(
self.selected_dir,
"neighbors-step" + str(self.model.step) + ".npy")
neighbors_weight_path = os.path.join(
self.selected_dir,
"neighbors_weight-step" + str(self.model.step) + ".npy")
test_neighbors_path = os.path.join(
self.selected_dir,
"test_neighbors-step" + str(self.model.step) + ".npy")
test_neighbors_weight_path = os.path.join(
self.selected_dir,
"test_neighbors_weight-step" + str(self.model.step) + ".npy")
if os.path.exists(neighbors_model_path) and \
os.path.exists(neighbors_path) and \
os.path.exists(test_neighbors_path) and rebuild == False and DEBUG == False:
logger.info("neighbors and neighbor_weight exist!!!")
self.neighbors = np.load(neighbors_path)
self.neighbors_weight = np.load(neighbors_weight_path)
self.test_neighbors = np.load(test_neighbors_path)
return
logger.info("neighbors and neighbor_weight "
"do not exist, preprocessing!")
train_X = self.get_full_train_X()
train_num = train_X.shape[0]
train_y = self.get_full_train_label()
train_y = np.array(train_y)
test_X = self.get_test_X()
test_num = test_X.shape[0]
self.max_neighbors = min(len(train_y), self.max_neighbors)
logger.info("data shape: {}, labeled_num: {}".format(
str(train_X.shape), sum(train_y != -1)))
nn_fit = NearestNeighbors(7, n_jobs=-4).fit(train_X)
logger.info("nn construction finished!")
neighbor_result = nn_fit.kneighbors_graph(
nn_fit._fit_X,
self.max_neighbors,
# 2,
mode="distance")
test_neighbors_result = nn_fit.kneighbors_graph(test_X,
self.max_neighbors,
mode="distance")
logger.info("neighbor_result got!")
self.neighbors, self.neighbors_weight = self.csr_to_impact_matrix(
neighbor_result, train_num, self.max_neighbors)
self.test_neighbors, test_neighbors_weight = self.csr_to_impact_matrix(
test_neighbors_result, test_num, self.max_neighbors)
logger.info("preprocessed neighbors got!")
# save neighbors information
if save:
pickle_save_data(neighbors_model_path, nn_fit)
np.save(neighbors_path, self.neighbors)
np.save(neighbors_weight_path, self.neighbors_weight)
np.save(test_neighbors_path, self.test_neighbors)
np.save(test_neighbors_weight_path, test_neighbors_weight)
return self.neighbors, self.test_neighbors
def joint_information(x, y, n_neighbors=3, random_noise=0.3):
n_samples = x.size
if random_noise:
x = with_added_white_noise(x, random_noise)
y = with_added_white_noise(y, random_noise)
x = x.reshape((-1, 1))
y = y.reshape((-1, 1))
xy = np.hstack((x, y))
# Here we rely on NearestNeighbors to select the fastest algorithm.
nn = NearestNeighbors(metric='chebyshev', n_neighbors=n_neighbors)
nn.fit(xy)
radius = nn.kneighbors()[0]
radius = np.nextafter(radius[:, -1], 0)
# Algorithm is selected explicitly to allow passing an array as radius
# later (not all algorithms support this).
nn.set_params(algorithm='kd_tree')
nn.fit(x)
ind = nn.radius_neighbors(radius=radius, return_distance=False)
nx = np.array([i.size for i in ind])
nn.fit(y)
ind = nn.radius_neighbors(radius=radius, return_distance=False)
ny = np.array([i.size for i in ind])
mi = (digamma(n_samples) + digamma(n_neighbors) -
np.mean(digamma(nx + 1)) - np.mean(digamma(ny + 1)))
return max(0, mi)
def fit(self, X, y, training_indices=None):
"""Specify data to be plotted, and fit classifier only if required (the
specified clasifier is only trained if it has not been trained yet).
All the input data is provided in the matrix X, and corresponding
binary labels (values taking 0 or 1) in the vector y
Parameters
----------
X : array-like, shape = [n_samples, n_features]
A {n_samples by n_samples} size matrix containing data
y : array-like, shape = [n_samples]
Labels
training_indices : array-like or float, optional (default=None)
Indices on which the classifier has been trained / should be trained.
