def test_monotonic_likelihood():
# We check that each step of the each step of variational inference without
# regularization improve monotonically the training set of the bound
rng = np.random.RandomState(0)
rand_data = RandomData(rng, scale=7)
n_components = rand_data.n_components
for covar_type in COVARIANCE_TYPE:
X = rand_data.X[covar_type]
bgmm = BayesianGaussianMixture(
n_components=2 * n_components,
covariance_type=covar_type,
warm_start=True,
max_iter=1,
random_state=rng,
tol=1e-4,
)
current_lower_bound = -np.infty
# Do one training iteration at a time so we can make sure that the
# training log likelihood increases after each iteration.
for _ in range(500):
prev_lower_bound = current_lower_bound
current_lower_bound = bgmm.fit(X).lower_bound_
assert_greater_equal(current_lower_bound, prev_lower_bound)
if bgmm.converged_:
break
assert bgmm.converged_
def BayesianGaussianMixture(V, **kwargs):
"""Performs clustering on *V* by using Gaussian mixture models with variational inference. The function uses :func:`sklearn.micture.GaussianMixture`. See sklearn documents
for details.
:arg V: row-normalized eigenvectors for the purpose of clustering.
:type V: :class:`numpy.ndarray`
:arg n_clusters: specifies the number of clusters.
:type n_clusters: int
"""
try:
from sklearn.mixture import BayesianGaussianMixture
except ImportError:
raise ImportError('Use of this function (BayesianGaussianMixture) requires the '
'installation of sklearn.')
n_components = kwargs.pop('n_components', None)
if n_components == None:
n_components = kwargs.pop('n_clusters',None)
if n_components == None:
n_components = 1
n_init = kwargs.pop('n_init', 1)
mixture = BayesianGaussianMixture(n_init=n_init, **kwargs).fit(V)
return mixture.fit_predict(V)
def test_bayesian_mixture_fit_predict_n_init():
# Check that fit_predict is equivalent to fit.predict, when n_init > 1
X = np.random.RandomState(0).randn(1000, 5)
gm = BayesianGaussianMixture(n_components=5, n_init=10, random_state=0)
y_pred1 = gm.fit_predict(X)
y_pred2 = gm.predict(X)
assert_array_equal(y_pred1, y_pred2)
def test_check_covariance_precision():
# We check that the dot product of the covariance and the precision
# matrices is identity.
rng = np.random.RandomState(0)
rand_data = RandomData(rng, scale=7)
n_components, n_features = 2 * rand_data.n_components, 2
# Computation of the full_covariance
bgmm = BayesianGaussianMixture(n_components=n_components,
max_iter=100, random_state=rng, tol=1e-3,
reg_covar=0)
for covar_type in COVARIANCE_TYPE:
bgmm.covariance_type = covar_type
bgmm.fit(rand_data.X[covar_type])
if covar_type == 'full':
for covar, precision in zip(bgmm.covariances_, bgmm.precisions_):
assert_almost_equal(np.dot(covar, precision),
np.eye(n_features))
elif covar_type == 'tied':
assert_almost_equal(np.dot(bgmm.covariances_, bgmm.precisions_),
np.eye(n_features))
elif covar_type == 'diag':
assert_almost_equal(bgmm.covariances_ * bgmm.precisions_,
np.ones((n_components, n_features)))
else:
assert_almost_equal(bgmm.covariances_ * bgmm.precisions_,
np.ones(n_components))
def test_bayesian_mixture_predict_predict_proba():
# this is the same test as test_gaussian_mixture_predict_predict_proba()
rng = np.random.RandomState(0)
rand_data = RandomData(rng)
for prior_type in PRIOR_TYPE:
for covar_type in COVARIANCE_TYPE:
X = rand_data.X[covar_type]
Y = rand_data.Y
bgmm = BayesianGaussianMixture(
n_components=rand_data.n_components,
random_state=rng,
weight_concentration_prior_type=prior_type,
covariance_type=covar_type)
# Check a warning message arrive if we don't do fit
assert_raise_message(NotFittedError,
"This BayesianGaussianMixture instance"
" is not fitted yet. Call 'fit' with "
"appropriate arguments before using "
"this method.", bgmm.predict, X)
bgmm.fit(X)
Y_pred = bgmm.predict(X)
Y_pred_proba = bgmm.predict_proba(X).argmax(axis=1)
assert_array_equal(Y_pred, Y_pred_proba)
assert_greater_equal(adjusted_rand_score(Y, Y_pred), .95)
def fit(self, X, Y):
# assume classes are numbered 0...K-1
self.K = len(set(Y))
self.gaussians = []
self.p_y = np.zeros(self.K)
for k in range(self.K):
print("Fitting gmm", k)
Xk = X[Y == k]
self.p_y[k] = len(Xk)
gmm = BayesianGaussianMixture(10)
gmm.fit(Xk)
self.gaussians.append(gmm)
# normalize p(y)
self.p_y /= self.p_y.sum()
def test_bayesian_mixture_fit_predict(seed, max_iter, tol):
rng = np.random.RandomState(seed)
rand_data = RandomData(rng, scale=7)
n_components = 2 * rand_data.n_components
for covar_type in COVARIANCE_TYPE:
bgmm1 = BayesianGaussianMixture(n_components=n_components,
max_iter=max_iter, random_state=rng,
tol=tol, reg_covar=0)
bgmm1.covariance_type = covar_type
bgmm2 = copy.deepcopy(bgmm1)
X = rand_data.X[covar_type]
Y_pred1 = bgmm1.fit(X).predict(X)
Y_pred2 = bgmm2.fit_predict(X)
assert_array_equal(Y_pred1, Y_pred2)
def kde_entropy_sklearn_gmm(points, n_est=None, n_components=None):
"""
Use sklearn.neigbors.KernelDensity pdf to estimate entropy.
