def svm_gradient_batch_fast(X_pred, X_exp, y, X_pred_ids, X_exp_ids, w, C=.0001, sigma=1.):
# sample Kernel
rnpred = X_pred_ids#sp.random.randint(low=0,high=len(y),size=n_pred_samples)
rnexpand = X_exp_ids#sp.random.randint(low=0,high=len(y),size=n_expand_samples)
#K = GaussKernMini_fast(X_pred.T,X_exp.T,sigma)
X1 = X_pred.T
X2 = X_exp.T
if sp.sparse.issparse(X1):
G = sp.outer(X1.multiply(X1).sum(axis=0), sp.ones(X2.shape[1]))
else:
G = sp.outer((X1 * X1).sum(axis=0), sp.ones(X2.shape[1]))
if sp.sparse.issparse(X2):
H = sp.outer(X2.multiply(X2).sum(axis=0), sp.ones(X1.shape[1]))
else:
H = sp.outer((X2 * X2).sum(axis=0), sp.ones(X1.shape[1]))
K = sp.exp(-(G + H.T - 2. * fast_dot(X1.T, X2)) / (2. * sigma ** 2))
# K = sp.exp(-(G + H.T - 2.*(X1.T.dot(X2)))/(2.*sigma**2))
if sp.sparse.issparse(X1) | sp.sparse.issparse(X2): K = sp.array(K)
# compute predictions
yhat = fast_dot(K,w[rnexpand])
# compute whether or not prediction is in margin
inmargin = (yhat * y[rnpred]) <= 1
# compute gradient
G = C * w[rnexpand] - fast_dot((y[rnpred] * inmargin), K)
return G,rnexpand
def _gradient_func(self, w):
bias, wf = self._split_coefficents(w)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(wf)
x = self._counter.x
xw = self._xw
z = numexpr.evaluate(
'(l_plus + l_minus) * xw - xv_plus - xv_minus - l_minus + l_plus')
grad = wf + self._rank_penalty * fast_dot(x.T, z)
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
xcs = numpy.dot(xc, wf)
grad += self._regr_penalty * (fast_dot(xc.T, xcs) +
xc.sum(axis=0) * bias -
fast_dot(xc.T, self.y_compressed))
# intercept
if self._fit_intercept:
grad_intercept = self._regr_penalty * (
xcs.sum() + xc.shape[0] * bias - self.y_compressed.sum())
grad = numpy.concatenate(([grad_intercept], grad))
return grad
def gain(self, X, y):
H = np.zeros(X.shape[1])
y = np.array(y)
i = 0
batch = 100000
n_features = X.shape[1]
n_examples = X.shape[0]
p = np.zeros(shape=(2, 2, X.shape[1]))
while i < n_features:
if n_features - i < batch:
batch = n_features - i
X_batch_raw = X[:, i : (i + batch)]
X_batch = X_batch_raw.toarray()
p[1, 1, i : (i + batch)] = y * X_batch_raw
p[1, 0, i : (i + batch)] = np.fabs(y - 1) * X_batch_raw
p[0, 1, i : (i + batch)] = fast_dot(y, np.fabs(X_batch - 1))
p[0, 0, i : (i + batch)] = fast_dot(np.fabs(y - 1), np.fabs(X_batch - 1))
p_batch = p[:, :, i : (i + batch)] / n_examples
p_sum = np.sum(p_batch, axis=0)
s = X_batch_raw.sum(axis=0)
p_x = np.array([s, 1 - s])
H[i : (i + batch)] = np.sum(p_batch * np.log(p_batch + self.smoother)) - 4 * np.sum(np.multiply(p_x, p_sum))
i += batch
print(i / X.shape[1])
return H
def _ica_par(X, tol, g, fun_args, max_iter, w_init):
"""Parallel FastICA.
Used internally by FastICA --main loop
"""
W = _sym_decorrelation(w_init)
del w_init
p_ = float(X.shape[1])
for ii in moves.xrange(max_iter):
U = fast_dot(W, X)
gwtx, g_wtx = g(np.abs(U)**2, fun_args)
W1 = _sym_decorrelation(
fast_dot(gwtx * U, np.conj(X.T)) / p_ - g_wtx[:, np.newaxis] * W)
del gwtx, g_wtx
# builtin max, abs are faster than numpy counter parts.
lim = max(abs(abs(np.diag(fast_dot(W1, np.conj(W.T)))) - 1))
W = W1
if lim < tol:
break
else:
warnings.warn(
'FastICA did not converge. Consider increasing tolerance or the maximum number of iterations.'
)
return W, ii + 1
def _gradient_func(self, beta_bias):
bias, beta = self._split_coefficents(beta_bias)
K = self._counter.x
Kw = self._Kw
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(beta)
z = numexpr.evaluate(
'(l_plus + l_minus) * Kw - xv_plus - xv_minus - l_minus + l_plus')
gradient = Kw + self._rank_penalty * fast_dot(K, z)
if self._has_time:
K_comp = K.compress(self.regr_mask, axis=0)
K_comp_beta = numpy.dot(K_comp, beta)
gradient += self._regr_penalty * (
fast_dot(K_comp.T, K_comp_beta) + K_comp.sum(axis=0) * bias -
fast_dot(K_comp.T, self.y_compressed))
# intercept
if self._fit_intercept:
grad_intercept = self._regr_penalty * (K_comp_beta.sum() +
K_comp.shape[0] * bias -
self.y_compressed.sum())
gradient = numpy.concatenate(([grad_intercept], gradient))
return gradient
def _hessian_func(self, w, s):
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(s)
x = self._counter.x
xs = numpy.dot(x, s)
xs = numexpr.evaluate('(l_plus + l_minus) * xs - xv_plus - xv_minus')
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
s = s + self._regr_penalty * fast_dot(xc.T, numpy.dot(xc, s))
return s + self._rank_penalty * fast_dot(x.T, xs)
def svm_gradient_batch(X_pred,X_exp,y,X_pred_ids,X_exp_ids,w,C=.0001,sigma=1.):
# sample Kernel
rnpred = X_pred_ids#sp.random.randint(low=0,high=len(y),size=n_pred_samples)
rnexpand = X_exp_ids#sp.random.randint(low=0,high=len(y),size=n_expand_samples)
K = GaussKernMini(X_pred.T,X_exp.T,sigma)
# compute predictions
yhat = fast_dot(K,w[rnexpand])
# compute whether or not prediction is in margin
inmargin = (yhat * y[rnpred]) <= 1
# compute gradient
G = C * w[rnexpand] - fast_dot((y[rnpred] * inmargin), K)
return G,rnexpand
def zca_whitening(self, image, eps):
"""
N = 1
X = image[:,:].reshape((N, -1)).astype(np.float64)
X = check_array(X, dtype=[np.float64], ensure_2d=True, copy=True)
# Center data
self.mean_ = np.mean(X, axis=0)
print(X.shape)
X -= self.mean_
U, S, V = linalg.svd(X, full_matrices=False)
# flip eigenvectors' sign to enforce deterministic output
U, V = svd_flip(U, V)
zca_matrix = U.dot(np.diag(1.0/np.sqrt(np.diag(S) + 1))).dot(U.T) #ZCA Whitening matrix
return fast_dot(zca_matrix, X).reshape(image.shape) #Data whitening
"""
image = self.local_contrast_normalization(image)
N = 1
X = image.reshape((N, -1))
pca = PCA(whiten=True, svd_solver='full', n_components=X.shape[-1])
transformed = pca.fit_transform(X) # return U
pca.whiten = False
zca = fast_dot(transformed, pca.components_ + eps) + pca.mean_
# zca = pca.inverse_transform(transformed)
return zca.reshape(image.shape)
def _update_coordinate_descent(X, W, Ht, l1_reg, l2_reg, shuffle,
random_state):
"""Helper function for _fit_coordinate_descent
Update W to minimize the objective function, iterating once over all
coordinates. By symmetry, to update H, one can call
_update_coordinate_descent(X.T, Ht, W, ...)
