def fit(self, X, y=None):
# Compute the codebook
self.codebook = Codebook(K=self.K,
no_dump=self.no_dump_codebook,
force_reload=self.force_reload)
self.codebook.fit(X)
return self
class BoVWextractor(BaseEstimator):
def __init__(self, K=512, no_dump_codebook=False, force_reload=False):
self.K = K
self.no_dump_codebook=no_dump_codebook
self.force_reload = force_reload
def fit(self, X, y=None):
# Compute the codebook
self.codebook = Codebook(K=self.K,no_dump=self.no_dump_codebook, force_reload=self.force_reload)
self.codebook.fit(X['descriptors'])
return self
def transform(self, X):
print 'Getting BoVW representation'
init = time.time()
descriptors = X['descriptors']
visual_words = np.zeros((len(descriptors), self.K), dtype=np.float32)
for i in xrange(len(descriptors)):
words = self.codebook.predict(descriptors[i])
visual_words[i, :] = np.bincount(words, minlength=self.K)
end = time.time()
print '\tDone in ' + str(end - init) + ' secs.'
return visual_words
class SpatialPyramids(BaseEstimator):
def __init__(self, K=512, num_levels=3, no_dump_codebook=False):
self.K = K
self.num_levels = num_levels
self.no_dump_codebook = no_dump_codebook
def fit(self, X, y=None):
# Compute the codebook
self.codebook = Codebook(K=self.K, no_dump=self.no_dump_codebook)
self.codebook.fit(X['descriptors'])
return self
def transform(self, X):
print 'Getting Spatial Pyramid representation'
init = time.time()
descriptors = X['descriptors']
positions = X['positions']
imsizes = X['imsizes']
# Num. of cols/rows for each pyramid level
grid_ncolsrows = 2**np.arange(self.num_levels)
visual_words = np.zeros(
(len(descriptors), self.K * np.sum(grid_ncolsrows**2)),
dtype=np.float32)
for im in xrange(len(descriptors)):
# Compute the words
words = self.codebook.predict(descriptors[im])
# Compute the bincount for each grid cell in each pyramid level
current_vw = []
for l in range(self.num_levels):
r_vec = np.linspace(0,
imsizes[im][0] + 1,
num=grid_ncolsrows[l] + 1)
c_vec = np.linspace(0,
imsizes[im][1] + 1,
num=grid_ncolsrows[l] + 1)
for i in range(grid_ncolsrows[l]):
for j in range(grid_ncolsrows[l]):
rb = np.logical_and(positions[im][:, 0] >= r_vec[i],
positions[im][:, 0] < r_vec[i + 1])
cb = np.logical_and(positions[im][:, 1] >= c_vec[j],
positions[im][:, 1] < c_vec[j + 1])
current_vw.extend(
np.bincount(words[np.logical_and(rb, cb)],
minlength=self.K))
# Save the computed values
visual_words[im, :] = current_vw
end = time.time()
print '\tDone in ' + str(end - init) + ' secs.'
return visual_words
def train(self,
instances,
dev_set=None,
max_epoch=30,
learning_rate=.5,
batch_size=30):
# Construct a statistical model from labeled instances.
self.codebook = Codebook()
self.codebook.supervised_populate(instances)
self.parameters = np.zeros((self.codebook.dimension()))
self._train_sgd(instances, dev_set, max_epoch, learning_rate,
batch_size)
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy',
component_names=None):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
self._data = normalizer.normalize(data) if normalizer else data
self._normalizer = normalizer
self._dim = data.shape[1]
self._dlen = data.shape[0]
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask or np.ones([1, self._dim])
mapsize = self.calculate_map_size(lattice) if not mapsize else mapsize
self.codebook = Codebook(mapsize, lattice)
self.training = training
self._component_names = self.build_component_names(
) if component_names is None else [component_names]
self._distance_matrix = self.calculate_map_dist()
def __init__(self, shape):
self.codebooks = []
self.height = shape[1]
self.width = shape[0]
for i in range(self.width):
for j in range(self.height):
self.codebooks.append(Codebook())
def __init__(self, Y, D, sigma, b=1e-8, scale=1):
self.input_shape = Y.shape
self.sigma = sigma
self.scale = scale
m, r, c = D.shape
_, _, self.h, self.w = Y.shape
self.hx = int(np.ceil(self.h * scale)) if self.h > 1 else self.h
self.wx = int(np.ceil(self.w * scale)) if self.w > 1 else self.w
bh = int(2 * np.ceil(3 * sigma[0] * scale) + 1) if self.h > 1 else 1
bw = int(2 * np.ceil(3 * sigma[1] * scale) + 1) if self.w > 1 else 1
self.psf_support_scaled = (bh, bw)
bh = int(2 * np.ceil(3 * sigma[0]) + 1) if self.h > 1 else 1
bw = int(2 * np.ceil(3 * sigma[1]) + 1) if self.w > 1 else 1
self.psf_support = (bh, bw)
# Set up X
x_shape = (m, self.hx, self.wx)
self.X = torch.ones(x_shape).float()
# Set up Y
self.Y = torch.tensor(Y).float().flatten(start_dim=0, end_dim=1)
# Set up b
self.b = torch.tensor(b).float()
if self.b.ndim == 4:
self.b = self.b.flatten(start_dim=0, end_dim=1)
# Set up codebook
self.codebook = Codebook(D)
# Set up spatial blurr
self.psf = PSF((self.h, self.w), sigma, scale)
# Prepare constant
ones_channels = torch.ones((r * c, 1, 1))
ones_space = torch.ones((1, self.h, self.w))
self.denominator = self.codebook.matmul_t(ones_channels) * \
self.psf.matmul_t(ones_space)
# Compute Yhat = DXG
self.Yhat = self.__forward(self.X)
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy'):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows, as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
self._data = normalizer.normalize(data) if normalizer else data
self._normalizer = normalizer
self._dim = data.shape[1]
self._dlen = data.shape[0]
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask or np.ones([1, self._dim])
self.codebook = Codebook(mapsize, lattice)
self.training = training
self._component_names = self.build_component_names()
self._distance_matrix = self.calculate_map_dist()
"../images/Crowd_PETS09/S0/Background/View_001/Time_13-06/"))
EPSILON_1 = 1
TRAIN_N_IMG = 10
if __name__ == '__main__':
img_list = os.listdir(TRAINING_PATH)[:TRAIN_N_IMG]
for it, file in enumerate(img_list):
img = cv2.cvtColor(cv2.imread(os.path.join(TRAINING_PATH, file)),
cv2.COLOR_BGR2RGB).astype(float)
if it == 0:
codebooks = np.array(
[Codebook() for pixel in range(img.shape[0] * img.shape[1])])
for px_row in img:
for px_idx, px in enumerate(px_row):
R = px[0]
G = px[1]
B = px[2]
X = np.array([R, G, B])
I = sqrt(R**2 + G**2 + B**2)
pixel = Pixel(X, I)
codebook_empty = True
no_match = True
for codeword in codebooks[px_idx].codewords:
def codebook(self):
""" Export a code book of categories and codes.
"""
Codebook(self.settings, self.ui.textEdit)
def fit(self, X, y=None):
# Compute the codebook
self.codebook = Codebook(K=self.K, no_dump=self.no_dump_codebook)
self.codebook.fit(X['descriptors'])
return self
class SOM(object):
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy'):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
print "DDDD", data
self._data = data # normalizer.normalize(data) if normalizer else data
print "data : \n", self._data
print "map size : ", mapsize
self._normalizer = normalizer
print "_normalizer is a object to class normalizer : ", self._normalizer
self._dim = data.shape[1]
print "data shape[1] : ", self._dim
self._dlen = data.shape[0]
print "data shape[0] : ", data.shape[0]
self._dlabel = None
print "data label : ", self._dlabel
self._bmu = None
print "intial bmu : ", self._bmu
self.name = name
print "name : ", self.name
self.data_raw = data
print "data_raw : ", self.data_raw
#print "data_raw : ", self.data_raw
self.neighborhood = neighborhood
self.mapshape = mapshape
print "mapshape : ", self.mapshape
self.initialization = initialization
print "initilization : ", self.initialization
self.mask = mask or np.ones([1, self._dim])
print "mask : ", self.mask
self.codebook = Codebook(mapsize, lattice)
print "--------codebook is a object to class Codebook---------\n", self.codebook
self.training = training
print "training mode : ", self.training
self._component_names = self.build_component_names()
print "Initial component name : ", self._component_names
self._distance_matrix = self.calculate_map_dist()
print "initial distance matrix : ", self._distance_matrix
#self._data_labels=self.build_data_labels()
#print "data labels : ", self._data_labels
@property
def component_names(self):
return self._component_names
@component_names.setter
def component_names(self, compnames):
if self._dim == len(compnames):
self._component_names = np.asarray(compnames)[np.newaxis, :]
else:
raise ComponentNamesError('Component names should have the same '
'size as the data dimension/features')
def build_component_names(self):
cc = ['Variable-' + str(i + 1) for i in range(0, self._dim)]
return np.asarray(cc)[np.newaxis, :]
@property
def data_labels(self):
return self._dlabel
@data_labels.setter
def data_labels(self, labels):
"""
Set labels of the training data, it should be in the format of a list
of strings
"""
if labels.shape == (1, self._dlen):
label = labels.T
elif labels.shape == (self._dlen, 1):
label = labels
elif labels.shape == (self._dlen, ):
label = labels[:, np.newaxis]
else:
raise LabelsError('wrong label format')
self._dlabel = label
def build_data_labels(self):
cc = ['dlabel-' + str(i) for i in range(0, self._dlen)]
return np.asarray(cc)[:, np.newaxis]
def calculate_map_dist(self):
"""
Calculates the grid distance, which will be used during the training
steps. It supports only planar grids for the moment
"""
nnodes = self.codebook.nnodes
distance_matrix = np.zeros((nnodes, nnodes))
for i in range(nnodes):
distance_matrix[i] = self.codebook.grid_dist(i).reshape(1, nnodes)
return distance_matrix
@timeit()
def train(self, n_job=1, shared_memory=False, verbose='info'):
"""
Trains the som
:param n_job: number of jobs to use to parallelize the traning
:param shared_memory: flag to active shared memory
:param verbose: verbosity, could be 'debug', 'info' or None
"""
logging.root.setLevel(
getattr(logging, verbose.upper()) if verbose else logging.ERROR)
logging.info(" Training...")
logging.debug((
"--------------------------------------------------------------\n"
" details: \n"
" > data len is {data_len} and data dimension is {data_dim}\n"
" > map size is {mpsz0},{mpsz1}\n"
" > array size in log10 scale is {array_size}\n"
" > number of jobs in parallel: {n_job}\n"
" -------------------------------------------------------------\n"
).format(data_len=self._dlen,
data_dim=self._dim,
mpsz0=self.codebook.mapsize[0],
mpsz1=self.codebook.mapsize[1],
array_size=np.log10(self._dlen * self.codebook.nnodes *
self._dim),
n_job=n_job))
if self.initialization == 'random':
self.codebook.random_initialization(self._data)
elif self.initialization == 'pca':
self.codebook.pca_linear_initialization(self._data)
self.rough_train(njob=n_job, shared_memory=shared_memory)
self.finetune_train(njob=n_job, shared_memory=shared_memory)
logging.debug(
" --------------------------------------------------------------")
logging.info(" Final quantization error: %f" % np.mean(self._bmu[1]))
def _calculate_ms_and_mpd(self):
print "--------------------------------------------"
print "_calculation of ms and mpd .........\n"
mn = np.min(self.codebook.mapsize)
max_s = max(self.codebook.mapsize[0], self.codebook.mapsize[1])
if mn == 1:
mpd = float(self.codebook.nnodes * 10) / float(self._dlen)
else:
mpd = float(self.codebook.nnodes) / float(self._dlen)
ms = max_s / 2.0 if mn == 1 else max_s
print "ms , mpd : ", ms, mpd
return ms, mpd
def rough_train(self, njob=1, shared_memory=False):
print "--------------------------------------------"
logging.info(" Rough training...")