Data is standardized before kde.
Sample points drawn from gaussian mixture model from original points.
Fails for bimodal and dirichlet, similar to statsmodels kde.
"""
from sklearn.mixture import BayesianGaussianMixture as GMM
n, d = points.shape
# Default to the full set
if n_est is None:
n_est = n
# reduce size of draw to n_est
if n_est >= n:
x = points
else:
x = points[permutation(n)[:n_est]]
n = n_est
if n_components is None:
n_components = int(5*sqrt(d))
predictor = GMM(n_components=n_components, covariance_type='full',
#verbose=True,
max_iter=1000)
predictor.fit(x)
evaluation_points, _ = predictor.sample(n_est)
logp = sklearn_log_density(x, evaluation_points=evaluation_points)
H = -np.mean(logp)
return H / LN2
def gmm_entropy(points, n_est=None, n_components=None):
#from sklearn.mixture import GaussianMixture as GMM
from sklearn.mixture import BayesianGaussianMixture as GMM
n, d = points.shape
# Default to the full set
if n_est is None:
n_est = n
# reduce size of draw to n_est
if n_est >= n:
x = points
else:
x = points[permutation(n)[:n_est]]
n = n_est
if n_components is None:
n_components = int(5*sqrt(d))
## Standardization doesn't seem to help
## Note: sigma may be zero
#x, mu, sigma = standardize(x) # if standardized
predictor = GMM(n_components=n_components, covariance_type='full',
#verbose=True,
max_iter=1000)
predictor.fit(x)
eval_x, _ = predictor.sample(n_est)
weight_x = predictor.score_samples(eval_x)
H = -np.mean(weight_x)
#with np.errstate(divide='ignore'): H = H + np.sum(np.log(sigma)) # if standardized
dH = 0.
## cross-check against own calcs
#alt = GaussianMixture(predictor.weights_, mu=predictor.means_, sigma=predictor.covariances_)
#print("alt", H, alt.entropy())
#print(np.vstack((weight_x[:10], alt.logpdf(eval_x[:10]))).T)
return H / LN2, dH / LN2
def test_bayesian_mixture_precisions_prior_initialisation():
rng = np.random.RandomState(0)
n_samples, n_features = 10, 2
X = rng.rand(n_samples, n_features)
# Check raise message for a bad value of degrees_of_freedom_prior
bad_degrees_of_freedom_prior_ = n_features - 1.
bgmm = BayesianGaussianMixture(
degrees_of_freedom_prior=bad_degrees_of_freedom_prior_,
random_state=rng)
assert_raise_message(
ValueError, "The parameter 'degrees_of_freedom_prior' should be "
"greater than %d, but got %.3f." %
(n_features - 1, bad_degrees_of_freedom_prior_), bgmm.fit, X)
# Check correct init for a given value of degrees_of_freedom_prior
degrees_of_freedom_prior = rng.rand() + n_features - 1.
bgmm = BayesianGaussianMixture(
degrees_of_freedom_prior=degrees_of_freedom_prior,
random_state=rng).fit(X)
assert_almost_equal(degrees_of_freedom_prior,
bgmm.degrees_of_freedom_prior_)
# Check correct init for the default value of degrees_of_freedom_prior
degrees_of_freedom_prior_default = n_features
bgmm = BayesianGaussianMixture(
degrees_of_freedom_prior=degrees_of_freedom_prior_default,
random_state=rng).fit(X)
assert_almost_equal(degrees_of_freedom_prior_default,
bgmm.degrees_of_freedom_prior_)
# Check correct init for a given value of covariance_prior
covariance_prior = {
'full': np.cov(X.T, bias=1) + 10,
'tied': np.cov(X.T, bias=1) + 5,
'diag': np.diag(np.atleast_2d(np.cov(X.T, bias=1))) + 3,
'spherical': rng.rand()
}
bgmm = BayesianGaussianMixture(random_state=rng)
for cov_type in ['full', 'tied', 'diag', 'spherical']:
bgmm.covariance_type = cov_type
bgmm.covariance_prior = covariance_prior[cov_type]
bgmm.fit(X)
assert_almost_equal(covariance_prior[cov_type], bgmm.covariance_prior_)
# Check raise message for a bad spherical value of covariance_prior
bad_covariance_prior_ = -1.