If float, it is converted to a random sample with the specified proportion
of the full dataset.
Returns
-------
self : returns an instance of self.
"""
if set(np.array(y, dtype=int).tolist()) != set([0, 1]):
raise Exception(
"Currently only implemented for binary classification. Make sure you pass in two classes (0 and 1)")
if training_indices == None:
train_idx = range(len(y))
elif type(training_indices) == float:
train_idx, test_idx = train_test_split(range(len(y)), test_size=0.5)
else:
train_idx = training_indices
self.X = X
self.y = y
self.train_idx = train_idx
#self.test_idx = np.setdiff1d(np.arange(len(y)), self.train_idx, assume_unique=False)
self.test_idx = list(set(range(len(y))).difference(set(self.train_idx)))
# fit classifier if necessary
try:
self.classifier.predict([X[0]])
except:
self.classifier.fit(X[train_idx, :], y[train_idx])
self.y_pred = self.classifier.predict(self.X)
# fit DR method if necessary
try:
self.dimensionality_reduction.transform([X[0]])
except:
self.dimensionality_reduction.fit(X, y)
try:
self.dimensionality_reduction.transform([X[0]])
except:
raise Exception(
"Please make sure your dimensionality reduction method has an exposed transform() method! If in doubt, use PCA or Isomap")
# transform data
self.X2d = self.dimensionality_reduction.transform(self.X)
self.mean_2d_dist = np.mean(pdist(self.X2d))
self.X2d_xmin, self.X2d_xmax = np.min(self.X2d[:, 0]), np.max(self.X2d[:, 0])
self.X2d_ymin, self.X2d_ymax = np.min(self.X2d[:, 1]), np.max(self.X2d[:, 1])
self.majorityclass = 0 if list(y).count(0) > list(y).count(1) else 1
self.minorityclass = 1 - self.majorityclass
minority_idx, majority_idx = np.where(y == self.minorityclass)[
0], np.where(y == self.majorityclass)[0]
self.Xminor, self.Xmajor = X[minority_idx], X[majority_idx]
self.Xminor2d, self.Xmajor2d = self.X2d[minority_idx], self.X2d[majority_idx]
# set up efficient nearest neighbor models for later use
self.nn_model_2d_majorityclass = NearestNeighbors(n_neighbors=2)
self.nn_model_2d_majorityclass.fit(self.X2d[majority_idx, :])
self.nn_model_2d_minorityclass = NearestNeighbors(n_neighbors=2)
self.nn_model_2d_minorityclass.fit(self.X2d[minority_idx, :])
# step 1. look for decision boundary points between corners of majority &
# minority class distribution
minority_corner_idx, majority_corner_idx = [], []
for extremum1 in [np.min, np.max]:
for extremum2 in [np.min, np.max]:
_, idx = self.nn_model_2d_minorityclass.kneighbors(
[[extremum1(self.Xminor2d[:, 0]), extremum2(self.Xminor2d[:, 1])]])
minority_corner_idx.append(idx[0][0])
_, idx = self.nn_model_2d_majorityclass.kneighbors(
[[extremum1(self.Xmajor2d[:, 0]), extremum2(self.Xmajor2d[:, 1])]])
majority_corner_idx.append(idx[0][0])
# optimize to find new db keypoints between corners
self._linear_decision_boundary_optimization(
minority_corner_idx, majority_corner_idx, all_combinations=True, step=1)
# step 2. look for decision boundary points on lines connecting randomly
# sampled points of majority & minority class
n_samples = int(self.n_connecting_keypoints)
from_idx = list(random.sample(list(np.arange(len(self.Xminor))), n_samples))
to_idx = list(random.sample(list(np.arange(len(self.Xmajor))), n_samples))
# optimize to find new db keypoints between minority and majority class
self._linear_decision_boundary_optimization(
from_idx, to_idx, all_combinations=False, step=2)
if len(self.decision_boundary_points_2d) < 2:
print("Failed to find initial decision boundary. Retrying... If this keeps happening, increasing the acceptance threshold might help. Also, make sure the classifier is able to find a point with 0.5 prediction probability (usually requires an even number of estimators/neighbors/etc).")