#print "rough training start here ---------------\n"
ms, mpd = self._calculate_ms_and_mpd()
trainlen, radiusin, radiusfin = int(np.ceil(30 * mpd)), None, None
if self.initialization == 'random':
radiusin = max(1, np.ceil(ms / 3.))
radiusfin = max(1, radiusin / 6.)
print "trainlen, radiusin , radiusfin if intialization is random ", trainlen, radiusin, radiusfin
elif self.initialization == 'pca':
radiusin = max(1, np.ceil(ms / 8.))
radiusfin = max(1, radiusin / 4.)
print "trainlen, radiusin , radiusfin if intialization is pca", trainlen, radiusin, radiusfin
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory)
print "rough training END here ---------------\n"
print "++++++++++++++++++++++++++++++++++++++++++++"
def finetune_train(self, njob=1, shared_memory=False):
print "#################################################--START"
logging.info(" Finetune training...")
ms, mpd = self._calculate_ms_and_mpd()
trainlen, radiusin, radiusfin = None, None, None
if self.initialization == 'random':
trainlen = int(np.ceil(50 * mpd))
radiusin = max(1, ms / 12.) # from radius fin in rough training
radiusfin = max(1, radiusin / 25.)
print "trainlen, radiusin , radiusfin if intialization is random ", trainlen, radiusin, radiusfin
elif self.initialization == 'pca':
trainlen = int(np.ceil(40 * mpd))
radiusin = max(1, np.ceil(ms / 8.) / 4)
radiusfin = 1 # max(1, ms/128)
print "trainlen, radiusin , radiusfin if intialization is pca", trainlen, radiusin, radiusfin
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory)
print "#################################################--END"
def _batchtrain(self,
trainlen,
radiusin,
radiusfin,
njob=1,
shared_memory=False):
print "--------------------------------------------"
radius = np.linspace(radiusin, radiusfin, trainlen)
print "radius : ", radius
if shared_memory:
data = self._data
data_folder = tempfile.mkdtemp()
data_name = os.path.join(data_folder, 'data')
dump(data, data_name)
data = load(data_name, mmap_mode='r')
else:
data = self._data
print "data : ", data
bmu = None
# X2 is part of euclidean distance (x-y)^2 = x^2 +y^2 - 2xy that we use
# for each data row in bmu finding.
# Since it is a fixed value we can skip it during bmu finding for each
# data point, but later we need it calculate quantification error
fixed_euclidean_x2 = np.einsum('ij,ij->i', data, data)
#print "fixed_euclidean_x2 : ", fixed_euclidean_x2
#print "len of fixed_euclidean_x2 : ", len(fixed_euclidean_x2)
#print data
logging.info(" radius_ini: %f , radius_final: %f, trainlen: %d\n" %
(radiusin, radiusfin, trainlen))
print "traing starts here up to trainlen................."
for i in range(trainlen):
t1 = time()
neighborhood = self.neighborhood.calculate(self._distance_matrix,
radius[i],
self.codebook.nnodes)
#print "neighbour \n: ", neighborhood
#print "CODEBOOK MATRIX \n: ", self.codebook.matrix
bmu = self.find_bmu(data, njb=njob)
#print "bmu \n: ", bmu
print "intial codebook matrix : \n", self.codebook.matrix
self.codebook.matrix = self.update_codebook_voronoi(
data, bmu, neighborhood)
#print "codebook matrix after updation : \n", self.codebook.matrix
print "updated codebook matrix : \n", self.codebook.matrix
qerror = (i + 1, round(time() - t1, 3),
np.mean(np.sqrt(bmu[1] + fixed_euclidean_x2)))
logging.info(
" epoch: %d ---> elapsed time: %f, quantization error: %f\n" %
qerror)
break
bmu[1] = np.sqrt(bmu[1] + fixed_euclidean_x2)
self._bmu = bmu
print "bmu after training : ", self._bmu
print "++++++++++++++++++++++++++++++++++++++++++++"
@timeit(logging.DEBUG)
def find_bmu(self, input_matrix, njb=1):
"""
Finds the best matching unit (bmu) for each input data from the input
matrix. It does all at once parallelizing the calculation instead of
going through each input and running it against the codebook.
:param input_matrix: numpy matrix representing inputs as rows and
features/dimension as cols
:param njb: number of jobs to parallelize the search
:returns: the best matching unit for each input
"""
print "--------------------------------------------"
dlen = input_matrix.shape[0]
#print "codebook matrix : \n ", self.codebook.matrix
y2 = np.einsum('ij,ij->i', self.codebook.matrix, self.codebook.matrix)
#print "np.einsum('ij,ij->i', self.codebook.matrix, self.codebook.matrix) :\n", np.einsum('ij,ij->i', self.codebook.matrix, self.codebook.matrix)
parallelizer = Parallel(n_jobs=njb, pre_dispatch='3*n_jobs')
chunk_bmu_finder = delayed(_chunk_based_bmu_find)
def row_chunk(part):
return part * dlen // njb
def col_chunk(part):
return min((part + 1) * dlen // njb, dlen)
b = parallelizer(
chunk_bmu_finder(input_matrix[row_chunk(i):col_chunk(i)],
self.codebook.matrix, y2) for i in range(njb))
bmu = np.asarray(list(itertools.chain(*b))).T
del b
#print "bmu \n: ", bmu
return bmu
print "++++++++++++++++++++++++++++++++++++++++++++"
@timeit(logging.DEBUG)
def update_codebook_voronoi(self, training_data, bmu, neighborhood):
"""
Updates the weights of each node in the codebook that belongs to the
bmu's neighborhood.
First finds the Voronoi set of each node. It needs to calculate a
smaller matrix.
Super fast comparing to classic batch training algorithm, it is based
on the implemented algorithm in som toolbox for Matlab by Helsinky
University.
:param training_data: input matrix with input vectors as rows and
vector features as cols
:param bmu: best matching unit for each input data. Has shape of
(2, dlen) where first row has bmu indexes
:param neighborhood: matrix representing the neighborhood of each bmu
:returns: An updated codebook that incorporates the learnings from the
input data
"""
print "--------------------------------------------"
print "strating of update_codebook_voronoi function ...."
row = bmu[0].astype(int)
print "rows (contain bmu index) : ", row
col = np.arange(self._dlen)
print "coulms (contain data sample no.) : ", col
val = np.tile(1, self._dlen)
P = csr_matrix((val, (row, col)),
shape=(self.codebook.nnodes, self._dlen))
print "csr matrix in which csr_matrix[k, k]=val[k] : ", P
S = P.dot(
training_data
) #each neuron has sum of data samples which are matched to that neuron
print "sparse matrix after multiplying with training data : \n", S
# neighborhood has nnodes*nnodes and S has nnodes*dim
# ---> Nominator has nnodes*dim
nom = neighborhood.T.dot(S)
print "nominator : ", nom
print "P.sum(axis=1) and its shape : ", P.sum(axis=1), P.sum(
axis=1).shape
nV = P.sum(axis=1).reshape(
1, self.codebook.nnodes
) #indicate how many data sample match to a paritcular neuron
print "nV : ", nV
print "neighborhood.T : \n", neighborhood.T
print "nV.dot(neighborhood.T) and its shape \n : ", nV.dot(
neighborhood.T), nV.dot(neighborhood.T).shape
denom = nV.dot(neighborhood.T).reshape(self.codebook.nnodes, 1)
print "denominator : \n", denom
new_codebook = np.divide(nom, denom)
print "new codebook : \n ", new_codebook
print "new codebokk round up to decimal 6 : ", np.around(new_codebook,
decimals=6)
return np.around(new_codebook, decimals=6)
def project_data(self, data):
"""
Projects a data set to a trained SOM. It is based on nearest
neighborhood search module of scikitlearn, but it is not that fast.
"""
clf = neighbors.KNeighborsClassifier(n_neighbors=1)
#print "clf : ", clf
#print "codebook.matrix.shape[0] :",self.codebook.matrix.shape[0]
labels = np.arange(0, self.codebook.matrix.shape[0])
#print "labels : ",labels
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
data = self._normalizer.normalize_by(self.data_raw, data)
print "------"
print "normal data : ", clf.predict(data)
print "++++++"
return clf.predict(data)
def predict_by(self, data, target, k=5, wt='distance'):
# here it is assumed that target is the last column in the codebook
# and data has dim-1 columns
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indX = ind[ind != target]
x = self.codebook.matrix[:, indX]
y = self.codebook.matrix[:, target]
n_neighbors = k
clf = neighbors.KNeighborsRegressor(n_neighbors, weights=wt)
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indX]
elif dimdata == dim - 1:
data = self._normalizer.normalize_by(self.data_raw[:, indX], data)
predicted_values = clf.predict(data)
predicted_values = self._normalizer.denormalize_by(
self.data_raw[:, target], predicted_values)
return predicted_values
def predict(self, x_test, k=5, wt='distance'):
"""
Similar to SKlearn we assume that we have X_tr, Y_tr and X_test. Here
it is assumed that target is the last column in the codebook and data
has dim-1 columns
:param x_test: input vector
:param k: number of neighbors to use
:param wt: method to use for the weights
(more detail in KNeighborsRegressor docs)
:returns: predicted values for the input data
"""
target = self.data_raw.shape[1] - 1
x_train = self.codebook.matrix[:, :target]
y_train = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights=wt)
clf.fit(x_train, y_train)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
x_test = self._normalizer.normalize_by(self.data_raw[:, :target],
x_test)
predicted_values = clf.predict(x_test)
return self._normalizer.denormalize_by(self.data_raw[:, target],
predicted_values)
def find_k_nodes(self, data, k=5):
from sklearn.neighbors import NearestNeighbors
# we find the k most similar nodes to the input vector
neighbor = NearestNeighbors(n_neighbors=k, metric='euclidean')
neighbor.fit(self.codebook.matrix)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
return neighbor.kneighbors(
self._normalizer.normalize_by(self.data_raw, data))
def bmu_ind_to_xy(self, bmu_ind):
"""
Translates a best matching unit index to the corresponding
matrix x,y coordinates.