bgmm = BayesianGaussianMixture(covariance_type='spherical',
covariance_prior=bad_covariance_prior_,
random_state=rng)
assert_raise_message(
ValueError, "The parameter 'spherical covariance_prior' "
"should be greater than 0., but got %.3f." % bad_covariance_prior_,
bgmm.fit, X)
# Check correct init for the default value of covariance_prior
covariance_prior_default = {
'full': np.atleast_2d(np.cov(X.T)),
'tied': np.atleast_2d(np.cov(X.T)),
'diag': np.var(X, axis=0, ddof=1),
'spherical': np.var(X, axis=0, ddof=1).mean()
}
bgmm = BayesianGaussianMixture(random_state=0)
for cov_type in ['full', 'tied', 'diag', 'spherical']:
bgmm.covariance_type = cov_type
bgmm.fit(X)
assert_almost_equal(covariance_prior_default[cov_type],
bgmm.covariance_prior_)
def find_nuclear_contours(cell_rois, cell_stack, nuclear_stack, um_to_px):
nuclear_rois = {}
unknown_rois = {}
nuclear_centroids = {}
status_tags = {}
for ID, cell_roi in cell_rois.items():
print(f'ID: {ID}')
# although, 8 seems pyknotic
# 45, and 41, are instances of multis
ids = [7]
ids = [id - 1 for id in ids]
#if ID not in ids: continue
cell_contours = cell_roi['contours']
nuclear_contours = []
unknown_contours = []
nuclear_centroids_list = []
status_tags_list = []
for tp, cell_contour in enumerate(cell_contours):
#if tp > 0: break
# Acquire cellular slice and offset.
cell_slice, cell_offset = get_slice_and_offset(cell_contour)
# Use slice to acquire cell subimage.
cell_subimg = cell_stack[tp][cell_slice]
tmp = cell_contour - np.array([cell_offset[::-1]])
cx, cy = extract_improved_cell_centroid(cell_subimg, tmp)
cell_offset = cell_slice[0].start, cell_slice[1].start
nuclear_centroids_list.append(np.array([[cx, cy]]) + np.array([cell_offset[::-1]]))
nuc_subimg = nuclear_stack[tp][cell_slice]
cell_mask = np.zeros_like(cell_subimg)
tmp = cell_contour - np.array([cell_offset[::-1]])
cv2.drawContours(cell_mask, [tmp], -1, 255, -1)
nuc_subimg[cell_mask == 0] = 0
# Remove everything outside of the cell.
nuc_subimg[nuc_subimg < np.percentile(nuc_subimg, 10)] = 0
nuc_subimg = rmp.subtract(nuc_subimg)
nuc_subimg[nuc_subimg < np.percentile(nuc_subimg, 30)] = 0
nuc_subimg = morphology.grey_opening(nuc_subimg, size=(3,3))
DELTA = um_to_px(10)
mask = np.zeros_like(nuc_subimg)
cv2.circle(mask, (cx, cy), DELTA, 255, -1)
nuc_subimg[mask == 0] = 0
median_filtered = skimage.filters.median(nuc_subimg)
# Median filtering leads to pixels with 0 value having much larger nonzero values. Ensure the pixels
# with low values continue to be low.
median_filtered[nuc_subimg <= np.percentile(nuc_subimg, 5)] = 0
nuc_subimg = median_filtered
def plt_center():
plt.plot(cx, cy, 'r.')
def fail_func(tag):
nuclear_contours.append(np.empty(0))
unknown_contours.append(np.empty(0))
status_tags_list.append(tag)
COMP = 5
# 1D GMM AREA
# Fit only to the nonzero portion of the image to attain a sharper fit.
gmm = BayesianGaussianMixture(n_components=COMP)
nonzero_nuc_subimg = nuc_subimg[nuc_subimg != 0]
if nonzero_nuc_subimg.size == 0:
fail_func('insufficient')
continue
try:
gmm.fit(nonzero_nuc_subimg.reshape(-1, 1))
except ValueError:
fail_func('insufficient-2')
continue
#visualize('nuc_subimg', nuc_subimg)
gpred = gmm.predict(nuc_subimg.reshape(-1, 1)).reshape(nuc_subimg.shape).astype(np.uint8)
label_of_min = np.argmin(gmm.means_)
# If the minimum's label is not zero, then switch it to zero.
if label_of_min != 0:
# Temporarily set zero-label to value guaranteed to exceed total Gaussian count.
tmp = COMP + 1
gpred[gpred == 0] = tmp
# Ensure the minimum label is zero.
gpred[gpred == label_of_min] = 0
# Ensure whatever label was zero is now set to the minimum label.
gpred[gpred == tmp] = label_of_min
# Switch the means to reflect the change.
tmp = np.copy(gmm.means_[0])
gmm.means_[0] = gmm.means_[label_of_min]
gmm.means_[label_of_min] = tmp
# Ensure whatever was background in the nuclear image remains as background. Without this line, it is possible
# that GMM prediction, which was fed nonzero pixels only, would incorrectly classify background. This is necessary.
gpred[nuc_subimg == 0] = 0
#visualize(f'{ID+1}-T{tp+1}gpred', gpred, plt_center)
# Extract contours and choose the one closest to the SCE.
P = np.ones_like(nuc_subimg, dtype=np.uint8)
P[gpred == 0] = 0
_, contours, _ = cv2.findContours(P, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
compute_roi_centroid_dist = lambda contour: compute_dist(np.array([cx, cy]), contour)
try:
nuclear_contour = min(contours, key=compute_roi_centroid_dist)
except ValueError:
fail_func('unfound')
continue
#visualize('gpred', gpred)
# Upon having found the closest contour, remove rest from image.
Q = np.zeros_like(nuc_subimg, dtype=np.uint8)
cv2.drawContours(Q, [nuclear_contour], -1, 255, -1)
#visualize('Q', Q)
gpred[Q == 0] = 0
# Check for pyknotic.
if gpred[cy, cx] == 0:
fail_func('pyknotic')
continue
# Crop the image.
x, y, w, h = cv2.boundingRect(nuclear_contour)
max_y, max_x = gpred.shape
crop_y, crop_x = max(0, y-1), max(0, x-1)
cropped_gpred = gpred[crop_y:min(max_y, y+h+1), crop_x:min(max_x, x+w+1)]
# Update cellular offset and SCE.
cell_offset += np.array([crop_y, crop_x])
cx -= crop_x
cy -= crop_y
#visualize('gpred', gpred)
img, label_num = skimage.measure.label(cropped_gpred, return_num=True)
if label_num == 1:
pass
#print(np.where(img == 1))
img = img.astype(np.uint8)
# Add one to label number because 0 is not included in count. 0 is background.
label_num += 1
def traverse_row_left_to_right(row, inward_links, incidence_counts):
# Moving left-to-right, find each index where its right-neighbor differs from it.
diffs = np.where(row[:-1] != row[1:])[0]
# Make tuples of these labels and their right-neighbors; of the label that changes upon
# right-traversal, and what it changes to.
transitions = list(zip(row[diffs], row[diffs+1]))
# The directed graph:
labels_within = [0]
for left, right in transitions:
incidence_counts[left, right] += 1
if right not in labels_within:
inward_links[left].add(right)
labels_within.append(right)
def traverse_rows(img, inward_links, incidence_counts):
for row in img:
traverse_row_left_to_right(row, inward_links, incidence_counts)
# Now reverse row and do same.
traverse_row_left_to_right(row[::-1], inward_links, incidence_counts)
def build_graphs(img, label_num):
inward_links = defaultdict(set)
incidence_counts = np.zeros((label_num,) * 2)
traverse_rows(img, inward_links, incidence_counts)
# Transpose image and to do same for columns.
traverse_rows(img.T, inward_links, incidence_counts)
# Turn incidence counts into proportions.
incidence_proportions = np.array([c / np.sum(c) for c in incidence_counts])
return inward_links, incidence_proportions
#visualize('img', img)
inward_links, incidence_proportions = build_graphs(img, label_num)
'''
_labels = sorted(inward_links.keys())
for l in _labels:
print(f'{l}: {inward_links[l]}')
for ix, row in enumerate(incidence_proportions):
_row = [str(round(x, 2)) for x in row]
print(f'{ix}: {_row}')
'''
# Determine which label is to be taken as the background interface.
bg_interface_label = np.argmax(incidence_proportions[0])
# Now will follow a series of rules that use the graph information to properly setup the subsequent portions.
#visualize('img', img, plt_center)
# Ensure that any ROI that is fully contained within another label is coalesced into that label.
for label in range(label_num):
# If the label has no inward links, then it is fully contained in another label.
if label not in inward_links:
# Find which label it is contained inside.
for _label, links in inward_links.items():
if label in links:
#print(f'{label} not found. It is inside {_label}.')