"""
n_components = Ht.shape[1]
HHt = fast_dot(Ht.T, Ht)
XHt = safe_sparse_dot(X, Ht)
# L2 regularization corresponds to increase of the diagonal of HHt
if l2_reg != 0.:
# adds l2_reg only on the diagonal
HHt.flat[::n_components + 1] += l2_reg
# L1 regularization corresponds to decrease of each element of XHt
if l1_reg != 0.:
XHt -= l1_reg
if shuffle:
permutation = random_state.permutation(n_components)
else:
permutation = np.arange(n_components)
# The following seems to be required on 64-bit Windows w/ Python 3.5.
permutation = np.asarray(permutation, dtype=np.intp)
return _update_cdnmf_fast(W, HHt, XHt, permutation)
def sliding_window(image, window_size, step_size):
pair = np.mgrid[890:1400:step_size[0], 250:730:step_size[1],
0.0:6.28:5].reshape(3, -1).T
for pts in pair:
xx = int(pts[0])
yy = int(pts[1])
tt = pts[2]
#crop = image[yy-halfy:yy+halfy+1,xx-halfx:xx+halfx+1]
crop = np.mgrid[yy - halfy:yy + halfy + 1,
xx - halfx:xx + halfx + 1].reshape(2, -1).T
crop[:, [0, 1]] = crop[:, [1, 0]]
col = crop.shape[0]
newp = np.ones((col, 3))
newp[:, :-1] = crop
transform = np.array([[
np.cos(tt), -np.sin(tt), -xx * np.cos(tt) + xx + yy * np.sin(tt)
], [np.sin(tt),
np.cos(tt), -xx * np.sin(tt) - yy * np.cos(tt) + yy], [0, 0, 1]])
newp = fast_dot(transform, newp.T)
newp = newp[0:2, :]
newp = np.transpose(newp).astype(int)
imx = newp[:, 0]
imy = newp[:, 1]
newim = np.array(image[imy, imx]).reshape(
(window_size[1], window_size[0]))
print(xx, yy, tt)
newim = read_one_image(newim)
#print(xx,yy,newim.shape)
yield (xx, yy, tt, newim)
def _update_coordinate_descent(X, W, Ht, l1_reg, l2_reg, shuffle,
random_state):
"""Helper function for _fit_coordinate_descent
Update W to minimize the objective function, iterating once over all
coordinates. By symmetry, to update H, one can call
_update_coordinate_descent(X.T, Ht, W, ...)
"""
n_components = Ht.shape[1]
HHt = fast_dot(Ht.T, Ht)
XHt = safe_sparse_dot(X, Ht)
# L2 regularization corresponds to increase of the diagonal of HHt
if l2_reg != 0.:
# adds l2_reg only on the diagonal
HHt.flat[::n_components + 1] += l2_reg
# L1 regularization corresponds to decrease of each element of XHt
if l1_reg != 0.:
XHt -= l1_reg
if shuffle:
permutation = random_state.permutation(n_components)
else:
permutation = np.arange(n_components)
# The following seems to be required on 64-bit Windows w/ Python 3.5.
permutation = np.asarray(permutation, dtype=np.intp)
return _update_cdnmf_fast(W, HHt, XHt, permutation)
def _gradient_func(self, w):
# sum over columns without running into overflow problems
# scipy.sparse.spmatrix.sum uses dtype of matrix, which is too small
col_sum = numpy.asmatrix(numpy.ones((1, self.Aw.shape[0]), dtype=numpy.int_)) * self.Aw
v = numpy.asarray(col_sum).squeeze()
z = fast_dot(self.data_x.T, self.Aw.T.dot(self.AXw) - v)
return w + self.alpha * z
def get_bmu(self, yn):
"""Returns the ID of the best matching unit.
Best is determined from the cosine similarity of the
sample with the normalized Kohonen network.
See https://en.wikipedia.org/wiki/Cosine_similarity
for cosine similarity documentation.
TODO: make possible the finding the second best matching unit
Parameters
----------
KN : sparse matrix
Shape = [n_nodes, n_features] must be normalized according to
l2 norm as used in the sklearn Normalizer()
y : vector of dimension 1 x nfeatures
Target sample.
Returns
-------
tuple : (loc, cosine_distance)
index of the matching unit, with the corresponding cosine distance
"""
#d = ((self.K_-y)**2).sum(axis=1)
#loc = np.argmin(d)
#qe = np.sqrt(d[loc])
similarity = fast_dot(self.KN_, yn.T)
loc = np.argmax(similarity)
qe = 1/(1.0e-4+similarity[loc])-1
return loc, qe
def _update_delta(self, m, mask=None, drop_diag=False):
self.delta_DK_M[m][:, :] = self.alpha * self.beta_M[m]
if mask is None and not drop_diag:
self.sumE_MK[m, :] = 1.
self.delta_DK_M[m][:, :] += self.sumE_MK.prod(axis=0)
assert np.isfinite(self.delta_DK_M[m]).all()
elif mask is None and drop_diag:
assert self.mode_dims[0] == self.mode_dims[1]
mask = np.abs(np.identity(self.mode_dims[0]) - 1)
if m > 1:
tmp = np.zeros(self.n_components)
for k in xrange(self.n_components):
tmp[k] = (mask * np.outer(self.E_DK_M[0][:, k],
self.E_DK_M[1][:, k])).sum()
assert tmp.shape == (self.n_components, )
self.sumE_MK[m, :] = 1.
else:
tmp = np.dot(mask, self.E_DK_M[np.abs(m - 1)])
assert tmp.shape == self.E_DK_M[m].shape
self.delta_DK_M[m][:, :] += self.sumE_MK[2:].prod(axis=0) * tmp
assert np.isfinite(self.delta_DK_M[m]).all()
else:
if drop_diag:
diag_idx = np.identity(self.mode_dims[0]).astype(bool)
assert (mask[diag_idx] == 0).all()
tmp = mask.copy()
tmp, order = make_first_mode(tmp, m)
tmp = fast_dot(tmp, self.E_DK_M[order[-1]])
for i in xrange(self.n_modes - 2, 0, -1):
tmp *= self.E_DK_M[order[i]]
tmp = tmp.sum(axis=-2)
self.delta_DK_M[m][:, :] += tmp
assert np.isfinite(self.delta_DK_M[m]).all()
def _beta_divergence_dense(X, W, H, beta):
"""Compute the beta-divergence of X and W.H for dense array only.