:param bmu_ind: node index of the best matching unit
(number of node from top left node)
:returns: corresponding (x,y) coordinate
"""
rows = self.codebook.mapsize[0]
cols = self.codebook.mapsize[1]
# bmu should be an integer between 0 to no_nodes
out = np.zeros((bmu_ind.shape[0], 3))
out[:, 2] = bmu_ind
out[:, 0] = rows - 1 - bmu_ind / cols
out[:, 0] = bmu_ind / cols
out[:, 1] = bmu_ind % cols
return out.astype(int)
def cluster(self, n_clusters=3):
import sklearn.cluster as clust
cl_labels = clust.KMeans(n_clusters=n_clusters).fit_predict(
self._normalizer.denormalize_by(self.data_raw,
self.codebook.matrix))
self.cluster_labels = cl_labels
return cl_labels
def predict_probability(self, data, target, k=5):
"""
Predicts probability of the input data to be target
:param data: data to predict, it is assumed that 'target' is the last
column in the codebook, so data hould have dim-1 columns
:param target: target to predict probability
:param k: k parameter on KNeighborsRegressor
:returns: probability of data been target
"""
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indx = ind[ind != target]
x = self.codebook.matrix[:, indx]
y = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights='distance')
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indx]
elif dimdata == dim - 1:
data = self._normalizer.normalize_by(self.data_raw[:, indx], data)
weights, ind = clf.kneighbors(data,
n_neighbors=k,
return_distance=True)
weights = 1. / weights
sum_ = np.sum(weights, axis=1)
weights = weights / sum_[:, np.newaxis]
labels = np.sign(self.codebook.matrix[ind, target])
labels[labels >= 0] = 1
# for positives
pos_prob = labels.copy()
pos_prob[pos_prob < 0] = 0
pos_prob *= weights
pos_prob = np.sum(pos_prob, axis=1)[:, np.newaxis]
# for negatives
neg_prob = labels.copy()
neg_prob[neg_prob > 0] = 0
neg_prob = neg_prob * weights * -1
neg_prob = np.sum(neg_prob, axis=1)[:, np.newaxis]
return np.concatenate((pos_prob, neg_prob), axis=1)
def node_activation(self, data, target=None, wt='distance'):
weights, ind = None, None
if not target:
clf = neighbors.KNeighborsClassifier(
n_neighbors=self.codebook.nnodes)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
data = self._normalizer.normalize_by(self.data_raw, data)
weights, ind = clf.kneighbors(data)
print "--------------\n", weights, ind
# Softmax function
weights = 1. / weights
return weights, ind
class MaxEnt:
# -*- mode: Python; coding: utf-8 -*-
def __init__(self):
self.parameters = np.zeros((0, 0))
def train(self,
instances,
dev_set=None,
max_epoch=30,
learning_rate=.5,
batch_size=30):
# Construct a statistical model from labeled instances.
self.codebook = Codebook()
self.codebook.supervised_populate(instances)
self.parameters = np.zeros((self.codebook.dimension()))
self._train_sgd(instances, dev_set, max_epoch, learning_rate,
batch_size)
def _mini_batch(self, instances, batch_size):
# Yield mini-batches from the original data
shuffle(instances)
for i in range(0, len(instances), batch_size):
yield instances[i:i + batch_size]
def _compute_gradient(self, batch):
# Compute the gradient given the current batch of data
log_likelihood = 0
observed_count = np.zeros(self.codebook.dimension())
expected_count = np.zeros(self.codebook.dimension())
for datapoint in batch:
feature_map = [
self.codebook.feature_index(feature)
for feature in datapoint.features()
]
observed_count[feature_map,
self.codebook.label_index(datapoint.label)] += 1
lambda_vector = self.parameters[feature_map, :].sum(0)
log_likelihood -= sum(lambda_vector) - logsumexp(lambda_vector)
posterior = np.exp(
lambda_vector[self.codebook.label_index(datapoint.label)] -
logsumexp(lambda_vector))
expected_count[
feature_map,
self.codebook.label_index(datapoint.label)] += posterior
return observed_count - expected_count, log_likelihood
def _train_sgd(self, train_instances, dev_set, max_epoch, learning_rate,
batch_size):
# Train MaxEnt model with Mini-batch Gradient Descent
for epoch in range(1, max_epoch + 1):
for batch in self._mini_batch(train_instances, batch_size):
gradient, log_likelihood = self._compute_gradient(batch)
self.parameters += gradient * learning_rate
if dev_set:
print("(Epoch, accuracy):", (epoch, self.accuracy(dev_set)))
def accuracy(self, instances):
# Simple accuracy test for the dev set
current_state = [self.classify(x) == x.label for x in instances]
return float(sum(current_state)) / len(current_state)
def classify(self, instance):
feature_map = [
self.codebook.feature_index(feature)
for feature in instance.features()
if feature in self.codebook._features2index
]
lambda_vector = self.parameters[feature_map, :].sum(0)
posteriors = np.exp(lambda_vector - logsumexp(lambda_vector))
return self.codebook.get_label(np.argmax(posteriors))
class ISTDeco:
'''
ISTDECO - Deconvovles 1D or 2D spatial transcriptomic data
Parameters
----------
Y : float
Input image data of shape (rounds, channels, height, width)
D : float
Codebook of shape (ncodes, rounds, channels)
sigma : tuple(float,float)
Tuple of values corresponding to the standard deviation
of the gaussian shaped psf.
b : float
Background offset parameter. Can be a constant or same shape as Y.
Must be positive. Default: 1e-5
scale : float
We can deconvolve the image data to higher/lower spatial resolution.
Scale determines the scale factor. Defalut: 1.0 (no upscaling)
Example
----------
model = ISTDeco(Y, D, sigma, b=1e-5, upscale=1)
model = model.to('cuda') # If we want GPU
X, Q, loss = model.run(niter=100)
'''
def __init__(self, Y, D, sigma, b=1e-8, scale=1):
self.input_shape = Y.shape
self.sigma = sigma
self.scale = scale
m, r, c = D.shape
_, _, self.h, self.w = Y.shape
self.hx = int(np.ceil(self.h * scale)) if self.h > 1 else self.h
self.wx = int(np.ceil(self.w * scale)) if self.w > 1 else self.w
bh = int(2 * np.ceil(3 * sigma[0] * scale) + 1) if self.h > 1 else 1
bw = int(2 * np.ceil(3 * sigma[1] * scale) + 1) if self.w > 1 else 1
self.psf_support_scaled = (bh, bw)
bh = int(2 * np.ceil(3 * sigma[0]) + 1) if self.h > 1 else 1
bw = int(2 * np.ceil(3 * sigma[1]) + 1) if self.w > 1 else 1
self.psf_support = (bh, bw)
# Set up X
x_shape = (m, self.hx, self.wx)
self.X = torch.ones(x_shape).float()
# Set up Y
self.Y = torch.tensor(Y).float().flatten(start_dim=0, end_dim=1)
# Set up b
self.b = torch.tensor(b).float()
if self.b.ndim == 4:
self.b = self.b.flatten(start_dim=0, end_dim=1)
# Set up codebook
self.codebook = Codebook(D)
# Set up spatial blurr
self.psf = PSF((self.h, self.w), sigma, scale)
# Prepare constant
ones_channels = torch.ones((r * c, 1, 1))
ones_space = torch.ones((1, self.h, self.w))
self.denominator = self.codebook.matmul_t(ones_channels) * \
self.psf.matmul_t(ones_space)
# Compute Yhat = DXG
self.Yhat = self.__forward(self.X)
def to(self, device):
'''
Puts tensors on a device. See pytorch doc for more info.
Useful for moving tensors to cuda.
Example
----------
model = ISTDECO(Y,D,sigma)
model.to('cuda') # Put tensors on GPU
model.to('cpu') # Put tensors on CPU
Parameters
----------
device : str
The device, for instance 'cpu' or 'cuda'
'''
self.Y = self.Y.to(device)
self.Yhat = self.Yhat.to(device)
self.X = self.X.to(device)
self.denominator = self.denominator.to(device)
self.codebook = self.codebook.to(device)
self.psf = self.psf.to(device)
self.b = self.b.to(device)
return self
def __forward(self, tensor):
return self.psf.matmul(self.codebook.matmul(tensor)) + self.b
def __compute_quality(self, tensor):
# Pool intensities spatially
tensor_blurr = torch.nn.functional.avg_pool2d(tensor, \
self.psf_support_scaled,\
stride=1,\
divisor_override=1,\
padding=tuple(t//2 for t in self.psf_support_scaled))
tensor_blurr2 = self.psf.up_pool(
torch.nn.functional.relu(self.Y - self.b))
# Compute quality feature
Q = tensor_blurr / self.codebook.matmul_t(tensor_blurr2)
Q[torch.isnan(Q)] = 0
return Q
def run(self, niter=100, acc=1.0, suppress_radius=1):
'''
Run the optimization
Parameters
----------
niter : int
Number of iterations
acc : float
Factor for acceleration the multiplicative updates
If too large, a convergence may be unstalbe. Usually a
value between 1.0 and 1.5 is fine. Default: 1.0
suppress_width : int
Integer indicating the radius of the non-maxima supression footprint.
Default: 1.
Outputs
---------
X : numpy array
A multi-channel image of shape (m, sy, sx) where
m is the number of barcodes, sy and sx are the scaled height and width.
The values in X corresponds to the intensity of different barcodes.
For instance, X_kij is the intensity of the barcode with code k, localized
at i and j.
Q : numpy array
A multi-channel image of shape (m, sy, sx) where
m is the number of barcodes, sy and sx are the scaled height and width.
The values in Q are useful for elimintaing false-positive detections during
pos-processing.
loss : numpy array
An (niter,) shaped array with values
'''
loss = torch.zeros((niter, ))
for i in range(niter):
loss[i] = torch.sum(self.Yhat) - \
torch.sum(self.Y*torch.log(self.Yhat+1e-9))
self.X = self.X * (self.codebook.matmul_t(
self.psf.matmul_t(self.Y / self.Yhat)) / self.denominator)**acc
self.Yhat = self.__forward(self.X)
Q = self.__compute_quality(self.X)
if suppress_radius is not None:
mask = self.__nonmaxsuppress(suppress_radius, self.X)
self.X = self.X * mask
Q = Q * mask
return self.X.cpu().numpy(), Q.cpu().numpy(), loss
def __nonmaxsuppress(self, radius, tensor):
padd = [radius if self.h > 1 else 0, radius]
kernel_sz = (2 * radius + 1 if self.h > 1 else 1, 2 * radius + 1)
mask = torch.nn.functional.max_pool2d(tensor,
kernel_sz,
stride=1,
padding=padd) == tensor
#ints = torch.nn.functional.avg_pool2d(tensor, kernel_sz, stride=1, padding=padd, divisor_override=1)
return mask
class SOM(object):
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy'):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows, as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
self._data = normalizer.normalize(data) if normalizer else data
self._normalizer = normalizer
self._dim = data.shape[1]
self._dlen = data.shape[0]
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask or np.ones([1, self._dim])
self.codebook = Codebook(mapsize, lattice)
self.training = training
self._component_names = self.build_component_names()
self._distance_matrix = self.calculate_map_dist()
#self.set_data_labels() # slow for large data sets
@property
def component_names(self):
return self._component_names
@component_names.setter
def component_names(self, compnames):
if self._dim == len(compnames):
self._component_names = np.asarray(compnames)[np.newaxis, :]
else:
raise ComponentNamesError('Component names should have the same size as the data dimension/features')
def build_component_names(self):
cc = ['Variable-' + str(i+1) for i in range(0, self._dim)]
return np.asarray(cc)[np.newaxis, :]
@property
def data_labels(self):
return self._dlabel
@data_labels.setter
def data_labels(self, labels):
"""
Set labels of the training data, it should be in the format of a list of strings
"""
if labels.shape == (1, self._dlen):
label = labels.T
elif labels.shape == (self._dlen, 1):
label = labels
elif labels.shape == (self._dlen,):
label = labels[:, np.newaxis]
else:
raise LabelsError('wrong label format')
self._dlabel = label
def build_data_labels(self):
cc = ['dlabel-' + str(i) for i in range(0, self._dlen)]
return np.asarray(cc)[:, np.newaxis]
def calculate_map_dist(self):
"""
Calculates the grid distance, which will be used during the training steps.