img[img == label] = _label
break
# Ensure everything that is not background or background interface is same label.
# Multiplying by 2 here because: a) need label to be different than the background interface label,
# and it helps visualization for the values to be separated.
signal_label = bg_interface_label * 2
img[np.logical_and(img != 0, img != bg_interface_label)] = signal_label
# Find all signal ROIs. Keep one closest to SCE and eliminate rest.
_img = np.copy(img)
_img[img == bg_interface_label] = 0
_, contours, _ = cv2.findContours(_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
compute_roi_centroid_dist = lambda contour: compute_dist(np.array([cx, cy]), contour)
#visualize('img', _img)
try: contour = min(contours, key=compute_roi_centroid_dist)
except:
fail_func('strange-1')
continue
# Eliminate any other signal ROIs.
mask = np.zeros_like(img)
cv2.drawContours(mask, [contour], -1, 255, -1)
img[np.logical_and(mask == 0, img != bg_interface_label)] = 0
# Check that SCE is on signal. If not, strange.
sce_label = img[cy, cx]
if sce_label in [0, bg_interface_label]:
fail_func('strange-2')
continue
# Will now compute the centroid of the signal ROI to expand around for attaining the nuclear ROI estimate.
# To better estimate the centroid of the nuclear region, will erode the signal ROI. This will remove portions
# of the ROI which may come from either other nuclei or other non-nuclear regions that make it non-circular.
# These portions would skew the centroid estimate.
# Acquire the contour's bounding rectangle.
x, y, w, h = cv2.boundingRect(contour)
min_dim = min(w, h)
# Will use a 3x3 structuring element to erode. To be conservative, only erode a quarter of the minimum dimension
# of the bounding rectangle. Erosion will function on both sides of the ROI, so eroding half of the minimum dimension
# on either side would make the ROI vanish. Eroding half of half will ensure the ROI persists. This is done somewhat
# heuristically to improve results.
iter = int(min_dim / 4)
#visualize('mask', mask)
mask = binary_erosion(mask, structure=np.array([[0,1,0],[1,1,1],[0,1,0]]), iterations=iter).astype(np.uint8)
_, contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# Try to find a contour after erosion. If this fails, continue to use contour extracted before erosion.
try: contour = contours[0]
except IndexError: contour = contour
#visualize('postmask', mask)
# Compute the centroid of the chosen signal ROI.
my, mx = map(int, [np.mean(contour[:, :, 1]), np.mean(contour[:, :, 0])])
# Now begin drawing concentric circles centered at this centroid with increasing radius.
# When the circle overlaps background, either the background interface or otherwise, determine that
# it has encompassed the full extent of the nucleus that will be measured.
center = mx, my
mask = np.zeros_like(img)
radius = 0
# 2 is currently used as the signal ROI label.
non_signal_labels = np.empty(0)
while non_signal_labels.size == 0:
radius += 1
cv2.circle(mask, center, radius, 255, -1)
non_signal_labels = img[np.logical_and(mask != 0, img != signal_label)]
# Subtract 1 as it just failed the loop condition.
radius -= 1
# After fitting a circle, try to expand into an ellipse.
# First try one axis, then the other.
radius1 = radius
mask = np.zeros_like(img)
non_signal_labels = np.empty(0)
while non_signal_labels.size == 0:
radius1 += 1
cv2.ellipse(mask, center, (radius1, radius), 0, 0, 360, 255, -1)
non_signal_labels = img[np.logical_and(mask != 0, img != signal_label)]
# Subtract 1 as it just failed the loop condition.
radius1 -= 1
radius2 = radius
mask = np.zeros_like(img)
non_signal_labels = np.empty(0)
while non_signal_labels.size == 0:
radius2 += 1
cv2.ellipse(mask, center, (radius1, radius2), 0, 0, 360, 255, -1)
non_signal_labels = img[np.logical_and(mask != 0, img != signal_label)]
# Subtract 1 as it just failed the loop condition.
radius2 -= 1
_, contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
nuclear_contour = contours[0]
if nuclear_contour.size == 0:
fail_func('unfound')
continue
# If SCE is not within nuclear contour, do not choose contour.
if cv2.pointPolygonTest(nuclear_contour, (cx, cy), False) != 1:
fail_func('external-SCE')
continue
def plt_m():
plt.plot(mx, my, 'b.')
# Now find all ROIs that are not background or interface. Choose one closest to SCE.
#visualize('img', img, plt_m)
# Now specify the indeterminate region. This will be entire contour, sans background.
img[img != 0] = 255
_, contours, _ = cv2.findContours(img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
unknown_contour = contours[0]
#itercount = max(1, int(DELTA / 10))
# Currently leaving out the containing own centroid portion.
# If made it this far, believe that we've extracted the SCE containing contour.
# NOT E: this could have failed. Need to have additional steps to account for failure.
# Will need to check for own centroid.
# If all contained are of either this label or whatever other label is found, then have found a region containing this one.