Used as a reference for testing nmf._beta_divergence.
"""
if isinstance(X, numbers.Number):
W = np.array([[W]])
H = np.array([[H]])
X = np.array([[X]])
WH = fast_dot(W, H)
if beta == 2:
return squared_norm(X - WH) / 2
WH_Xnonzero = WH[X != 0]
X_nonzero = X[X != 0]
np.maximum(WH_Xnonzero, 1e-9, out=WH_Xnonzero)
if beta == 1:
res = np.sum(X_nonzero * np.log(X_nonzero / WH_Xnonzero))
res += WH.sum() - X.sum()
elif beta == 0:
div = X_nonzero / WH_Xnonzero
res = np.sum(div) - X.size - np.sum(np.log(div))
else:
res = (X_nonzero ** beta).sum()
res += (beta - 1) * (WH ** beta).sum()
res -= beta * (X_nonzero * (WH_Xnonzero ** (beta - 1))).sum()
res /= beta * (beta - 1)
return res
def _beta_divergence_dense(X, W, H, beta):
"""Compute the beta-divergence of X and W.H for dense array only.
Used as a reference for testing nmf._beta_divergence.
"""
if isinstance(X, numbers.Number):
W = np.array([[W]])
H = np.array([[H]])
X = np.array([[X]])
WH = fast_dot(W, H)
if beta == 2:
return squared_norm(X - WH) / 2
WH_Xnonzero = WH[X != 0]
X_nonzero = X[X != 0]
np.maximum(WH_Xnonzero, 1e-9, out=WH_Xnonzero)
if beta == 1:
res = np.sum(X_nonzero * np.log(X_nonzero / WH_Xnonzero))
res += WH.sum() - X.sum()
elif beta == 0:
div = X_nonzero / WH_Xnonzero
res = np.sum(div) - X.size - np.sum(np.log(div))
else:
res = (X_nonzero**beta).sum()
res += (beta - 1) * (WH**beta).sum()
res -= beta * (X_nonzero * (WH_Xnonzero**(beta - 1))).sum()
res /= beta * (beta - 1)
return res
def repeated_corr(X, y, dtype=float):
"""Computes pearson correlations between a vector and a matrix.
Adapted from Jona-Sassenhagen's PR #L1772 on mne-python.
Parameters
----------
y : np.array, shape (n_samples)
Data vector.
X : np.array, shape (n_samples, n_measures)
Data matrix onto which the vector is correlated.
dtype : type, optional
Data type used to compute correlation values to optimize memory.
Returns
-------
rho : np.array, shape (n_measures)
"""
from sklearn.utils.extmath import fast_dot
if X.ndim not in [1, 2] or y.ndim != 1 or X.shape[0] != y.shape[0]:
raise ValueError('y must be a vector, and X a matrix with an equal'
'number of rows.')
if X.ndim == 1:
X = X[:, None]
y -= np.array(y.mean(0), dtype=dtype)
X -= np.array(X.mean(0), dtype=dtype)
y_sd = y.std(0, ddof=1)
X_sd = X.std(0, ddof=1)[:, None if y.shape == X.shape else Ellipsis]
return (fast_dot(y.T, X) / float(len(y) - 1)) / (y_sd * X_sd)
def _ica_def(X, tol, g, fun_args, max_iter, w_init):
"""Deflationary FastICA using fun approx to neg-entropy function
Used internally by FastICA.
"""
n_components = w_init.shape[0]
W = np.zeros((n_components, n_components), dtype=X.dtype)
n_iter = []
# j is the index of the extracted component
for j in range(n_components):
w = w_init[j, :].copy()
w /= np.sqrt((w ** 2).sum())
for i in moves.xrange(max_iter):
gwtx, g_wtx = g(fast_dot(w.T, X), fun_args)
w1 = (X * gwtx).mean(axis=1) - g_wtx.mean() * w
_gs_decorrelation(w1, W, j)
w1 /= np.sqrt((w1 ** 2).sum())
lim = np.abs(np.abs((w1 * w).sum()) - 1)
w = w1
if lim < tol:
break
n_iter.append(i + 1)
W[j, :] = w
return W, max(n_iter)
def _update_delta(self, m, mask=None, drop_diag=False):
self.delta_DK_M[m][:, :] = self.alpha * self.beta_M[m]
if mask is None and not drop_diag:
self.sumE_MK[m, :] = 1.
self.delta_DK_M[m][:, :] += self.sumE_MK.prod(axis=0)
assert np.isfinite(self.delta_DK_M[m]).all()
elif mask is None and drop_diag:
assert self.mode_dims[0] == self.mode_dims[1]
mask = np.abs(np.identity(self.mode_dims[0]) - 1)
if m > 1:
tmp = np.zeros(self.n_components)
for k in xrange(self.n_components):
tmp[k] = (mask * np.outer(self.E_DK_M[0][:, k], self.E_DK_M[1][:, k])).sum()
assert tmp.shape == (self.n_components,)
self.sumE_MK[m, :] = 1.