return self.fit(X, y, training_indices)
# step 3. look for decision boundary points between already known db
# points that are too distant (search on connecting line first, then on
# surrounding hypersphere surfaces)
edges, gap_distances, gap_probability_scores = self._get_sorted_db_keypoint_distances() # find gaps
self.nn_model_decision_boundary_points = NearestNeighbors(n_neighbors=2)
self.nn_model_decision_boundary_points.fit(self.decision_boundary_points)
i = 0
retries = 0
while i < self.n_interpolated_keypoints:
if self.verbose:
print("Step 3/{}:{}/".format(self.steps, i, self.n_interpolated_keypoints))
if self.random_gap_selection:
# randomly sample from sorted DB keypoint gaps?
gap_idx = np.random.choice(len(gap_probability_scores),
1, p=gap_probability_scores)[0]
else:
# get largest gap
gap_idx = 0
from_point = self.decision_boundary_points[edges[gap_idx][0]]
to_point = self.decision_boundary_points[edges[gap_idx][1]]
# optimize to find new db keypoint along line connecting two db keypoints
# with large gap
db_point = self._find_decision_boundary_along_line(
from_point, to_point, penalize_tangent_distance=self.penalties_enabled)
if self.decision_boundary_distance(db_point) > self.acceptance_threshold:
if self.verbose:
print("No good solution along straight line - trying to find decision boundary on hypersphere surface around known decision boundary point")
# hypersphere radius half the distance between from and to db keypoints
R = euclidean(from_point, to_point) / 2.0
# search around either source or target keypoint, with 0.5 probability,
# hoping to find decision boundary in between
if random.random() > 0.5:
from_point = to_point
# optimize to find new db keypoint on hypersphere surphase around known keypoint
db_point = self._find_decision_boundary_on_hypersphere(from_point, R)
if self.decision_boundary_distance(db_point) <= self.acceptance_threshold:
db_point2d = self.dimensionality_reduction.transform([db_point])[0]
self.decision_boundary_points.append(db_point)
self.decision_boundary_points_2d.append(db_point2d)
i += 1
retries = 0
else:
retries += 1
if retries > self.hypersphere_max_retry_budget:
i += 1
dist = self.decision_boundary_distance(db_point)
msg = "Found point is too distant from decision boundary ({}), but retry budget exceeded ({})"
print(msg.format(dist, self.hypersphere_max_retry_budget))
elif self.verbose:
dist = self.decision_boundary_distance(db_point)
print("Found point is too distant from decision boundary ({}) retrying...".format(dist))
else:
db_point2d = self.dimensionality_reduction.transform([db_point])[0]
self.decision_boundary_points.append(db_point)
self.decision_boundary_points_2d.append(db_point2d)
i += 1
retries = 0
edges, gap_distances, gap_probability_scores = self._get_sorted_db_keypoint_distances() # reload gaps
self.decision_boundary_points = np.array(self.decision_boundary_points)
self.decision_boundary_points_2d = np.array(self.decision_boundary_points_2d)
if self.verbose:
print("Done fitting! Found {} decision boundary keypoints.".format(
len(self.decision_boundary_points)))
return self
#coding:utf-8
'''
Created on 2018年1月23日
@author: root
'''
from sklearn.neighbors.unsupervised import NearestNeighbors
import numpy as np
from KNNDateOnHand import *
datingDataMat,datingLabels = file2matrix('datingTestSet2.txt')
normMat,ranges,minVals = autoNorm(datingDataMat)
nbrs = NearestNeighbors(n_neighbors=3).fit(normMat)
input_man= [30000,5,0.5]
S = (input_man - minVals)/ranges
distances, indices = nbrs.kneighbors(S)
print(indices)
print(distances)
# classCount K：类别名 V：这个类别中的样本出现的次数
classCount = {}
for i in range(3):