It supports only planar grids for the moment
"""
nnodes = self.codebook.nnodes
distance_matrix = np.zeros((nnodes, nnodes))
for i in range(nnodes):
distance_matrix[i] = self.codebook.grid_dist(i).reshape(1, nnodes)
return distance_matrix
@timeit()
def train(self, n_job=1, shared_memory=False, verbose='info'):
"""
Trains the som
:param n_job: number of jobs to use to parallelize the traning
:param shared_memory: flag to active shared memory
:param verbose: verbosity, could be 'debug', 'info' or None
"""
logging.root.setLevel(getattr(logging, verbose.upper()) if verbose else logging.ERROR)
logging.info(" Training...")
logging.debug((
"--------------------------------------------------------------\n"
" details: \n"
" > data len is {data_len} and data dimension is {data_dim} \n"
" > map size is {mpsz0},{mpsz1}\n"
" > array size in log10 scale is {array_size}\n"
" > number of jobs in parallel: {n_job}\n"
" --------------------------------------------------------------\n")
.format(data_len=self._dlen,
data_dim=self._dim,
mpsz0=self.codebook.mapsize[0],
mpsz1=self.codebook.mapsize[1],
array_size=np.log10(self._dlen*self.codebook.nnodes*self._dim),
n_job=n_job))
if self.initialization == 'random':
self.codebook.random_initialization(self._data)
elif self.initialization == 'pca':
self.codebook.pca_linear_initialization(self._data)
self.rough_train(njob=n_job, shared_memory=shared_memory)
self.finetune_train(njob=n_job, shared_memory=shared_memory)
logging.debug(" --------------------------------------------------------------")
logging.info(" Final quantization error: %f" % np.mean(self._bmu[1]))
def _calculate_ms_and_mpd(self):
mn = np.min(self.codebook.mapsize)
max_s = max(self.codebook.mapsize[0], self.codebook.mapsize[1])
mpd = float(self.codebook.nnodes*10)/float(self._dlen) if mn == 1 else float(self.codebook.nnodes)/float(self._dlen)
ms = max_s/2.0 if mn == 1 else max_s
return ms, mpd
def rough_train(self, njob=1, shared_memory=False):
logging.info(" Rough training...")
ms, mpd = self._calculate_ms_and_mpd()
trainlen, radiusin, radiusfin = int(np.ceil(30*mpd)), None, None
if self.initialization == 'random':
radiusin = max(1, np.ceil(ms/3.))
radiusfin = max(1, radiusin/6.)
elif self.initialization == 'pca':
radiusin = max(1, np.ceil(ms/8.))
radiusfin = max(1, radiusin/4.)
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory)
def finetune_train(self, njob=1, shared_memory=False):
logging.info(" Finetune training...")
ms, mpd = self._calculate_ms_and_mpd()
trainlen, radiusin, radiusfin = None, None, None
if self.initialization == 'random':
trainlen = int(np.ceil(50*mpd))
radiusin = max(1, ms/12.) # from radius fin in rough training
radiusfin = max(1, radiusin/25.)
elif self.initialization == 'pca':
trainlen = int(np.ceil(40*mpd))
radiusin = max(1, np.ceil(ms/8.)/4)
radiusfin = 1 # max(1, ms/128)
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory)
def _batchtrain(self, trainlen, radiusin, radiusfin, njob=1, shared_memory=False):
radius = np.linspace(radiusin, radiusfin, trainlen)
if shared_memory:
data = self._data
data_folder = tempfile.mkdtemp()
data_name = os.path.join(data_folder, 'data')
dump(data, data_name)
data = load(data_name, mmap_mode='r')
else:
data = self._data
bmu = None
# X2 is part of euclidean distance (x-y)^2 = x^2 +y^2 - 2xy that we use for each data row in bmu finding.
# Since it is a fixed value we can skip it during bmu finding for each data point,
# but later we need it calculate quantification error
fixed_euclidean_x2 = np.einsum('ij,ij->i', data, data)
logging.info(" radius_ini: %f , radius_final: %f, trainlen: %d\n" % (radiusin, radiusfin, trainlen))
for i in range(trainlen):
t1 = time()
neighborhood = self.neighborhood.calculate(self._distance_matrix, radius[i], self.codebook.nnodes)
bmu = self.find_bmu(data, njb=njob)
self.codebook.matrix = self.update_codebook_voronoi(data, bmu, neighborhood)
qerror = (i+1, round(time() - t1, 3), np.mean(np.sqrt(bmu[1] + fixed_euclidean_x2)))
logging.info(" epoch: %d ---> elapsed time: %f, quantization error: %f\n" % qerror)
bmu[1] = np.sqrt(bmu[1] + fixed_euclidean_x2)
self._bmu = bmu
@timeit(logging.DEBUG)
def find_bmu(self, input_matrix, njb=1):
"""
Finds the best matching unit (bmu) for each input data from the input matrix. It does all at once parallelizing
the calculation instead of going through each input and running it against the codebook.
:param input_matrix: numpy matrix representing inputs as rows and features/dimension as cols
:param njb: number of jobs to parallelize the search
:returns: the best matching unit for each input
"""
dlen = input_matrix.shape[0]
y2 = np.einsum('ij,ij->i', self.codebook.matrix, self.codebook.matrix)
parallelizer = Parallel(n_jobs=njb, pre_dispatch='3*n_jobs')
chunk_bmu_finder = delayed(_chunk_based_bmu_find)
row_chunk = lambda part: part * dlen // njb
col_chunk = lambda part: min((part+1)*dlen // njb, dlen)
b = parallelizer(chunk_bmu_finder(input_matrix[row_chunk(i):col_chunk(i)], self.codebook.matrix, y2) for i in xrange(njb))
bmu = np.asarray(list(itertools.chain(*b))).T
del b
return bmu
@timeit(logging.DEBUG)
def update_codebook_voronoi(self, training_data, bmu, neighborhood):
"""
Updates the weights of each node in the codebook that belongs to the bmu's neighborhood.
First finds the Voronoi set of each node. It needs to calculate a smaller matrix.
Super fast comparing to classic batch training algorithm, it is based on the implemented algorithm in
som toolbox for Matlab by Helsinky university
:param training_data: input matrix with input vectors as rows and vector features as cols
:param bmu: best matching unit for each input data. Has shape of (2, dlen) where first row has bmu indexes
:param neighborhood: matrix representing the neighborhood of each bmu
:returns: An updated codebook that incorporates the learnings from the input data
"""
row = bmu[0].astype(int)
col = np.arange(self._dlen)
val = np.tile(1, self._dlen)
P = csr_matrix((val, (row, col)), shape=(self.codebook.nnodes, self._dlen))
S = P.dot(training_data)
# neighborhood has nnodes*nnodes and S has nnodes*dim ---> Nominator has nnodes*dim
nom = neighborhood.T.dot(S)
nV = P.sum(axis=1).reshape(1, self.codebook.nnodes)
denom = nV.dot(neighborhood.T).reshape(self.codebook.nnodes, 1)
new_codebook = np.divide(nom, denom)
return np.around(new_codebook, decimals=6)
def project_data(self, data):
"""
Projects a data set to a trained SOM. It is based on nearest neighborhood search module of scikitlearn,
but it is not that fast.
"""
clf = neighbors.KNeighborsClassifier(n_neighbors=1)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
data = self._normalizer.normalize_by(self.data_raw, data)
#data = normalize(data, method='var')
#plt.hist(data[:,2])
return clf.predict(data)
def predict_by(self, data, target, k=5, wt='distance'):
# here it is assumed that target is the last column in the codebook
# and data has dim-1 columns
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indX = ind[ind != target]
x = self.codebook.matrix[:, indX]
y = self.codebook.matrix[:, target]
n_neighbors = k
clf = neighbors.KNeighborsRegressor(n_neighbors, weights=wt)
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indX]
elif dimdata == dim-1:
data = self._normalizer.normalize_by(self.data_raw[:, indX], data)
predicted_values = clf.predict(data)
predicted_values = self._normalizer.denormalize_by(self.data_raw[:, target], predicted_values)
return predicted_values
def predict(self, x_test, k=5, wt='distance'):
"""
Similar to SKlearn we assume that we have X_tr, Y_tr and X_test. Here it is assumed that target is the last
column in the codebook and data has dim-1 columns
:param x_test: input vector
:param k: number of neighbors to use
:param wt: method to use for the weights (more detail in KNeighborsRegressor docs)
:returns: predicted values for the input data
"""
target = self.data_raw.shape[1]-1
x_train = self.codebook.matrix[:, :target]
y_train = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights=wt)
clf.fit(x_train, y_train)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
x_test = self._normalizer.normalize_by(self.data_raw[:, :target], x_test)
predicted_values = clf.predict(x_test)
return self._normalizer.denormalize_by(self.data_raw[:, target], predicted_values)
def find_k_nodes(self, data, k=5):
from sklearn.neighbors import NearestNeighbors
# we find the k most similar nodes to the input vector
neighbor = NearestNeighbors(n_neighbors=k)
neighbor.fit(self.codebook.matrix)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
return neighbor.kneighbors(self._normalizer.normalize_by(self.data_raw, data))
def bmu_ind_to_xy(self, bmu_ind):
"""
Translates a best matching unit index to the corresponding matrix x,y coordinates
:param bmu_ind: node index of the best matching unit (number of node from top left node)
:returns: corresponding (x,y) coordinate
"""
rows = self.codebook.mapsize[0]
cols = self.codebook.mapsize[1]
# bmu should be an integer between 0 to no_nodes
out = np.zeros((bmu_ind.shape[0], 3))
out[:, 2] = bmu_ind
out[:, 0] = rows-1-bmu_ind / cols
out[:, 0] = bmu_ind / cols
out[:, 1] = bmu_ind % cols
return out.astype(int)
def cluster(self, n_clusters=8):
import sklearn.cluster as clust
return clust.KMeans(n_clusters=n_clusters).fit_predict(self._normalizer.denormalize_by(self.data_raw,
self.codebook.matrix))
def predict_probability(self, data, target, k=5):
"""
Predicts probability of the input data to be target
:param data: data to predict, it is assumed that 'target' is the last column in the codebook,
so data hould have dim-1 columns
:param target: target to predict probability
:param k: k parameter on KNeighborsRegressor
:returns: probability of data been target
"""
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indx = ind[ind != target]
x = self.codebook.matrix[:, indx]
y = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights='distance')
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indx]
elif dimdata == dim-1:
data = self._normalizer.normalize_by(self.data_raw[:, indx], data)
weights, ind = clf.kneighbors(data, n_neighbors=k, return_distance=True)
weights = 1./weights
sum_ = np.sum(weights, axis=1)
weights = weights/sum_[:, np.newaxis]
labels = np.sign(self.codebook.matrix[ind, target])
labels[labels >= 0] = 1
# for positives
pos_prob = labels.copy()
pos_prob[pos_prob < 0] = 0
pos_prob *= weights
pos_prob = np.sum(pos_prob, axis=1)[:, np.newaxis]
# for negatives
neg_prob = labels.copy()
neg_prob[neg_prob > 0] = 0
neg_prob = neg_prob * weights * -1
neg_prob = np.sum(neg_prob, axis=1)[:, np.newaxis]
#predicted_values = clf.predict(data)
#predicted_values = denormalize_by(data_raw[:,Target], predicted_values)
return np.concatenate((pos_prob, neg_prob), axis=1)
def node_activation(self, data, target=None, wt='distance'):
weights, ind = None, None
if not target:
clf = neighbors.KNeighborsClassifier(n_neighbors=self.codebook.nnodes)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of original data
data = self._normalizer.normalize_by(self.data_raw, data)
weights, ind = clf.kneighbors(data)
# Softmax function
weights = 1./weights
#S_ = np.sum(np.exp(weights),axis=1)[:,np.newaxis]
#weights = np.exp(weights)/S_
return weights, ind
def codebook(self):
""" Export a text file code book of categories and codes.