else:
tmp = np.dot(mask, self.E_DK_M[np.abs(m-1)])
assert tmp.shape == self.E_DK_M[m].shape
self.delta_DK_M[m][:, :] += self.sumE_MK[2:].prod(axis=0) * tmp
assert np.isfinite(self.delta_DK_M[m]).all()
else:
if drop_diag:
diag_idx = np.identity(self.mode_dims[0]).astype(bool)
assert (mask[diag_idx] == 0).all()
tmp = mask.copy()
tmp, order = make_first_mode(tmp, m)
tmp = fast_dot(tmp, self.E_DK_M[order[-1]])
for i in xrange(self.n_modes - 2, 0, -1):
tmp *= self.E_DK_M[order[i]]
tmp = tmp.sum(axis=-2)
self.delta_DK_M[m][:, :] += tmp
assert np.isfinite(self.delta_DK_M[m]).all()
def _gradient_func(self, w):
# sum over columns without running into overflow problems
# scipy.sparse.spmatrix.sum uses dtype of matrix, which is too small
col_sum = numpy.asmatrix(numpy.ones((1, self.Aw.shape[0]), dtype=numpy.int_)) * self.Aw
v = numpy.asarray(col_sum).squeeze()
z = fast_dot(self.data_x.T, self.Aw.T.dot(self.AXw) - v)
return w + self.alpha * z
def _gradient_func(self, w):
if self._last_w is not None and (w == self._last_w).all():
return self._last_gradient
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(w)
x = self._counter.x
xs = numpy.dot(x, w)
z = numexpr.evaluate(
'(l_plus + l_minus) * xs - xv_plus - xv_minus - l_minus + l_plus')
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
w = w + self._regr_penalty * (fast_dot(xc.T, numpy.dot(xc, w)) -
fast_dot(xc.T, self.y_compressed))
self._last_gradient = w + self._rank_penalty * fast_dot(x.T, z)
self._last_w = w
return self._last_gradient
def applyPCA(sampleData, mean, components): pca = PCA(n_components=components.shape[0]) pca.components_ = components pca.mean_ = mean transform = pca.transform(np.array([sampleData])) reconstructed = fast_dot(transform, pca.components_) + pca.mean_ reconstructed = reconstructed[0] return sampleData / reconstructed
def applyPCA(sampleData, mean, components):
pca = PCA(n_components=components.shape[0])
pca.components_ = components
pca.mean_ = mean
transform = pca.transform(np.array([sampleData]))
reconstructed = fast_dot(transform, pca.components_) + pca.mean_
reconstructed = reconstructed[0]
return sampleData / reconstructed
def transform(self, mf):
if self.sparse:
block = safe_sparse_dot(mf.as_csr_matrix(), self.rot)
else:
X = mf.as_np_array()
if self.means is not None:
X = X - self.means
block = fast_dot(X, self.rot) / np.sqrt(self.explained_variance)
mf.select_columns([]) # clear the multiframe
mf.append_np_block("pca_features", block, self.output_names)
return mf
def _gradient_func(self, w):
bias, wf = self._split_coefficents(w)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(wf)
x = self._counter.x
xw = self._xw
z = numexpr.evaluate('(l_plus + l_minus) * xw - xv_plus - xv_minus - l_minus + l_plus')
grad = wf + self._rank_penalty * fast_dot(x.T, z)
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
xcs = numpy.dot(xc, wf)
grad += self._regr_penalty * (fast_dot(xc.T, xcs) + xc.sum(axis=0) * bias - fast_dot(xc.T, self.y_compressed))
# intercept
if self._fit_intercept:
grad_intercept = self._regr_penalty * (xcs.sum() + xc.shape[0] * bias - self.y_compressed.sum())
grad = numpy.concatenate(([grad_intercept], grad))
return grad
def get_bmu(kn, yn, epsilon=1.0e-6):
"""Returns the ID of the best matching unit.
Best is determined from the cosine similarity of the
sample with the normalized Kohonen network.
See https://en.wikipedia.org/wiki/Cosine_similarity
for cosine similarity documentation.
"""
similarity = fast_dot(kn, yn.T)
loc = np.argmax(similarity)
qe = 1 / (epsilon + similarity[loc]) - 1
return loc, qe
def inverse_transform(self, X):
"""Transform data back to its original space, i.e.,
return an input X_original whose transform would be X
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples is the number of samples
and n_components is the number of components.
Returns
-------
X_original array-like, shape (n_samples, n_features)
"""
check_is_fitted(self, 'mean_')
if self.whiten:
return fast_dot(
X,
np.sqrt(self.explained_variance_[:, np.newaxis]) *
self.components_) + self.mean_
else:
return fast_dot(X, self.components_) + self.mean_
def _special_dot_X(W, H, X):
"""Computes np.dot(W, H) in a special way:
- If X is sparse, np.dot(W, H) is computed only where X is non zero,
and a sparse matrix is returned, with the same sparsity as X.
- If X is masked, np.dot(W, H) is computed entirely, and a masked array is
returned, with the same mask as X.
- If X is dense, np.dot(W, H) is computed entirely, and returned as a dense
array.
"""
if sp.issparse(X):
ii, jj = X.nonzero()
dot_vals = np.multiply(W[ii, :], H.T[jj, :]).sum(axis=1)
WH = sp.coo_matrix((dot_vals, (ii, jj)), shape=X.shape)
return WH.tocsr()
elif isinstance(X, np.ma.masked_array):
WH = np.ma.masked_array(fast_dot(W, H), mask=X.mask)
WH.unshare_mask()
return WH
else:
return fast_dot(W, H)
def _gradient_func(self, beta_bias):
bias, beta = self._split_coefficents(beta_bias)
K = self._counter.x
Kw = self._Kw
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(beta)
z = numexpr.evaluate('(l_plus + l_minus) * Kw - xv_plus - xv_minus - l_minus + l_plus')
gradient = Kw + self._rank_penalty * fast_dot(K, z)
if self._has_time:
K_comp = K.compress(self.regr_mask, axis=0)
K_comp_beta = numpy.dot(K_comp, beta)
gradient += self._regr_penalty * (fast_dot(K_comp.T, K_comp_beta)
+ K_comp.sum(axis=0) * bias - fast_dot(K_comp.T, self.y_compressed))
# intercept
if self._fit_intercept:
grad_intercept = self._regr_penalty * (K_comp_beta.sum() + K_comp.shape[0] * bias - self.y_compressed.sum())
gradient = numpy.concatenate(([grad_intercept], gradient))
return gradient
def GaussKernMini_fast(X1,X2,sigma):
if sp.sparse.issparse(X1):
G = sp.outer(X1.multiply(X1).sum(axis=0),sp.ones(X2.shape[1]))
else:
G = sp.outer((X1 * X1).sum(axis=0),sp.ones(X2.shape[1]))
if sp.sparse.issparse(X2):
H = sp.outer(X2.multiply(X2).sum(axis=0),sp.ones(X1.shape[1]))
else:
H = sp.outer((X2 * X2).sum(axis=0),sp.ones(X1.shape[1]))
K = sp.exp(-(G + H.T - 2.*fast_dot(X1.T,X2))/(2.*sigma**2))
# K = sp.exp(-(G + H.T - 2.*(X1.T.dot(X2)))/(2.*sigma**2))
if sp.sparse.issparse(X1) | sp.sparse.issparse(X2): K = sp.array(K)
return K
def _hessian_func(self, w, s):
s_bias, s_feat = self._split_coefficents(s)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(s_feat)
x = self._counter.x
xs = numpy.dot(x, s_feat)
xs = numexpr.evaluate('(l_plus + l_minus) * xs - xv_plus - xv_minus')
hessp = s_feat + self._rank_penalty * fast_dot(x.T, xs)
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
hessp += self._regr_penalty * fast_dot(xc.T, numpy.dot(xc, s_feat))
# intercept
if self._fit_intercept:
xsum = xc.sum(axis=0)
hessp += self._regr_penalty * xsum * s_bias
hessp_intercept = self._regr_penalty * xc.shape[0] * s_bias + self._regr_penalty * numpy.dot(xsum, s_feat)
hessp = numpy.concatenate(([hessp_intercept], hessp))
return hessp
def _hessian_func(self, w, s):
s_bias, s_feat = self._split_coefficents(s)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(s_feat)
x = self._counter.x
xs = numpy.dot(x, s_feat)
xs = numexpr.evaluate('(l_plus + l_minus) * xs - xv_plus - xv_minus')
hessp = s_feat + self._rank_penalty * fast_dot(x.T, xs)
if self._has_time:
xc = x.compress(self.regr_mask, axis=0)
hessp += self._regr_penalty * fast_dot(xc.T, numpy.dot(xc, s_feat))
# intercept
if self._fit_intercept:
xsum = xc.sum(axis=0)
hessp += self._regr_penalty * xsum * s_bias
hessp_intercept = self._regr_penalty * xc.shape[0] * s_bias + self._regr_penalty * numpy.dot(xsum, s_feat)
hessp = numpy.concatenate(([hessp_intercept], hessp))
return hessp
def transform(self, subspace, feats):
if self.debug:
print('-- coding features ...')