#mask1 = np.zeros_like(gpred)
#mask2 = np.zeros_like(gpred)
# Draw contour recently found, dilate it, and re-extract.
#cv2.drawContours(mask1, [contour], -1, 255, -1)
#cv2.drawContours(mask2, [contour], -1, 255, -1)
#mask2 = binary_dilation(mask2, structure=np.ones((3, 3))).astype(np.uint8)
#xor_mask = np.logical_xor(mask1, mask2)
# Visualize XOR
#tmp = np.copy(gpred)
#tmp[np.logical_not(xor_mask)] = 0
#visualize('gpred', gpred)
#visualize('xor', tmp)
#S = np.copy(tmp)
#visualize('pre', pred2)
#S = binary_opening(S)
#S = binary_dilation(S, iterations=itercount).astype(np.uint8)
#S[cell_mask == 0] = 0
#_, unknown_contour_estimates, _ = cv2.findContours(S, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
unknown_contours.append(unknown_contour + np.array([cell_offset[::-1]]))
nuclear_contours.append(nuclear_contour + np.array([cell_offset[::-1]]))
status_tags_list.append('')
nuclear_rois[ID] = nuclear_contours
unknown_rois[ID] = unknown_contours
nuclear_centroids[ID] = nuclear_centroids_list
status_tags[ID] = status_tags_list
return nuclear_rois, unknown_rois, nuclear_centroids, status_tags
def _mean(data): # aggregated data
gm = BayesianGaussianMixture(n_components=2)
gm.fit(data)
return gm.means_
import numpy as np
import matplotlib.pyplot as plt
from sklearn.datasets.samples_generator import make_blobs
from sklearn.mixture import GaussianMixture
from sklearn.mixture import BayesianGaussianMixture
from copy import deepcopy
aic = []
bic = []
n_com = 12
data, label = make_blobs(n_samples=300, n_features=2, centers=10)
gmm_without = GaussianMixture(n_components=4).fit_predict(data)
gmm_vbem = BayesianGaussianMixture(n_components=10).fit_predict(data)
plt.scatter(data[:, 0], data[:, 1], c=label)
plt.show()
low = 99999
for n in range(1, n_com + 1):
gmm = GaussianMixture(n_components=n)
gmm.fit(data)
aic.append(gmm.aic(data))
if aic[-1] < low:
low = aic[-1]
model = deepcopy(gmm)
low2 = 99999
for n in range(1, n_com + 1):
gmm = GaussianMixture(n_components=n)
gmm.fit(data)
class Pyxelate:
CONVOLUTIONS = np.array(
[[[2, 2], [2, 2]], [[11, -1], [-1, -1]], [[-1, 11], [-1, -1]],
[[-1, -1], [11, -1]], [[-1, -1], [-1, 11]], [[5, 5], [-1, -1]],
[[-1, -1], [5, 5]], [[5, -1], [5, -1]], [[-1, 5], [-1, 5]],
[[5, -1], [-1, 5]], [[-1, 5], [5, -1]], [[-1, 3], [3, 3]],
[[3, -1], [3, 3]], [[3, 3], [-1, 3]], [[3, 3], [3, -1]]],
dtype="int")
SOLUTIONS = np.array([
[[1, 1], [1, 1]],
[[0, 1], [1, 1]],
[[1, 0], [1, 1]],
[[1, 1], [0, 1]],
[[1, 1], [1, 0]],
[[1, 1], [0, 0]],
[[0, 0], [1, 1]],
[[1, 0], [1, 0]],
[[0, 1], [0, 1]],
[[1, 0], [1, 0]],
[[0, 1], [0, 1]],
[[1, 0], [0, 0]],
[[0, 1], [0, 0]],
[[0, 0], [1, 0]],
[[0, 0], [0, 1]],
],
dtype="bool")
ITER = 2
def __init__(self,
height,
width,
color=8,
dither=True,
alpha=.6,
regenerate_palette=True,
keyframe=.6,
sensitivity=.07,
random_state=0):
"""Create instance for generating similar pixel arts."""
self.height = int(height)
self.width = int(width)
if self.width < 1 or self.height < 1:
raise ValueError("Result can not be smaller than 1x1 pixels.")
self.color = int(color)
if self.color < 2:
raise ValueError("The minimum number of colors is 2.")
elif self.color > 32:
raise ValueError("The maximum number of colors is 32.")
if dither:
self.dither = 1 / (self.color + 1)
else:
self.dither = 0.
self.alpha = float(alpha) # threshold for opacity
self.regenerate_palette = bool(regenerate_palette)
self.keyframe = keyframe # threshold for differences between keyframes
self.sensitivity = sensitivity # threshold for differences between parts of keyframes
# BGM
self.is_fitted = False
self.random_state = int(random_state)
self.model = BayesianGaussianMixture(
n_components=self.color,
max_iter=256,
covariance_type="tied",
weight_concentration_prior_type="dirichlet_distribution",
mean_precision_prior=1. / 256.,
warm_start=False,
random_state=self.random_state)
def convert(self, image):
"""Generate pixel art from image"""
return self._convert(image, False, False)
def _convert(self, image, override_adapthist=False, override_dither=False):
"""Generate pixel art from image or sequence of images"""