"""
Codebook(self.app, self.ui.textEdit)
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy'):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
print "DDDD", data
self._data = data # normalizer.normalize(data) if normalizer else data
print "data : \n", self._data
print "map size : ", mapsize
self._normalizer = normalizer
print "_normalizer is a object to class normalizer : ", self._normalizer
self._dim = data.shape[1]
print "data shape[1] : ", self._dim
self._dlen = data.shape[0]
print "data shape[0] : ", data.shape[0]
self._dlabel = None
print "data label : ", self._dlabel
self._bmu = None
print "intial bmu : ", self._bmu
self.name = name
print "name : ", self.name
self.data_raw = data
print "data_raw : ", self.data_raw
#print "data_raw : ", self.data_raw
self.neighborhood = neighborhood
self.mapshape = mapshape
print "mapshape : ", self.mapshape
self.initialization = initialization
print "initilization : ", self.initialization
self.mask = mask or np.ones([1, self._dim])
print "mask : ", self.mask
self.codebook = Codebook(mapsize, lattice)
print "--------codebook is a object to class Codebook---------\n", self.codebook
self.training = training
print "training mode : ", self.training
self._component_names = self.build_component_names()
print "Initial component name : ", self._component_names
self._distance_matrix = self.calculate_map_dist()
print "initial distance matrix : ", self._distance_matrix
class SOM(object):
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
name='sompy',
component_names=None):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
self._data = normalizer.normalize(data) if normalizer else data
self._normalizer = normalizer
self._dim = data.shape[1]
self._dlen = data.shape[0]
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask or np.ones([1, self._dim])
mapsize = self.calculate_map_size(lattice) if not mapsize else mapsize
self.codebook = Codebook(mapsize, lattice)
self.training = training
self._component_names = self.build_component_names() if component_names is None else [component_names]
self._distance_matrix = self.calculate_map_dist()
@property
def component_names(self):
return self._component_names
@component_names.setter
def component_names(self, compnames):
if self._dim == len(compnames):
self._component_names = np.asarray(compnames)[np.newaxis, :]
else:
raise ComponentNamesError('Component names should have the same '
'size as the data dimension/features')
def build_component_names(self):
cc = ['Variable-' + str(i+1) for i in range(0, self._dim)]
return np.asarray(cc)[np.newaxis, :]
@property
def data_labels(self):
return self._dlabel
@data_labels.setter
def data_labels(self, labels):
"""
Set labels of the training data, it should be in the format of a list
of strings
"""
if labels.shape == (1, self._dlen):
label = labels.T
elif labels.shape == (self._dlen, 1):
label = labels
elif labels.shape == (self._dlen,):
label = labels[:, np.newaxis]
else:
raise LabelsError('wrong label format')
self._dlabel = label
def build_data_labels(self):
cc = ['dlabel-' + str(i) for i in range(0, self._dlen)]
return np.asarray(cc)[:, np.newaxis]
def calculate_map_dist(self):
"""
Calculates the grid distance, which will be used during the training
steps. It supports only planar grids for the moment
"""
nnodes = self.codebook.nnodes
distance_matrix = np.zeros((nnodes, nnodes))
for i in range(nnodes):
distance_matrix[i] = self.codebook.grid_dist(i).reshape(1, nnodes)
return distance_matrix
@timeit()
def train(self,
n_job=1,
shared_memory=False,
verbose='info',
train_rough_len=None,
train_rough_radiusin=None,
train_rough_radiusfin=None,
train_finetune_len=None,
train_finetune_radiusin=None,
train_finetune_radiusfin=None,
nth=1):
"""
Trains the som
:param n_job: number of jobs to use to parallelize the traning
:param shared_memory: flag to active shared memory
:param verbose: verbosity, could be 'debug', 'info' or None
"""
logging.root.setLevel(
getattr(logging, verbose.upper()) if verbose else logging.ERROR)
logging.info(" Training...")
logging.debug((
"--------------------------------------------------------------\n"
" details: \n"
" > data len is {data_len} and data dimension is {data_dim}\n"
" > map size is {mpsz0},{mpsz1}\n"
" > array size in log10 scale is {array_size}\n"
" > number of jobs in parallel: {n_job}\n"
" -------------------------------------------------------------\n")
.format(data_len=self._dlen,
data_dim=self._dim,
mpsz0=self.codebook.mapsize[0],
mpsz1=self.codebook.mapsize[1],
array_size=np.log10(
self._dlen * self.codebook.nnodes * self._dim),
n_job=n_job))
if self.initialization == 'random':
self.codebook.random_initialization(self._data)
elif self.initialization == 'pca':
self.codebook.pca_linear_initialization(self._data)
self.rough_train(njob=n_job, shared_memory=shared_memory, trainlen=train_rough_len,
radiusin=train_rough_radiusin, radiusfin=train_rough_radiusfin, nth=nth)
self.finetune_train(njob=n_job, shared_memory=shared_memory, trainlen=train_finetune_len,
radiusin=train_finetune_radiusin, radiusfin=train_finetune_radiusfin, nth=nth)
logging.debug(
" --------------------------------------------------------------")
logging.info(" Final quantization error: %f" % np.mean(self._bmu[1]))
def _calculate_ms_and_mpd(self):
mn = np.min(self.codebook.mapsize)
max_s = max(self.codebook.mapsize[0], self.codebook.mapsize[1])
if mn == 1:
mpd = float(self.codebook.nnodes*10)/float(self._dlen)
else:
mpd = float(self.codebook.nnodes)/float(self._dlen)
ms = max_s/2.0 if mn == 1 else max_s
return ms, mpd
def rough_train(self, njob=1, shared_memory=False, trainlen=None, radiusin=None, radiusfin=None,nth=1):
logging.info(" Rough training...")
ms, mpd = self._calculate_ms_and_mpd()
trainlen = int(np.ceil(30*mpd)) if not trainlen else trainlen
if self.initialization == 'random':
radiusin = max(1, np.ceil(ms/3.)) if not radiusin else radiusin
radiusfin = max(1, radiusin/6.) if not radiusfin else radiusfin
elif self.initialization == 'pca':
radiusin = max(1, np.ceil(ms/8.)) if not radiusin else radiusin
radiusfin = max(1, radiusin/4.) if not radiusfin else radiusfin
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory, nth)
def finetune_train(self, njob=1, shared_memory=False, trainlen=None, radiusin=None, radiusfin=None,nth=1):
logging.info(" Finetune training...")
ms, mpd = self._calculate_ms_and_mpd()
if self.initialization == 'random':
trainlen = int(np.ceil(50*mpd)) if not trainlen else trainlen
radiusin = max(1, ms/12.) if not radiusin else radiusin # from radius fin in rough training
radiusfin = max(1, radiusin/25.) if not radiusfin else radiusfin
elif self.initialization == 'pca':
trainlen = int(np.ceil(40*mpd)) if not trainlen else trainlen
radiusin = max(1, np.ceil(ms/8.)/4) if not radiusin else radiusin
radiusfin = 1 if not radiusfin else radiusfin # max(1, ms/128)
self._batchtrain(trainlen, radiusin, radiusfin, njob, shared_memory, nth)
def _batchtrain(self, trainlen, radiusin, radiusfin, njob=1,
shared_memory=False, nth=1):
radius = np.linspace(radiusin, radiusfin, trainlen)
if shared_memory:
data = self._data
data_folder = tempfile.mkdtemp()
data_name = os.path.join(data_folder, 'data')
dump(data, data_name)
data = load(data_name, mmap_mode='r')
else:
data = self._data
bmu = None
# X2 is part of euclidean distance (x-y)^2 = x^2 +y^2 - 2xy that we use
# for each data row in bmu finding.
# Since it is a fixed value we can skip it during bmu finding for each
# data point, but later we need it calculate quantification error
fixed_euclidean_x2 = np.einsum('ij,ij->i', data, data)
logging.info(" radius_ini: %f , radius_final: %f, trainlen: %d\n" %
(radiusin, radiusfin, trainlen))
for i in range(trainlen):
t1 = time()
neighborhood = self.neighborhood.calculate(
self._distance_matrix, radius[i], self.codebook.nnodes)
bmu = self.find_bmu(data, njb=njob, nth=nth)
self.codebook.matrix = self.update_codebook_voronoi(data, bmu,
neighborhood)
qerror = (i + 1, round(time() - t1, 3),
np.mean(np.sqrt(bmu[1] + fixed_euclidean_x2)))
logging.info(
" epoch: %d ---> elapsed time: %f, quantization error: %f\n" %
qerror)
bmu[1] = np.sqrt(bmu[1] + fixed_euclidean_x2)
self._bmu = bmu
@timeit(logging.DEBUG)
def find_bmu(self, input_matrix, njb=1, nth=1):
"""
Finds the best matching unit (bmu) for each input data from the input
matrix. It does all at once parallelizing the calculation instead of
going through each input and running it against the codebook.
:param input_matrix: numpy matrix representing inputs as rows and
features/dimension as cols
:param njb: number of jobs to parallelize the search
:returns: the best matching unit for each input
"""
dlen = input_matrix.shape[0]
y2 = np.einsum('ij,ij->i', self.codebook.matrix, self.codebook.matrix)
if njb == -1:
njb = cpu_count()
pool = Pool(njb)
chunk_bmu_finder = _chunk_based_bmu_find
def row_chunk(part):
return part * dlen // njb
def col_chunk(part):
return min((part+1)*dlen // njb, dlen)
chunks = [input_matrix[row_chunk(i):col_chunk(i)] for i in range(njb)]
b = pool.map(lambda chk: chunk_bmu_finder(chk, self.codebook.matrix, y2, nth=nth), chunks)
pool.close()
pool.join()
bmu = np.asarray(list(itertools.chain(*b))).T
del b
return bmu
@timeit(logging.DEBUG)
def update_codebook_voronoi(self, training_data, bmu, neighborhood):
"""
Updates the weights of each node in the codebook that belongs to the
bmu's neighborhood.
First finds the Voronoi set of each node. It needs to calculate a
smaller matrix.
Super fast comparing to classic batch training algorithm, it is based
on the implemented algorithm in som toolbox for Matlab by Helsinky
University.
:param training_data: input matrix with input vectors as rows and
vector features as cols
:param bmu: best matching unit for each input data. Has shape of
(2, dlen) where first row has bmu indexes
:param neighborhood: matrix representing the neighborhood of each bmu
:returns: An updated codebook that incorporates the learnings from the
input data
"""
row = bmu[0].astype(int)
col = np.arange(self._dlen)
val = np.tile(1, self._dlen)
P = csr_matrix((val, (row, col)), shape=(self.codebook.nnodes,
self._dlen))
S = P.dot(training_data)
# neighborhood has nnodes*nnodes and S has nnodes*dim
# ---> Nominator has nnodes*dim
nom = neighborhood.T.dot(S)
nV = P.sum(axis=1).reshape(1, self.codebook.nnodes)
denom = nV.dot(neighborhood.T).reshape(self.codebook.nnodes, 1)
new_codebook = np.divide(nom, denom)
return np.around(new_codebook, decimals=6)
def project_data(self, data):
"""
Projects a data set to a trained SOM. It is based on nearest
neighborhood search module of scikitlearn, but it is not that fast.