sys.stdout.flush()
feats -= subspace['mean']
transformed_feats = fast_dot(feats, subspace['components'].T)
# transformed_feats /= np.sqrt(subspace['explained_variance'])
return transformed_feats
def fit_transform(self, X, y=None):
"""Fit the model with X and apply the dimensionality reduction on X.
Parameters
----------
X : array-like, shape (n_samples, n_features)
New data, where n_samples in the number of samples
and n_features is the number of features.
Returns
-------
X_new : array-like, shape (n_samples, n_components)
"""
X = check_array(X)
X = self._fit(X)
return fast_dot(X, self.components_.T)
def _hessian_func(self, beta, s):
s_bias, s_feat = self._split_coefficents(s)
K = self._counter.x
Ks = numpy.dot(K, s_feat)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(s_feat)
xs = numexpr.evaluate('(l_plus + l_minus) * Ks - xv_plus - xv_minus')
hessian = Ks + self._rank_penalty * fast_dot(K, xs)
if self._has_time:
K_comp = K.compress(self.regr_mask, axis=0)
hessian += self._regr_penalty * fast_dot(K_comp.T, numpy.dot(K_comp, s_feat))
# intercept
if self._fit_intercept:
xsum = K_comp.sum(axis=0)
hessian += self._regr_penalty * xsum * s_bias
hessian_intercept = self._regr_penalty * K_comp.shape[0] * s_bias \
+ self._regr_penalty * numpy.dot(xsum, s_feat)
hessian = numpy.concatenate(([hessian_intercept], hessian))
return hessian
def _hessian_func(self, beta, s):
s_bias, s_feat = self._split_coefficents(s)
K = self._counter.x
Ks = numpy.dot(K, s_feat)
l_plus, xv_plus, l_minus, xv_minus = self._counter.calculate(s_feat)
xs = numexpr.evaluate('(l_plus + l_minus) * Ks - xv_plus - xv_minus')
hessian = Ks + self._rank_penalty * fast_dot(K, xs)
if self._has_time:
K_comp = K.compress(self.regr_mask, axis=0)
hessian += self._regr_penalty * fast_dot(K_comp.T, numpy.dot(K_comp, s_feat))
# intercept
if self._fit_intercept:
xsum = K_comp.sum(axis=0)
hessian += self._regr_penalty * xsum * s_bias
hessian_intercept = self._regr_penalty * K_comp.shape[0] * s_bias \
+ self._regr_penalty * numpy.dot(xsum, s_feat)
hessian = numpy.concatenate(([hessian_intercept], hessian))
return hessian
def transform_PCA(pca, k, X):
X_reduced = X - pca.mean_
X_reduced = fast_dot(X_reduced, pca.components_[0:k].T)
# Transform test data with principal components:
#X_reduced = pca.transform(test_X)
# Reconstruct:
X_rec = np.dot(X_reduced, pca.components_[0:k])
# Restore mean:
X_rec += pca.mean_
print "Variance Explained: {}".format(np.sum(pca.explained_variance_ratio_[:k]))
return X_reduced, X_rec
def nmf_predict_direct(rate_matrix,user_distribution,item_distribution,user_ids_list,item_ids_list,top_n,fout_str):
fout = open(fout_str,'w')
#method 1 : w*h
for u_ix,u in enumerate(user_distribution):
predict_vec = fast_dot(u,item_distribution)
filter_vec = np.where(rate_matrix.getrow(u_ix).toarray()>0,0,1)
predict_vec = predict_vec * filter_vec
sort_ix_vec = np.argpartition(-predict_vec[0],top_n)[:top_n]
candidate_item_list = list()
for i_ix in sort_ix_vec:
item_id = item_ids_list[i_ix]
candidate_item_list.append(item_id)
user_id = user_ids_list[u_ix]
print >> fout,'%s,%s' %(user_id,'#'.join(candidate_item_list))
fout.close()
def fit_transform(self, X, y=None):
"""Fit the model with X and apply the dimensionality reduction on X.
Parameters
----------
X : array-like, shape (n_samples, n_features)
Training data, where n_samples is the number of samples
and n_features is the number of features.
Returns
-------
X_new : array-like, shape (n_samples, n_components)
"""
self._fit(X)
if self.copy and self.center_ is not None:
X = X - self.center_
return fast_dot(X, self.components_.T)
def fit_transform(self, X, y=None):
"""Fit the model with X and apply the dimensionality reduction on X.
Parameters
----------
X : array-like, shape (n_samples, n_features)
Training data, where n_samples is the number of samples
and n_features is the number of features.
Returns
-------
X_new : array-like, shape (n_samples, n_components)
"""
self._fit(X)
if self.copy and self.center_ is not None:
X = X - self.center_
return fast_dot(X, self.components_.T)
def _reorth(basis, target, rows=None, alpha=0.5):
"""Reorthogonalize a vector using iterated Gram-Schmidt
Parameters
----------
basis: ndarray, shape (n_features, n_basis)
The matrix whose rows are a set of basis to reorthogonalize against
target: ndarray, shape (n_features,)
The target vector to be reorthogonalized
rows: {array-like, None}, default None
Indices of rows from basis to use. Use all if None
alpha: float, default 0.5
Parameter for determining whether to do a second reorthogonalization.
Returns
-------
reorthed_target: ndarray, shape (n_features,)
The reorthogonalized vector
"""
if rows is not None:
basis = basis[rows]
norm_target = norm(target)
norm_target_old = 0
n_reorth = 0
while norm_target < alpha * norm_target_old or n_reorth == 0:
for row in basis:
t = fast_dot(row, target)
target = target - t * row
norm_target_old = norm_target
norm_target = norm(target)
n_reorth += 1
if n_reorth > 4:
# target in span(basis) => accpet target = 0
target = np.zeros(basis.shape[0])
break
return target
def _reorth(basis, target, rows=None, alpha=0.5):
"""Reorthogonalize a vector using iterated Gram-Schmidt
Parameters
----------
basis: ndarray, shape (n_features, n_basis)
The matrix whose rows are a set of basis to reorthogonalize against
target: ndarray, shape (n_features,)
The target vector to be reorthogonalized
rows: {array-like, None}, default None
Indices of rows from basis to use. Use all if None
alpha: float, default 0.5
Parameter for determining whether to do a second reorthogonalization.