# does the image have alpha channel?
if self._is_transparent(image):
# remove artifacts from transparent edges
image = self._dilate(image)
# create alpha mask
mask = resize(image[:, :, 3], (self.height, self.width),
anti_aliasing=True)
# mask for colors
color_mask = resize(image[:, :, 3], (32, 32),
anti_aliasing=False).ravel()
else:
mask = None
color_mask = None
# apply adaptive contrast
if not override_adapthist:
image = self._fix_hist(image)
# create sample for finding palette
if self.regenerate_palette or not self.is_fitted:
examples = resize(image[:, :, :3], (32, 32),
anti_aliasing=False).reshape(-1, 3).astype("int")
if color_mask is not None:
# transparent colors should be ignored
examples = examples[color_mask >= self.alpha]
self._fit_model(examples)
# resize image to 4 times the desired width and height
image = resize(
image[:, :, :3],
(self.height * self.ITER * 2, self.width * self.ITER * 2),
anti_aliasing=True)
# generate pixelated image with desired width / height
t = time()
image = self._reduce(image)
print('SS:', time() - t)
# apply palette
height, width, depth = image.shape
reshaped = np.reshape(image, (height * width, depth))
probs = self.model.predict_proba(reshaped)
y = np.argmax(probs, axis=1)
# increase hue and snap color values to multiples of 8
palette = rgb2hsv(self.model.means_.reshape(-1, 1, 3))
palette[:, :, 1] *= 1.14 # empirical magic number
palette = hsv2rgb(palette).reshape(self.color, 3) // 8 * 8
palette[palette ==
248] = 255 # clamping // 8 * 8 would rarely allow 255 values
# generate recolored image
image = palette[y]
# apply dither over threshold if it's not zero
if not override_dither and self.dither:
# get second best probability by removing the best one
probs[np.arange(len(y)), y] = 0
# get new best and values
v = np.max(probs, axis=1) > self.dither
y = np.argmax(probs, axis=1)
# replace every second pixel with second best color
pad = not bool(width % 2)
if pad:
# make sure to alternate between starting positions
# bottleneck
for i in range(0, len(image), 2):
i += (i // width) % 2
if v[i]:
image[i] = palette[y[i]]
else:
i = np.argwhere(v[::2]) * 2
image[i] = palette[y[i]]
image = np.reshape(image, (height, width, depth))
if mask is not None:
# use transparency from original image, but make it either 0 or 255
mask[mask >= self.alpha] = 255
mask[mask < self.alpha] = 0
image = np.dstack(
(image, mask)) # result has lost its alpha channel
return np.clip(image.astype("int"), 0, 255).astype("uint8")
def convert_sequence(self, images):
"""Generates sequence of pixel arts from a list of images"""
try:
_ = np.array(images, dtype=float)
except ValueError:
# image sizes are different == setting an array element with a sequence
raise ValueError("Shape of images in list are different.")
# apply adaptive histogram on each
images = [self._fix_hist(image) for image in images]
transparent = self._is_transparent(images[0])
keyframe_limit = self.keyframe * np.prod(images[0].shape) * 255.
sensitivity_limit = self.sensitivity * 255.
diff_images, key_frames = [], []
# create new images that are just the differences between sequences
for image in images:
# add first image
if diff_images:
diff = np.abs(image[:, :, :3] - diff_images[-1][:, :, :3])
# image is not too different, from previous one, create mask
if np.sum(diff) < keyframe_limit:
diff = resize(np.mean(diff, axis=2),
(self.height, self.width),
anti_aliasing=True)
over, under = diff > sensitivity_limit, diff <= sensitivity_limit
diff[over], diff[under] = 255, 0.
diff = resize(diff, (image.shape[0], image.shape[1]),
anti_aliasing=False)