"""
clf = neighbors.KNeighborsClassifier(n_neighbors=1)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
data = self._normalizer.normalize_by(self.data_raw, data)
return clf.predict(data)
def predict_by(self, data, target, k=5, wt='distance'):
# here it is assumed that target is the last column in the codebook
# and data has dim-1 columns
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indX = ind[ind != target]
x = self.codebook.matrix[:, indX]
y = self.codebook.matrix[:, target]
n_neighbors = k
clf = neighbors.KNeighborsRegressor(n_neighbors, weights=wt)
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indX]
elif dimdata == dim-1:
data = self._normalizer.normalize_by(self.data_raw[:, indX], data)
predicted_values = clf.predict(data)
predicted_values = self._normalizer.denormalize_by(
self.data_raw[:, target], predicted_values)
return predicted_values
def predict(self, x_test, k=5, wt='distance'):
"""
Similar to SKlearn we assume that we have X_tr, Y_tr and X_test. Here
it is assumed that target is the last column in the codebook and data
has dim-1 columns
:param x_test: input vector
:param k: number of neighbors to use
:param wt: method to use for the weights
(more detail in KNeighborsRegressor docs)
:returns: predicted values for the input data
"""
target = self.data_raw.shape[1]-1
x_train = self.codebook.matrix[:, :target]
y_train = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights=wt)
clf.fit(x_train, y_train)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
x_test = self._normalizer.normalize_by(
self.data_raw[:, :target], x_test)
predicted_values = clf.predict(x_test)
return self._normalizer.denormalize_by(
self.data_raw[:, target], predicted_values)
def find_k_nodes(self, data, k=5):
from sklearn.neighbors import NearestNeighbors
# we find the k most similar nodes to the input vector
neighbor = NearestNeighbors(n_neighbors=k)
neighbor.fit(self.codebook.matrix)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
return neighbor.kneighbors(
self._normalizer.normalize_by(self.data_raw, data))
def bmu_ind_to_xy(self, bmu_ind):
"""
Translates a best matching unit index to the corresponding
matrix x,y coordinates.
:param bmu_ind: node index of the best matching unit
(number of node from top left node)
:returns: corresponding (x,y) coordinate
"""
rows = self.codebook.mapsize[0]
cols = self.codebook.mapsize[1]
# bmu should be an integer between 0 to no_nodes
out = np.zeros((bmu_ind.shape[0], 3))
out[:, 2] = bmu_ind
out[:, 0] = rows-1-bmu_ind / cols
out[:, 0] = bmu_ind / cols
out[:, 1] = bmu_ind % cols
return out.astype(int)
def cluster(self, n_clusters=8):
import sklearn.cluster as clust
cl_labels = clust.KMeans(n_clusters=n_clusters).fit_predict(
self._normalizer.denormalize_by(self.data_raw,
self.codebook.matrix))
self.cluster_labels = cl_labels
return cl_labels
def predict_probability(self, data, target, k=5):
"""
Predicts probability of the input data to be target
:param data: data to predict, it is assumed that 'target' is the last
column in the codebook, so data hould have dim-1 columns
:param target: target to predict probability
:param k: k parameter on KNeighborsRegressor
:returns: probability of data been target
"""
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indx = ind[ind != target]
x = self.codebook.matrix[:, indx]
y = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights='distance')
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indx]
elif dimdata == dim-1:
data = self._normalizer.normalize_by(self.data_raw[:, indx], data)
weights, ind = clf.kneighbors(data, n_neighbors=k,
return_distance=True)
weights = 1./weights
sum_ = np.sum(weights, axis=1)
weights = weights/sum_[:, np.newaxis]
labels = np.sign(self.codebook.matrix[ind, target])
labels[labels >= 0] = 1
# for positives
pos_prob = labels.copy()
pos_prob[pos_prob < 0] = 0
pos_prob *= weights
pos_prob = np.sum(pos_prob, axis=1)[:, np.newaxis]
# for negatives
neg_prob = labels.copy()
neg_prob[neg_prob > 0] = 0
neg_prob = neg_prob * weights * -1
neg_prob = np.sum(neg_prob, axis=1)[:, np.newaxis]
return np.concatenate((pos_prob, neg_prob), axis=1)
def node_activation(self, data, target=None, wt='distance'):
weights, ind = None, None
if not target:
clf = neighbors.KNeighborsClassifier(
n_neighbors=self.codebook.nnodes)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
data = self._normalizer.normalize_by(self.data_raw, data)
weights, ind = clf.kneighbors(data)
# Softmax function
weights = 1./weights
return weights, ind
def calculate_topographic_error(self):
bmus1 = self.find_bmu(self.data_raw, njb=1, nth=1)
bmus2 = self.find_bmu(self.data_raw, njb=1, nth=2)
bmus_gap = np.abs((self.bmu_ind_to_xy(np.array(bmus1[0]))[:, 0:2] - self.bmu_ind_to_xy(np.array(bmus2[0]))[:, 0:2]).sum(axis=1))
return np.mean(bmus_gap != 1)
def calculate_map_size(self, lattice):
"""
Calculates the optimal map size given a dataset using eigenvalues and eigenvectors. Matlab ported
:lattice: 'rect' or 'hex'
:return: map sizes
"""
D = self.data_raw.copy()
dlen = D.shape[0]
dim = D.shape[1]
munits = np.ceil(5 * (dlen ** 0.5))
A = np.ndarray(shape=[dim, dim]) + np.Inf
for i in range(dim):
D[:, i] = D[:, i] - np.mean(D[np.isfinite(D[:, i]), i])
for i in range(dim):
for j in range(dim):
c = D[:, i] * D[:, j]
c = c[np.isfinite(c)]
A[i, j] = sum(c) / len(c)
A[j, i] = A[i, j]
VS = np.linalg.eig(A)
eigval = sorted(np.linalg.eig(A)[0])
if eigval[-1] == 0 or eigval[-2] * munits < eigval[-1]:
ratio = 1
else:
ratio = np.sqrt(eigval[-1] / eigval[-2])
if lattice == "rect":
size1 = min(munits, round(np.sqrt(munits / ratio)))
else:
size1 = min(munits, round(np.sqrt(munits / ratio*np.sqrt(0.75))))
size2 = round(munits / size1)
return [int(size1), int(size2)]
from codebook import Codebook codebooks = Codebook.load_codebooks() Codebook.test_codebooks(codebooks)
from codebook import Codebook codebooks = Codebook.train_codebooks() Codebook.save_codebooks(codebooks)
def build_codebook(encoder, dataset, args):
embed_bb = args.getboolean('Embedding', 'EMBED_BB')
codebook = Codebook(encoder, dataset, embed_bb)
return codebook
class SOMMap(object):
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
radius_train='linear',
name='sompy',
component_names=None,
components_to_plot=None,
isNormalized=False):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
if data is not None: # ajout LB
if normalizer and isNormalized == False:
for i in range(len(normalizer)):
data[:, i] = normalizer[i].normalize(data[:, i])
self._data = data
else:
self._data = data
self._data = self._data.astype('double') # ajout LB
self._dim = data.shape[1]
self._dlen = data.shape[0]
else:
self._data = None
self._dim = None
self._dlen = None
self._normalizer = normalizer
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask
mapsize = self.calculate_map_size(lattice) if not mapsize else mapsize
self.mapsize = mapsize
self.codebook = Codebook(mapsize, lattice)
self.training = training
self.radius_train = radius_train
self._component_names = self.build_component_names(
) if component_names is None else [component_names]
self._distance_matrix = self.calculate_map_dist()
self.components_to_plot = components_to_plot
def __str__(self):
return f'mapsize={self.mapsize}\nname={self.name}\nNormaliser={self._normalizer.name}\nMap shape={self.mapshape}'
def attach_data(self, data, comp_names): # ajout LB
self.data_raw = data
self._dim = data.shape[1]
self._dlen = data.shape[0]
self.component_names = comp_names
self._data = self._normalizer.normalize(data)
self._data = self._data.astype('double')
self._bmu = self.find_bmu(data, njb=1)
def save(self, file):
dico = {
'name': self.name,
'codebook': self.codebook.matrix,
'lattice': self.codebook.lattice,
'mapsize': self.codebook.mapsize,
'normalization': self._normalizer,
'norm_params': self._normalizer.params,
'comp_names': self._component_names[0],
'mask': self.mask,
'neighborhood': self.neighborhood.name,
'codebookinitialized': self.codebook.initialized,
'initialization': self.initialization,
'bmu': self._bmu,
'mapshape': self.mapshape,
'training': self.training,
'radius_train': self.radius_train,
'dim': self._dim
}
import pickle
pickle.dump(dico, open(file, 'wb'))
#def plotepoch(self,comp0,comp1):
def plotplanes(self): # ajout LB
if self.components_to_plot is None:
return
comps = self.components_to_plot
n = {1: (1, 1), 2: (1, 2), 3: (2, 2), 4: (2, 2)} # lignes, colonnes
nplots = len(comps)
if nplots > 4:
raise ValueError(
'Le nombre de comp doit etre inferieur ou égal à 4')
nl = n[nplots][0]
nc = n[nplots][1]
neighbours_nodes = self.calculate_neighbours_nodes()
edges = []
for i in range(self.codebook.nnodes):
for j in range(len(neighbours_nodes[i])):
edges.append((i, neighbours_nodes[i][j]))
nodes = [i for i in range(self.codebook.nnodes)]
G = nx.Graph()
G.add_nodes_from(nodes)
G.add_edges_from(edges)
plt.clf()
for i in range(nplots):
refs = [self.codebook.matrix[:, comps[i][0]]
], [self.codebook.matrix[:, comps[i][1]]]
pos = {}
for k in range(self.codebook.nnodes):
name = int(k)
pos[name] = (refs[0][0][k], refs[1][0][k])
plt.subplot(nl, nc, i + 1)
plt.scatter(self._data[:, comps[i][0]],
self._data[:, comps[i][1]],
marker='x',
c='b')
#plt.scatter([self.codebook.matrix[:,comps[i][0]]],[self.codebook.matrix[:,comps[i][1]]],marker='o',c='r')
nx.draw_networkx(G,
pos,
arrows=False,
with_labels=False,
node_size=50)
plt.xlabel(self._component_names[0][comps[i][0]])
plt.ylabel(self._component_names[0][comps[i][1]])
plt.title('comp. {} - {}'.format(comps[i][0], comps[i][1]))
plt.pause(0.1)
def plot_tsne(self, init="pca", perplexity=10, verbose=2):
T_SNE = TSNE(2, init=init, perplexity=perplexity, verbose=verbose)
x2d = T_SNE.fit_transform(self.codebook.matrix)
neighbours_nodes = self.calculate_neighbours_nodes()
edges = []
for i in range(self.codebook.nnodes):
for j in range(len(neighbours_nodes[i])):
edges.append((i, neighbours_nodes[i][j]))
nodes = [i for i in range(self.codebook.nnodes)]
G = nx.Graph()
G.add_nodes_from(nodes)
for i in range(self.codebook.nnodes):
G.nodes[i]['label'] = self._nlabel[i]
G.add_edges_from(edges)
pos = {}
for i in range(self.codebook.nnodes):
name = int(i)
pos[name] = (x2d[i][0], x2d[i][1])
colorlist = []
for i in range(len(np.unique(self._nlabel))):
colorlist.append('#%06X' % randint(0, 0xFFFFFF))
plt.figure()
for i in range(len(np.unique(self._nlabel))):
nodelist = []
for j in range(self.codebook.nnodes):
if G.nodes[j]['label'] == np.unique(self._nlabel)[i]:
nodelist.append(j)
nx.draw_networkx(G,
pos,
arrows=False,
nodelist=nodelist,
node_color=colorlist[i],
with_labels=False,
node_size=50,
label=np.unique(self._nlabel)[i])
plt.xlabel('tsne-2d-one')
plt.ylabel('tsne-2d-two')
plt.title('T-SNE des neurones référants')
plt.show()
@property
def component_names(self):
return self._component_names
@component_names.setter
def component_names(self, compnames):
if self._dim == len(compnames):
self._component_names = np.asarray(compnames)[np.newaxis, :]
else:
raise ComponentNamesError('Component names should have the same '
'size as the data dimension/features')
def build_component_names(self):
return ['Variable-' + str(i + 1) for i in range(0, self._dim)]
@property
def node_labels(self):
return self._nlabel
@node_labels.setter
def node_labels(self, labels):
"""
Set labels of the training data, it should be in the format of a list
of strings
"""
if labels.shape == (1, self.codebook.nnodes):
label = labels.T
elif labels.shape == (self.codebook.nnodes, 1):
label = labels
elif labels.shape == (self.codebook.nnodes, ):
label = labels[:, np.newaxis]
else:
raise LabelsError('wrong label format')
self._nlabel = label
def build_node_labels(self):
return ['nlabel-' + str(i) for i in range(0, self.codebook.nnodes)]
def node_labels_from_data(self, sData):
nlabels = []
for i in range(self.codebook.nnodes):
ind = get_index_positions(self._bmu[0], i)
if ind != []:
subData = [sData._dlabel[k] for k in ind]
nlabels.append(Counter(subData).most_common(1)[0][0])
else:
nlabels.append("Nan")
self._nlabel = nlabels
def calculate_map_dist(self): # CALCUL MATRICE DIST
"""
Calculates the grid distance, which will be used during the training
steps. It supports only planar grids for the moment
"""
nnodes = self.codebook.nnodes
distance_matrix = np.zeros((nnodes, nnodes))
for i in range(nnodes):
#distance_matrix[i] = self.codebook.grid_dist(i).reshape(1, nnodes)
distance_matrix[i] = self.codebook.grid_dist(i).T.reshape(
1, nnodes) #attention c'et la distance au carré
return distance_matrix
@timeit()
def train(self,
n_job=1,
shared_memory=False,
verbose='info',
train_rough_len=None,
train_rough_radiusin=None,
train_rough_radiusfin=None,
train_finetune_len=None,
train_finetune_radiusin=None,
train_finetune_radiusfin=None,
train_len_factor=1,
maxtrainlen=np.Inf,
alreadyinit=False,
watch_evolution=True):
"""
Trains the som
:param n_job: number of jobs to use to parallelize the traning
:param shared_memory: flag to active shared memory
:param verbose: verbosity, could be 'debug', 'info' or None
:param train_len_factor: Factor that multiply default training lenghts (similar to "training" parameter in the matlab version). (lbugnon)
"""
logging.root.setLevel(
getattr(logging, verbose.upper()) if verbose else logging.ERROR)
logging.info(" Training...")