Returns
-------
reorthed_target: ndarray, shape (n_features,)
The reorthogonalized vector
"""
if rows is not None:
basis = basis[rows]
norm_target = norm(target)
norm_target_old = 0
n_reorth = 0
while norm_target < alpha * norm_target_old or n_reorth == 0:
for row in basis:
t = fast_dot(row, target)
target = target - t * row
norm_target_old = norm_target
norm_target = norm(target)
n_reorth += 1
if n_reorth > 4:
# target in span(basis) => accpet target = 0
target = np.zeros(basis.shape[0])
break
return target
def repeated_corr(X, y, dtype=float):
"""Computes pearson correlations between a vector and a matrix.
Adapted from Jona-Sassenhagen's PR #L1772 on mne-python.
Parameters
----------
X : np.array, shape (n_samples, n_measures)
Data matrix onto which the vector is correlated.
y : np.array, shape (n_samples)
Data vector.
dtype : type, optional
Data type used to compute correlation values to optimize memory.
Returns
-------
rho : np.array, shape (n_measures)
"""
from sklearn.utils.extmath import fast_dot
if not isinstance(X, np.ndarray):
X = np.array(X)
if X.ndim == 1:
X = X[:, None]
shape = X.shape
X = np.reshape(X, [shape[0], -1])
if X.ndim not in [1, 2] or y.ndim != 1 or X.shape[0] != y.shape[0]:
raise ValueError('y must be a vector, and X a matrix with an equal'
'number of rows.')
if X.ndim == 1:
X = X[:, None]
ym = np.array(y.mean(0), dtype=dtype)
Xm = np.array(X.mean(0), dtype=dtype)
y -= ym
X -= Xm
y_sd = y.std(0, ddof=1)
X_sd = X.std(0, ddof=1)[:, None if y.shape == X.shape else Ellipsis]
R = (fast_dot(y.T, X) / float(len(y) - 1)) / (y_sd * X_sd)
R = np.reshape(R, shape[1:])
# cleanup variable changed in place
y += ym
X += Xm
return R
def inverse_transform(self, X):
"""Transform data back to its original space, i.e.,
return an input X_original whose transform would be X
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples is the number of samples
and n_components is the number of components.
Returns
-------
X_original: array-like, shape (n_samples, n_features)
"""
check_is_fitted(self, 'center_')
X_original = fast_dot(X, self.components_)
if self.center_ is not None:
X_original = X_original + self.center_
return X_original
def inverse_transform(self, X):
"""Transform data back to its original space, i.e.,
return an input X_original whose transform would be X
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples is the number of samples
and n_components is the number of components.
Returns
-------
X_original array-like, shape (n_samples, n_features)
Notes
-----
If whitening is enabled, inverse_transform does not compute the
exact inverse operation as transform.
"""
return fast_dot(X, self.components_) + self.mean_
def inverse_transform(self, X):
"""Transform data back to its original space, i.e.,
return an input X_original whose transform would be X
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples is the number of samples
and n_components is the number of components.
Returns
-------
X_original: array-like, shape (n_samples, n_features)
"""
check_is_fitted(self, 'center_')
X_original = fast_dot(X, self.components_)
if self.center_ is not None:
X_original = X_original + self.center_
return X_original
def inverse_transform(self, X):
"""Transform data back to its original space, i.e.,
return an input X_original whose transform would be X
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples is the number of samples
and n_components is the number of components.
Returns
-------
X_original array-like, shape (n_samples, n_features)
Notes
-----
If whitening is enabled, inverse_transform does not compute the
exact inverse operation as transform.
"""
return fast_dot(X, self.components_) + self.mean_
def transform(self, X, y=None):
"""Apply dimensionality reduction on X.
X is projected on the first principal components previous extracted
from a training set.
Parameters
----------
X : array-like, shape (n_samples, n_features)
New data, where n_samples in the number of samples
and n_features is the number of features.
Returns
-------
X_new : array-like, shape (n_samples, n_components)
"""
check_is_fitted(self, 'mean_')
X = check_array(X)
if self.mean_ is not None:
X = X - self.mean_
X = fast_dot(X, self.components_.T)
return X
def _bound(self, mask=None, drop_diag=False):
"""Evidence Lower Bound (ELBO)"""
if mask is None and not drop_diag:
B = self.sumE_MK.prod(axis=0).sum()
elif mask is None and drop_diag:
assert self.mode_dims[0] == self.mode_dims[1]
mask = np.abs(np.identity(self.mode_dims[0]) - 1)
tmp = np.zeros(self.n_components)
for k in xrange(self.n_components):
tmp[k] = (mask * np.outer(self.E_DK_M[0][:, k],
self.E_DK_M[1][:, k])).sum()
B = (self.sumE_MK[2:].prod(axis=0) * tmp).sum()
else:
tmp = mask.copy()
tmp = fast_dot(tmp, self.E_DK_M[-1])
for i in xrange(self.n_modes - 2, 0, -1):
tmp *= self.E_DK_M[i]
tmp = tmp.sum(axis=-2)
assert tmp.shape == self.E_DK_M[0].shape
B = (tmp * self.E_DK_M[0]).sum()
B -= np.log(self.sparse_data + 1).sum()
B += (self.sparse_data * np.log(self._sparse_zeta())).sum()
K = self.n_components
for m in xrange(self.n_modes):
D = self.mode_dims[m]
B += (self.alpha - 1.) * (np.log(self.G_DK_M[m]).sum())
B -= (self.alpha * self.beta_M[m]) * (self.sumE_MK[m, :].sum())
B -= K * D * (sp.gammaln(self.alpha) -
self.alpha * np.log(self.alpha * self.beta_M[m]))
gamma_DK = self.gamma_DK_M[m]
delta_DK = self.delta_DK_M[m]
B += (-(gamma_DK - 1.) * sp.psi(gamma_DK) - np.log(delta_DK) +
gamma_DK + sp.gammaln(gamma_DK)).sum()
return B
def inverse_transform(self, X, y=None):
"""Transform data back to its original space.
Returns an array X_original whose transform would be X.
Parameters
----------
X : array-like, shape (n_samples, n_components)
New data, where n_samples in the number of samples
and n_components is the number of components.
Returns
-------
X_original array-like, shape (n_samples, n_features)
Notes
-----
If whitening is enabled, inverse_transform does not compute the
exact inverse operation of transform.