voteLabel = datingLabels[indices[0][i]]
classCount[voteLabel] = classCount.get(voteLabel,0) + 1
sortedClassCount = sorted(classCount.iteritems(),key=operator.itemgetter(1),reverse=True)
resultList = ['没感觉', '看起来还行','极具魅力']
print(resultList[sortedClassCount[0][0]-1])
class BaseLabelPropagation(BaseEstimator, ClassifierMixin, metaclass=ABCMeta):
def __init__(self,
kernel='rbf',
gamma=20,
n_neighbors=7,
alpha=1,
max_iter=30,
tol=1e-3,
n_jobs=None):
self.max_iter = max_iter
self.tol = tol
# kernel parameters
self.kernel = kernel
self.gamma = gamma
self.n_neighbors = n_neighbors
# clamping factor
self.alpha = alpha
self.n_jobs = n_jobs
def _get_kernel(self, X, y=None):
if self.kernel == "rbf":
if y is None:
return rbf_kernel(X, X, gamma=self.gamma)
else:
return rbf_kernel(X, y, gamma=self.gamma)
elif self.kernel == "knn":
if self.nn_fit is None:
self.nn_fit = NearestNeighbors(self.n_neighbors,
n_jobs=self.n_jobs).fit(X)
if y is None:
return self.nn_fit.kneighbors_graph(self.nn_fit._fit_X,
self.n_neighbors,
mode='connectivity')
else:
return self.nn_fit.kneighbors(y, return_distance=False)
elif callable(self.kernel):
if y is None:
return self.kernel(X, X)
else:
return self.kernel(X, y)
else:
raise ValueError("%s is not a valid kernel. Only rbf and knn"
" or an explicit function "
" are supported at this time." % self.kernel)
def fit(self, X, y):
"""
Parameters
----------
X : array-like ,shape = [n_samples, n_features]
input data matrix
y : array-like, shape = [n_samples]
n_labeled_samples (unlabeled = -1)
Returns
----------
self : returns an instance of self.
"""
# initialize X_
self.X_ = X
# actual graph construction
graph_matrix = self._build_graph()
# initialize classes
classes = np.unique(y)
classes = (classes[classes != -1]) ## self indexing array
self.classes_ = classes
# set n size
n_samples, n_classes = len(y), len(classes)
# set unlabeled to -1
y = np.asarray(y)
unlabeled = y == -1
# initialize distributions
self.label_distributions_ = np.zeros((n_samples, n_classes))
for label in classes:
self.label_distributions_[y == label, classes == label] = 1
y_static = np.copy(self.label_distributions_)
if self._variant == 'propagation':
y_static[unlabeled] = 0
# initialize l_previous
l_previous = np.zeros((self.X_.shape[0], n_classes))
# add a dimension to unlabeled
unlabeled = unlabeled[:, np.newaxis]
for self.n_iter_ in range(self.max_iter):
if np.abs(self.label_distributions_ - l_previous).sum() < self.tol:
break
l_previous = self.label_distributions_
self.label_distributions_ = safe_sparse_dot(
graph_matrix, self.label_distributions_) ## BLAS dot
if self._variant == 'propagation':
normalizer = np.sum(self.label_distributions_,
axis=1)[:, np.newaxis]
self.label_distributions_ /= normalizer
self.label_distributions_ = np.where(unlabeled,
self.label_distributions_,
y_static)
else:
self.n_iter_ += 1
normalizer = np.sum(self.label_distributions_, axis=1)[:, np.newaxis]
self.label_distributions_ /= normalizer
# set the transduction item
transduction = self.classes_[np.argmax(self.label_distributions_,
axis=1)]
self.transduction_ = transduction.ravel()
return self
def _build_graph(self):
"""Matrix representing a fully connected graph between each sample
This basic implementation creates a non-stochastic affinity matrix, so
class distributions will exceed 1 (normalization may be desired).
"""
if self.kernel == 'knn':
self.nn_fit = None
affinity_matrix = self._get_kernel(self.X_)
normalizer = affinity_matrix.sum(axis=0)
if sparse.isspmatrix(affinity_matrix):
affinity_matrix.data /= np.diag(np.array(normalizer))
else:
affinity_matrix /= normalizer[:, np.newaxis]
return affinity_matrix