print('Training ...')
logging.debug((
"--------------------------------------------------------------\n"
" details: \n"
" > data len is {data_len} and data dimension is {data_dim}\n"
" > map size is {mpsz0},{mpsz1}\n"
" > array size in log10 scale is {array_size}\n"
" > number of jobs in parallel: {n_job}\n"
" -------------------------------------------------------------\n"
).format(data_len=self._dlen,
data_dim=self._dim,
mpsz0=self.codebook.mapsize[0],
mpsz1=self.codebook.mapsize[1],
array_size=np.log10(self._dlen * self.codebook.nnodes *
self._dim),
n_job=n_job))
if self.initialization == 'random' and alreadyinit == False:
self.codebook.random_initialization(self._data)
elif self.initialization == 'custom' and alreadyinit == False:
self.codebook.custom_initialization(self._data)
elif self.initialization == 'pca' and alreadyinit == False:
self.codebook.pca_linear_initialization(self._data, self.mask)
elif alreadyinit == False:
raise AttributeError('initialisation inconnue')
if train_rough_len > 0:
self.rough_train(njob=n_job,
shared_memory=shared_memory,
trainlen=train_rough_len,
radiusin=train_rough_radiusin,
radiusfin=train_rough_radiusfin,
trainlen_factor=train_len_factor,
maxtrainlen=maxtrainlen,
watch_evolution=watch_evolution)
if train_finetune_len > 0:
self.finetune_train(njob=n_job,
shared_memory=shared_memory,
trainlen=train_finetune_len,
radiusin=train_finetune_radiusin,
radiusfin=train_finetune_radiusfin,
trainlen_factor=train_len_factor,
maxtrainlen=maxtrainlen,
watch_evolution=watch_evolution)
if self._bmu is None: # calcul bmu meme si pas entrainé
self._bmu = self.find_bmu(self._data, njb=n_job)
#self.plotplanes2()
logging.debug(
" --------------------------------------------------------------")
logging.info(" Final quantization error: %f" % np.mean(self._bmu[1]))
def _calculate_ms_and_mpd(self):
mn = np.min(self.codebook.mapsize)
max_s = max(self.codebook.mapsize[0], self.codebook.mapsize[1])
if mn == 1:
mpd = float(self.codebook.nnodes * 10) / float(self._dlen)
else:
mpd = float(self.codebook.nnodes) / float(self._dlen)
ms = max_s / 2.0 if mn == 1 else max_s
return ms, mpd
def rough_train(self,
njob=1,
shared_memory=False,
trainlen=None,
radiusin=None,
radiusfin=None,
trainlen_factor=1,
maxtrainlen=np.Inf,
watch_evolution=True):
logging.info(" Rough training...")
print(" Rough training...")
ms, mpd = self._calculate_ms_and_mpd()
#lbugnon: add maxtrainlen
trainlen = min(int(np.ceil(30 * mpd)),
maxtrainlen) if not trainlen else trainlen
#print("maxtrainlen %d",maxtrainlen)
#lbugnon: add trainlen_factor
trainlen = int(trainlen * trainlen_factor)
if self.initialization == 'random':
radiusin = max(1, np.ceil(ms / 3.)) if not radiusin else radiusin
radiusfin = max(1, radiusin / 6.) if not radiusfin else radiusfin
elif self.initialization == 'pca':
radiusin = max(1, np.ceil(ms / 8.)) if not radiusin else radiusin
radiusfin = max(1, radiusin / 4.) if not radiusfin else radiusfin
self._batchtrain(trainlen,
radiusin,
radiusfin,
njob,
shared_memory,
watch_evolution=watch_evolution)
def finetune_train(self,
njob=1,
shared_memory=False,
trainlen=None,
radiusin=None,
radiusfin=None,
trainlen_factor=1,
maxtrainlen=np.Inf,
watch_evolution=True):
logging.info(" Finetune training...")
print('Finetune training')
ms, mpd = self._calculate_ms_and_mpd()
#lbugnon: add maxtrainlen
if self.initialization == 'random':
trainlen = min(int(np.ceil(50 * mpd)),
maxtrainlen) if not trainlen else trainlen
radiusin = max(
1, ms / 12.
) if not radiusin else radiusin # from radius fin in rough training
radiusfin = max(1, radiusin / 25.) if not radiusfin else radiusfin
elif self.initialization == 'pca':
trainlen = min(int(np.ceil(40 * mpd)),
maxtrainlen) if not trainlen else trainlen
radiusin = max(1,
np.ceil(ms / 8.) / 4) if not radiusin else radiusin
radiusfin = 1 if not radiusfin else radiusfin # max(1, ms/128)
#print("maxtrainlen %d",maxtrainlen)
#lbugnon: add trainlen_factor
trainlen = int(trainlen_factor * trainlen)
self._batchtrain(trainlen,
radiusin,
radiusfin,
njob,
shared_memory,
watch_evolution=watch_evolution)
def _batchtrain(self,
trainlen,
radiusin,
radiusfin,
njob=1,
shared_memory=False,
watch_evolution=True):
if self.radius_train == 'linear':
radius = np.linspace(radiusin, radiusfin, trainlen)
elif self.radius_train == 'power_series':
radius = []
ratio = radiusfin / radiusin
for i in range(trainlen):
radius.append(radiusin * ((ratio)**(i / trainlen)))
elif self.radius_train == 'inverse_of_time':
radius = []
B = trainlen / ((radiusin / radiusfin) - 1)
A = B * radiusin
for i in range(trainlen):
radius.append(A / (i + B))
else:
raise AttributeError('évolution du radius inconnue')
if shared_memory:
data = self._data
data_folder = tempfile.mkdtemp()
data_name = os.path.join(data_folder, 'data')
dump(data, data_name)
data = load(data_name, mmap_mode='r')
else:
data = self._data
bmu = None
# X2 is part of euclidean distance (x-y)^2 = x^2 +y^2 - 2xy that we use
# for each data row in bmu finding.
# Since it is a fixed value we can skip it during bmu finding for each
# data point, but later we need it calculate quantification error
logging.info(" radius_ini: %f , radius_final: %f, trainlen: %d\n" %
(radiusin, radiusfin, trainlen))
print(
"radius_ini: {:.3f} , radius_final: {:.3f}, trainlen: {}\n".format(
radiusin, radiusfin, trainlen))
for i in range(trainlen):
t1 = time()
neighborhood = self.neighborhood.calculate(self._distance_matrix,
radius[i],
self.codebook.nnodes)
bmu = self.find_bmu(data, njb=njob)
self.codebook.matrix = self.update_codebook_voronoi(
data, bmu, neighborhood)
#print('nombre de neurones activés : ',len(np.unique(bmu[0])))
qerror = self.calculate_quantization_error()
terror = self.calculate_topographic_error()
print('Epoch : {} qErr : {:.4f} tErr : {:.4f}'.format(
i, qerror, terror))
logging.info(
" epoch: {} ---> elapsed time: {:2.2f}, quantization error: {:2.4f}"
.format(i,
time() - t1, qerror))
if np.any(np.isnan(qerror)):
logging.info("nan quantization error, exit train\n")
#sys.exit("quantization error=nan, exit train")
if i % 1 == 0 and watch_evolution == True:
self.plotplanes() # ajout LB
if watch_evolution == False:
self.plotplanes()
#print('bmu = {}'.format(bmu[1] + fixed_euclidean_x2))
bmu = self.find_bmu(data,
njb=njob) # ajout LB : il faut remettre à jour
#tmp= bmu[1] + fixed_euclidean_x2
#tmp[tmp<0]=0 # ajout LB
#bmu[1] = np.sqrt(bmu[1] + fixed_euclidean_x2)
#bmu[1]=tmp
self._bmu = bmu
@timeit(logging.DEBUG)
def find_bmu(self, input_matrix, njb=1, nth=1):
"""
Finds the best matching unit (bmu) for each input data from the input
matrix. It does all at once parallelizing the calculation instead of
going through each input and running it against the codebook.
:param input_matrix: numpy matrix representing inputs as rows and
features/dimension as cols
:param njb: number of jobs to parallelize the search
:returns: the best matching unit for each input
"""
dlen = input_matrix.shape[0]
if self.mask is not None: # ajout LB
codebookmask = self.codebook.matrix * self.mask.squeeze()
datamask = input_matrix * self.mask
else:
codebookmask = self.codebook.matrix
datamask = input_matrix
y2 = np.einsum('ij,ij->i', self.codebook.matrix,
codebookmask) # somme des carrés Y**2
x2 = np.einsum('ij,ij->i', datamask, input_matrix)
if njb == -1:
njb = cpu_count()
pool = Pool(njb)
chunk_bmu_finder = _chunk_based_bmu_find
def row_chunk(part):
return part * dlen // njb
def col_chunk(part):
return min((part + 1) * dlen // njb, dlen)
chunks = [input_matrix[row_chunk(i):col_chunk(i)] for i in range(njb)]
#chunk_bmu_finder(input_matrix, self.codebook.matrix, y2, x2,nth=nth,mask=self.mask)
b = pool.map(lambda chk: chunk_bmu_finder(
chk, self.codebook.matrix, y2, x2, nth=nth, mask=self.mask),
chunks) # modif LB
pool.close()
pool.join()
bmu = np.asarray(list(itertools.chain(*b))).T
del b
return bmu
@timeit(logging.DEBUG)
def update_codebook_voronoi(self, training_data, bmu, neighborhood):
"""
Updates the weights of each node in the codebook that belongs to the
bmu's neighborhood.