"""
check_is_fitted(self, 'mean_')
X_original = fast_dot(X, self.components_)
if self.mean_ is not None:
X_original = X_original + self.mean_
return X_original
def transform(self, X):
"""Apply the dimensionality reduction on X.
X is projected on the first principal components previous extracted
from a training set.
Parameters
----------
X : array-like, shape (n_samples, n_features)
New data, where n_samples is the number of samples
and n_features is the number of features.
Returns
-------
X_new : array-like, shape (n_samples, n_components)
"""
X = array2d(X)
if self.mean_ is not None:
X = X - self.mean_
X_transformed = fast_dot(X, self.components_.T)
return X_transformed
def _bound(self, mask=None, drop_diag=False):
"""Evidence Lower Bound (ELBO)"""
if mask is None and not drop_diag:
B = self.sumE_MK.prod(axis=0).sum()
elif mask is None and drop_diag:
assert self.mode_dims[0] == self.mode_dims[1]
mask = np.abs(np.identity(self.mode_dims[0]) - 1)
tmp = np.zeros(self.n_components)
for k in xrange(self.n_components):
tmp[k] = (mask * np.outer(self.E_DK_M[0][:, k], self.E_DK_M[1][:, k])).sum()
B = (self.sumE_MK[2:].prod(axis=0) * tmp).sum()
else:
tmp = mask.copy()
tmp = fast_dot(tmp, self.E_DK_M[-1])
for i in xrange(self.n_modes - 2, 0, -1):
tmp *= self.E_DK_M[i]
tmp = tmp.sum(axis=-2)
assert tmp.shape == self.E_DK_M[0].shape
B = (tmp * self.E_DK_M[0]).sum()
B -= np.log(self.sparse_data + 1).sum()
B += (self.sparse_data * np.log(self._sparse_zeta())).sum()
K = self.n_components
for m in xrange(self.n_modes):
D = self.mode_dims[m]
B += (self.alpha - 1.) * (np.log(self.G_DK_M[m]).sum())
B -= (self.alpha * self.beta_M[m])*(self.sumE_MK[m, :].sum())
B -= K*D*(sp.gammaln(self.alpha) - self.alpha*np.log(self.alpha * self.beta_M[m]))
gamma_DK = self.gamma_DK_M[m]
delta_DK = self.delta_DK_M[m]
B += (-(gamma_DK - 1.)*sp.psi(gamma_DK) - np.log(delta_DK)
+ gamma_DK + sp.gammaln(gamma_DK)).sum()
return B
def inverse_transform(self, X, copy=True):
"""Transform the sources back to the mixed data (apply mixing matrix).
Parameters
----------
X : array-like, shape (n_samples, n_components)
Sources, where n_samples is the number of samples
and n_components is the number of components.
copy : bool (optional)
If False, data passed to fit are overwritten. Defaults to True.
Returns
-------
X_new : array-like, shape (n_samples, n_features)
"""
check_is_fitted(self, 'mixing_')
####X = check_array(X, copy=(copy and self.whiten), dtype=FLOAT_DTYPES)
X = fast_dot(X, self.mixing_.T)
if self.whiten:
X += self.mean_
return X
def transform(self, X, y=None):
"""Apply dimensionality reduction on X.
X is projected on the principal components previous extracted
from a training set.
Parameters
----------
X : array-like, shape (n_samples, n_features)
New data, where n_samples in the number of samples
and n_features is the number of features.
Returns
-------
X_transformed : array-like, shape (n_samples, n_components)
"""
check_is_fitted(self, 'center_')
X = check_array(X)
if self.center_ is not None:
X = X - self.center_
X_transformed = fast_dot(X, self.components_.T)
return X_transformed
def test_fast_dot():
"""Check fast dot blas wrapper function"""
if fast_dot is np.dot:
return
rng = np.random.RandomState(42)
A = rng.random_sample([2, 10])
B = rng.random_sample([2, 10])
try:
linalg.get_blas_funcs(['gemm'])[0]
has_blas = True
except (AttributeError, ValueError):
has_blas = False
if has_blas:
# Test _fast_dot for invalid input.
# Maltyped data.
for dt1, dt2 in [['f8', 'f4'], ['i4', 'i4']]:
assert_raises(ValueError, _fast_dot, A.astype(dt1),
B.astype(dt2).T)
# Malformed data.
## ndim == 0
E = np.empty(0)
assert_raises(ValueError, _fast_dot, E, E)
## ndim == 1
assert_raises(ValueError, _fast_dot, A, A[0])
## ndim > 2
assert_raises(ValueError, _fast_dot, A.T, np.array([A, A]))
## min(shape) == 1
assert_raises(ValueError, _fast_dot, A, A[0, :][None, :])
# test for matrix mismatch error
assert_raises(ValueError, _fast_dot, A, A)
# Test cov-like use case + dtypes.
for dtype in ['f8', 'f4']:
A = A.astype(dtype)
B = B.astype(dtype)
# col < row
C = np.dot(A.T, A)
C_ = fast_dot(A.T, A)
assert_almost_equal(C, C_, decimal=5)
C = np.dot(A.T, B)
C_ = fast_dot(A.T, B)
assert_almost_equal(C, C_, decimal=5)
C = np.dot(A, B.T)
C_ = fast_dot(A, B.T)
assert_almost_equal(C, C_, decimal=5)
# Test square matrix * rectangular use case.
A = rng.random_sample([2, 2])
for dtype in ['f8', 'f4']:
A = A.astype(dtype)
B = B.astype(dtype)
C = np.dot(A, B)
C_ = fast_dot(A, B)
assert_almost_equal(C, C_, decimal=5)
C = np.dot(A.T, B)
C_ = fast_dot(A.T, B)
assert_almost_equal(C, C_, decimal=5)
if has_blas:
for x in [np.array([[d] * 10] * 2) for d in [np.inf, np.nan]]:
assert_raises(ValueError, _fast_dot, x, x.T)
A = rng.random_sample([2, 10])
B = rng.random_sample([2, 10])
try:
linalg.get_blas_funcs('gemm')
has_blas = True
except AttributeError, ValueError:
has_blas = False
if has_blas:
# test dispatch to np.dot
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter('always', NonBLASDotWarning)
# maltyped data
for dt1, dt2 in [['f8', 'f4'], ['i4', 'i4']]:
fast_dot(A.astype(dt1), B.astype(dt2).T)
assert_true(isinstance(w.pop(-1).message, NonBLASDotWarning))
# malformed data
# ndim == 0
E = np.empty(0)
fast_dot(E, E)
assert_true(isinstance(w.pop(-1).message, NonBLASDotWarning))
## ndim == 1
fast_dot(A, A[0])
assert_true(isinstance(w.pop(-1).message, NonBLASDotWarning))
## ndim > 2
fast_dot(A.T, np.array([A, A]))
assert_true(isinstance(w.pop(-1).message, NonBLASDotWarning))
## min(shape) == 1
assert_raises(ValueError, fast_dot, A, A[0, :][None, :])
# test for matrix mismatch error
def pca_biplot_with_clustering(data_matrix, feature_labels, mean_normalize = False, k_means_post = True, K = 5, n_components=2, f_out = "foo"):
"""
Inputs:
data_matrix: numpy array of shape n x f where n is the number of samples and f is the number of features (variables)
feature_labels: the list of human-interpretable labels for the rows in data_matrix. Used for the biplot
mean_normalize: if true, each row i of data_matrix is replaced with i - mean(i)
k_means_post: if True, k means is run after the PCA analysis on the projected data matrix. I.e., in the reduced space.
if False, it is run before the PCA, and before the mean_normalization, and then the clusters are projected into the space.
n_components: the number of PCA vectors you want.
f_out: the path to write the graph (as PDF) to
TODO: the graph currently only supports n=2.
Outputs:
r_loadings: the loading vectors
Side Effects:
file created on disk at f_out
References:
http://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html
https://github.com/teddyroland/python-biplot/blob/master/biplot.py
http://stackoverflow.com/questions/21217710/factor-loadings-using-sklearn
http://stackoverflow.com/questions/14716965/r-principle-component-analysis-label-of-component
http://nxn.se/post/36838219245/loadings-with-scikit-learn-pca
"""
data_matrix = np.nan_to_num(data_matrix) #clustering does not allow infs, nans
if not k_means_post:
centroids, labels = clustering.kmpp(data_matrix, K)
if mean_normalize:
for iindex, i in enumerate(data_matrix):
data_matrix[iindex] = i - np.mean(i)
rows, variables = np.shape(data_matrix)
pca = PCA(n_components=n_components)
pca.fit(data_matrix)
r_loadings = np.matrix.transpose(pca.components_) #components is n_components x f; transpase to get rows as features, like R does it
transformed_matrix = fast_dot(data_matrix, r_loadings)
loading_vectors = []
loading_labels = []
for i in range(0, variables):
loading_vectors.append((list(r_loadings[i])))
loading_labels.append(str(feature_labels[i]))
if k_means_post:
centroids, labels = clustering.kmpp(transformed_matrix, K)
#do the plot
fig, ax = plt.subplots()
for l in range(0, K):
transformed_matrix_i = [transformed_matrix[i] for i in range(0, rows) if labels[i] == l]
ax.scatter([x[0] for x in transformed_matrix_i],
[x[1] for x in transformed_matrix_i],
color = _colors[l % len(_colors)],
label = "cluster" + str(l))
pxs = [i[0] for i in loading_vectors]
pys = [i[1] for i in loading_vectors]
for xindex, x in enumerate(pxs):
ax.arrow(0, 0, pxs[xindex], pys[xindex], linewidth=3, width=0.0005, head_width=0.0025, color="r", label = "loading")
ax.arrow(0, 0, pxs[xindex]*5, pys[xindex]*5, alpha = .5, linewidth = 1, linestyle="dashed", width=0.0005, head_width=0.0025, color="r")
ax.text(pxs[xindex]*5, pys[xindex]*5, loading_labels[xindex], color="r")
ax.legend(loc='best')
x0,x1 = ax.get_xlim()
y0,y1 = ax.get_ylim()
ax.set_xlim(min(x0, y0), max(x1,y1)) #make it square
ax.set_ylim(min(x0, y0), max(x1,y1))
ax.set_xlabel("PCA[0]")
ax.set_ylabel("PCA[1]")
ax.set_title("", fontsize = 20)
ax.grid(b=True, which='major', color='k', linestyle='--')
fig.savefig(f_out + "{0}{1}{2}".format(K, "_meannormalized" if mean_normalize else "", "_kmpost" if k_means_post else "_kmfirst") + ".pdf", format="pdf")
plt.close() #THIS IS CRUCIAL SEE: http://stackoverflow.com/questions/26132693/matplotlib-saving-state-between-different-uses-of-io-bytesio
return r_loadings
def _multiplicative_update_h(X, W, H, beta_loss, l1_reg_H, l2_reg_H, gamma):
"""update H in Multiplicative Update NMF"""
X_mask = X.mask if isinstance(X, np.ma.masked_array) else False
if beta_loss == 2:
if X_mask is False:
numerator = safe_sparse_dot(W.T, X)
denominator = fast_dot(fast_dot(W.T, W), H)
else:
numerator = _safe_dot(W.T, X)
WH = _special_dot_X(W, H, X)
denominator = _safe_dot(W.T, WH)
else:
# Numerator
WH_safe_X = _special_dot_X(W, H, X)
if sp.issparse(X):
WH_safe_X_data = WH_safe_X.data
X_data = X.data
else:
WH_safe_X_data = WH_safe_X
X_data = X
# copy used in the Denominator
WH = WH_safe_X.copy()
if beta_loss - 1. < 0:
WH[np.logical_and(WH == 0, ~X_mask)] = EPSILON
# to avoid division by zero
if beta_loss - 2. < 0:
WH_safe_X_data[
np.logical_and(WH_safe_X_data == 0, ~X_mask)] = EPSILON
if beta_loss == 1:
# to work around spurious warnings coming out of masked arrays
with np.errstate(invalid='ignore'):
np.divide(X_data, WH_safe_X_data, out=WH_safe_X_data)
else:
WH_safe_X_data **= beta_loss - 2
# element-wise multiplication
WH_safe_X_data *= X_data
# here numerator = dot(W.T, (dot(W, H) ** (beta_loss - 2)) * X)
numerator = _safe_dot(W.T, WH_safe_X)
# Denominator
if beta_loss == 1:
if X_mask is False:
W_sum = np.sum(W, axis=0) # shape(n_components, )
W_sum = W_sum[:, np.newaxis]
else:
W_sum = np.dot(W.T, ~X_mask)
W_sum[W_sum == 0] = 1.
denominator = W_sum
else:
# computation of WtWH = dot(W.T, dot(W, H) ** beta_loss - 1)
if sp.issparse(X):
# memory efficient computation
# (compute column by column, avoiding the dense matrix WH)
WtWH = np.empty(H.shape)
for i in range(X.shape[1]):
WHi = fast_dot(W, H[:, i])
if beta_loss - 1 < 0:
WHi[WHi == 0] = EPSILON
WHi **= beta_loss - 1
WtWH[:, i] = fast_dot(W.T, WHi)
else:
WH **= beta_loss - 1
WtWH = _safe_dot(W.T, WH)
denominator = WtWH
# Add L1 and L2 regularization
if l1_reg_H > 0:
denominator += l1_reg_H
if l2_reg_H > 0:
denominator = denominator + l2_reg_H * H
denominator[denominator == 0] = EPSILON
numerator /= denominator
delta_H = numerator
# gamma is in ]0, 1]
if gamma != 1:
delta_H **= gamma
return delta_H