First finds the Voronoi set of each node. It needs to calculate a
smaller matrix.
Super fast comparing to classic batch training algorithm, it is based
on the implemented algorithm in som toolbox for Matlab by Helsinky
University.
:param training_data: input matrix with input vectors as rows and
vector features as cols
:param bmu: best matching unit for each input data. Has shape of
(2, dlen) where first row has bmu indexes
:param neighborhood: matrix representing the neighborhood of each bmu
:returns: An updated codebook that incorporates the learnings from the
input data
"""
row = bmu[0].astype(
int) # pour chape individu le numero du codebook associé
col = np.arange(self._dlen)
val = np.tile(1, self._dlen) # autant de 1 que d'individus
P = csr_matrix(
(val, (row, col)), shape=(self.codebook.nnodes, self._dlen)
) # nbr codebook x nbr individus avec des 1 lorsque un individu voisinage du codebook, 0 sinon
S = P.dot(
training_data
) # nbr codebook x dim codebook : formule 5 page 11 de somtoolbox 5 matlab
# neighborhood has nnodes*nnodes and S has nnodes*dim
# ---> Nominator has nnodes*dim
nom = neighborhood.T.dot(S)
nV = P.sum(axis=1).reshape(1, self.codebook.nnodes)
denom = nV.dot(neighborhood.T).reshape(self.codebook.nnodes,
1) # role de la transposée ???
new_codebook = np.divide(nom, denom)
if (denom == 0.0).sum() > 0:
print('denominateur nul', denom)
raise
#return np.around(new_codebook, decimals=6)
return np.asarray(new_codebook) # modif LB
def project_data(self, data, normalize=False): # modif LB
"""
Projects a data set to a trained SOM. It is based on nearest
neighborhood search module of scikitlearn, but it is not that fast.
"""
clf = neighbors.KNeighborsClassifier(n_neighbors=1)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
if normalize:
data = self._normalizer.normalize_by(self.data_raw, data)
return clf.predict(data)
def predict_by(self, data, target, k=5, wt='distance'):
# here it is assumed that target is the last column in the codebook
# and data has dim-1 columns
print('fonction predict_by est elle utilisée ?')
raise
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indX = ind[ind != target]
x = self.codebook.matrix[:, indX]
y = self.codebook.matrix[:, target]
n_neighbors = k
clf = neighbors.KNeighborsRegressor(n_neighbors, weights=wt)
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indX]
elif dimdata == dim - 1:
data = self._normalizer.normalize_by(self.data_raw[:, indX], data)
predicted_values = clf.predict(data)
predicted_values = self._normalizer.denormalize_by(
self.data_raw[:, target], predicted_values)
return predicted_values
def predict(self, x_test, k=5, wt='distance'):
"""
Similar to SKlearn we assume that we have X_tr, Y_tr and X_test. Here
it is assumed that target is the last column in the codebook and data
has dim-1 columns
:param x_test: input vector
:param k: number of neighbors to use
:param wt: method to use for the weights
(more detail in KNeighborsRegressor docs)
:returns: predicted values for the input data
"""
target = self.data_raw.shape[1] - 1
x_train = self.codebook.matrix[:, :target]
y_train = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights=wt)
clf.fit(x_train, y_train)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
x_test = self._normalizer.normalize_by(self.data_raw[:, :target],
x_test)
predicted_values = clf.predict(x_test)
return self._normalizer.denormalize_by(self.data_raw[:, target],
predicted_values)
def find_k_nodes(self, data, k=5):
from sklearn.neighbors import NearestNeighbors
# we find the k most similar nodes to the input vector
neighbor = NearestNeighbors(n_neighbors=k)
neighbor.fit(self.codebook.matrix)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
return neighbor.kneighbors(
self._normalizer.normalize_by(self.data_raw, data))
def bmu_ind_to_xy(self, bmu_ind):
"""
Translates a best matching unit index to the corresponding
matrix x,y coordinates.
:param bmu_ind: node index of the best matching unit
(number of node from top left node)
:returns: corresponding (x,y) coordinate
"""
rows = self.codebook.mapsize[0]
cols = self.codebook.mapsize[1]
# bmu should be an integer between 0 to no_nodes
out = np.zeros((bmu_ind.shape[0], 3))
out[:, 2] = bmu_ind
out[:, 0] = rows - 1 - bmu_ind / cols
out[:, 0] = bmu_ind / cols
out[:, 1] = bmu_ind % cols
return out.astype(int)
def cluster(self, n_clusters=8):
import sklearn.cluster as clust
#cl_labels = clust.KMeans(n_clusters=n_clusters).fit_predict(
# self._normalizer.denormalize_by(self.data_raw,
# self.codebook.matrix))
cl_labels = clust.KMeans(n_clusters=n_clusters).fit_predict(
self.codebook.matrix)
#print(cl_labels)
self.cluster_labels = cl_labels
return cl_labels
def predict_probability(self, data, target, k=5):
"""
Predicts probability of the input data to be target
:param data: data to predict, it is assumed that 'target' is the last
column in the codebook, so data hould have dim-1 columns
:param target: target to predict probability
:param k: k parameter on KNeighborsRegressor
:returns: probability of data been target
"""
dim = self.codebook.matrix.shape[1]
ind = np.arange(0, dim)
indx = ind[ind != target]
x = self.codebook.matrix[:, indx]
y = self.codebook.matrix[:, target]
clf = neighbors.KNeighborsRegressor(k, weights='distance')
clf.fit(x, y)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
dimdata = data.shape[1]
if dimdata == dim:
data[:, target] = 0
data = self._normalizer.normalize_by(self.data_raw, data)
data = data[:, indx]
elif dimdata == dim - 1:
data = self._normalizer.normalize_by(self.data_raw[:, indx], data)
weights, ind = clf.kneighbors(data,
n_neighbors=k,
return_distance=True)
weights = 1. / weights
sum_ = np.sum(weights, axis=1)
weights = weights / sum_[:, np.newaxis]
labels = np.sign(self.codebook.matrix[ind, target])
labels[labels >= 0] = 1
# for positives
pos_prob = labels.copy()
pos_prob[pos_prob < 0] = 0
pos_prob *= weights
pos_prob = np.sum(pos_prob, axis=1)[:, np.newaxis]
# for negatives
neg_prob = labels.copy()
neg_prob[neg_prob > 0] = 0
neg_prob = neg_prob * weights * -1
neg_prob = np.sum(neg_prob, axis=1)[:, np.newaxis]
return np.concatenate((pos_prob, neg_prob), axis=1)
def node_activation(self, data, target=None, wt='distance'):
weights, ind = None, None
if not target:
clf = neighbors.KNeighborsClassifier(
n_neighbors=self.codebook.nnodes)
labels = np.arange(0, self.codebook.matrix.shape[0])
clf.fit(self.codebook.matrix, labels)
# The codebook values are all normalized
# we can normalize the input data based on mean and std of
# original data
data = self._normalizer.normalize_by(self.data_raw, data)
weights, ind = clf.kneighbors(data)
# Softmax function
weights = 1. / weights
return weights, ind
def calculate_topographic_error(self):
bmus1 = self.find_bmu(self.data_raw, njb=1, nth=1)
bmus2 = self.find_bmu(self.data_raw, njb=1, nth=2)
topographic_error = None
if self.codebook.lattice == "rect":
bmus_gap = np.abs(
(self.bmu_ind_to_xy(np.array(bmus1[0]))[:, 0:2] -
self.bmu_ind_to_xy(np.array(bmus2[0]))[:, 0:2]).sum(axis=1))
topographic_error = np.mean(bmus_gap != 1)
elif self.codebook.lattice == "hexa":
dist_matrix_1 = self.codebook.lattice_distances[bmus1[0].astype(
int)].reshape(len(bmus1[0]), -1)
topographic_error = (np.array([
distances[bmu2]
for bmu2, distances in zip(bmus2[0].astype(int), dist_matrix_1)
]) > 2).mean()
return (topographic_error)
def calculate_quantization_error(self):
neuron_values = self.codebook.matrix[self.find_bmu(
self._data)[0].astype(int)]
quantization_error = np.mean(
np.abs(neuron_values - self._data)) # norme L1 et pas L2 ????
return quantization_error
def calculate_map_size(self, lattice):
"""
Calculates the optimal map size given a dataset using eigenvalues and eigenvectors. Matlab ported
:lattice: 'rect' or 'hex'
:return: map sizes
"""
D = self.data_raw.copy()
dlen = D.shape[0]
dim = D.shape[1]
munits = np.ceil(5 * (dlen**0.5))
A = np.ndarray(shape=[dim, dim]) + np.Inf
for i in range(dim):
D[:, i] = D[:, i] - np.mean(D[np.isfinite(D[:, i]), i])
for i in range(dim):
for j in range(dim):
c = D[:, i] * D[:, j]
c = c[np.isfinite(c)]
A[i, j] = sum(c) / len(c)
A[j, i] = A[i, j]
VS = np.linalg.eig(A)
eigval = sorted(VS[0])
if eigval[-1] == 0 or eigval[-2] * munits < eigval[-1]:
ratio = 1
else:
ratio = np.sqrt(eigval[-1] / eigval[-2])
if lattice == "rect":
size1 = min(munits, round(np.sqrt(munits / ratio)))
else:
size1 = min(munits, round(np.sqrt(munits / ratio * np.sqrt(0.75))))
size2 = round(munits / size1)
return [int(size1), int(size2)]
def calculate_neighbours_nodes(self):
res = []
for i in range(self.codebook.nnodes):
current = []
for j in range(self.codebook.nnodes):
if (self._distance_matrix[i][j] < 1.5
and self._distance_matrix[i][j] != 0):
current.append(j)
res.append(current)
return res
from codebook import Codebook import cv2 as cv import numpy as np FILE_NAME = "./examples/img.png" img = cv.imread(FILE_NAME, cv.CV_8UC1) count = Codebook.count_people(img) print(count)
def __init__(self,
data,
neighborhood,
normalizer=None,
mapsize=None,
mask=None,
mapshape='planar',
lattice='rect',
initialization='pca',
training='batch',
radius_train='linear',
name='sompy',
component_names=None,
components_to_plot=None,
isNormalized=False):
"""
Self Organizing Map
:param data: data to be clustered, represented as a matrix of n rows,
as inputs and m cols as input features
:param neighborhood: neighborhood object calculator.
:param normalizer: normalizer object calculator.
:param mapsize: tuple/list defining the dimensions of the som. If
single number is provided is considered as the number of nodes.
:param mask: mask
:param mapshape: shape of the som.
:param lattice: type of lattice.
:param initialization: method to be used for initialization of the som.
:param name: name used to identify the som
:param training: Training mode (seq, batch)
"""
if data is not None: # ajout LB
if normalizer and isNormalized == False:
for i in range(len(normalizer)):
data[:, i] = normalizer[i].normalize(data[:, i])
self._data = data
else:
self._data = data
self._data = self._data.astype('double') # ajout LB
self._dim = data.shape[1]
self._dlen = data.shape[0]
else:
self._data = None
self._dim = None
self._dlen = None
self._normalizer = normalizer
self._dlabel = None
self._bmu = None
self.name = name
self.data_raw = data
self.neighborhood = neighborhood
self.mapshape = mapshape
self.initialization = initialization
self.mask = mask
mapsize = self.calculate_map_size(lattice) if not mapsize else mapsize
self.mapsize = mapsize
self.codebook = Codebook(mapsize, lattice)
self.training = training
self.radius_train = radius_train
self._component_names = self.build_component_names(
) if component_names is None else [component_names]
self._distance_matrix = self.calculate_map_dist()
self.components_to_plot = components_to_plot