# was the image already transparent?
if transparent:
image[:, :, 3] = diff
else:
image = np.dstack((image, diff))
key_frames.append(False)
else:
key_frames.append(True)
else:
key_frames.append(True)
# add transparency layer for keyframes also, for easier broadcasting
if not self._is_transparent(image):
image = np.dstack(
(image, np.ones((image.shape[0], image.shape[1]))))
diff_images.append(image)
# create a palette from all images if possible
if self.regenerate_palette:
warnings.warn(
"using regenerate_palette=True will result in flickering, as the palette will be regenerated for each image!",
Warning)
else:
self._palette_from_list(diff_images)
# merge keyframes and differences
last = None
for image, key in zip(diff_images, key_frames):
current = self._convert(image, True,
~key) # pyxelate keyframe / change
if last is None:
last = current
else:
# merge differences to previous images
mask = ~np.logical_xor(last[:, :, 3], current[:, :, 3])
last[mask] = current[mask]
# generator
yield last.copy()
def _palette_from_list(self, images):
"""Fit model to find palette using all images in list at once"""
transparency = self._is_transparent(images[0])
examples = []
color_masks = []
# sample from all images
for image in images:
examples.append(
resize(image[:, :, :3], (16, 16),
anti_aliasing=False).reshape(-1, 3).astype("int"))
if transparency:
color_masks.append(
resize(images[0][:, :, 3], (16, 16), anti_aliasing=False))
# concatenate to a single matrix
examples = np.concatenate(examples)
if transparency:
# transparent colors should be ignored
color_masks = np.concatenate(color_masks).ravel()
examples = examples[color_masks >= self.alpha]
self._fit_model(examples)
def _fit_model(self, X):
"""Fit model while suppressing warnings from sklearn"""
converge = True
with warnings.catch_warnings(record=True) as w:
# fit model
self.model.fit(X)
if w and w[-1].category == ConvergenceWarning:
warnings.filterwarnings('ignore', category=ConvergenceWarning)
converge = False
if not converge:
warnings.warn(
"the model has failed to converge, try a different number of colors for better results!",
Warning)
self.is_fitted = True
def _reduce(self, image):
"""Apply convolutions on image ITER times and generate a smaller image
based on the highest magnitude of gradients"""
# self is visible to decorated function
@adapt_rgb(each_channel)
def _wrapper(dim):
# apply median filter for noise reduction
dim = median(dim, square(4))
for n in range(self.ITER):
h, w = dim.shape
h, w = h // 2, w // 2
tt = time()
flatten = view_as_blocks(dim, (2, 2)).reshape(-1, 2, 2)
print('flatten:', time() - tt)
# bottleneck
tt = time()
# slow one upper
new_image = np.fromiter(
(self._reduce_conv(f) for f in flatten),
flatten.dtype).reshape((h, w))
print('bottleneck:', time() - tt)
if n < self.ITER - 1:
dim = new_image.copy()
return new_image
return _wrapper(image)
def _reduce_conv(self, f):
# slow one lower
"""The actual function that selects the right pixels based on the gradients 2x2 square"""
return np.mean(f[self.SOLUTIONS[np.argmax(
np.sum(np.multiply(self.CONVOLUTIONS, f.reshape(-1, 2,
2)).reshape(-1, 4),
axis=1))]])
def _dilate(self, image):
"""Dilate semi-transparent edges to remove artifacts
(unwanted edges, caused by transparent pixels having different colors)"""
@adapt_rgb(each_channel)
def _wrapper(dim):
return dilation(dim, selem=square(4))
# use dilated pixels for semi-transparent ones
mask = image[:, :, 3]
alter = _wrapper(image[:, :, :3])
image[:, :, :3][mask < self.alpha] = alter[mask < self.alpha]
return image
@staticmethod
def _fix_hist(image):
"""Apply adaptive histogram"""
image = equalize_adapthist(
image) * 255 * 1.14 # empirical magic number
image[image <= 8.] = 0.
return image
@staticmethod
def _is_transparent(image):
"""Returns True if there is an additional dimension for transparency"""
return bool(image.shape[2] == 4)
def all_classifier_models():
models = []
metrix = []
c_report = []
train_accuracy = []
test_accuracy = []
models.append(('LogisticRegression', LogisticRegression(solver='liblinear', multi_class='ovr')))
models.append(('LinearDiscriminantAnalysis', LinearDiscriminantAnalysis()))
models.append(('KNeighborsClassifier', KNeighborsClassifier()))
models.append(('DecisionTreeClassifier', DecisionTreeClassifier()))
models.append(('GaussianNB', GaussianNB()))
models.append(('RandomForestClassifier', RandomForestClassifier(n_estimators=100)))
models.append(('SVM', SVC(gamma='auto')))
models.append(('Linear_SVM', LinearSVC()))
models.append(('XGB', XGBClassifier()))
models.append(('SGD', SGDClassifier()))
models.append(('Perceptron', Perceptron()))
models.append(('ExtraTreeClassifier', ExtraTreeClassifier()))
models.append(('OneClassSVM', OneClassSVM(gamma = 'auto')))
models.append(('NuSVC', NuSVC()))
models.append(('MLPClassifier', MLPClassifier(solver='lbfgs', alpha=1e-5, random_state=1)))
models.append(('RadiusNeighborsClassifier', RadiusNeighborsClassifier(radius=2.0)))
models.append(('OutputCodeClassifier', OutputCodeClassifier(estimator=RandomForestClassifier(random_state=0),random_state=0)))
models.append(('OneVsOneClassifier', OneVsOneClassifier(estimator = RandomForestClassifier(random_state=1))))
models.append(('OneVsRestClassifier', OneVsRestClassifier(estimator = RandomForestClassifier(random_state=1))))
models.append(('LogisticRegressionCV', LogisticRegressionCV()))
models.append(('RidgeClassifierCV', RidgeClassifierCV()))
models.append(('RidgeClassifier', RidgeClassifier()))
models.append(('PassiveAggressiveClassifier', PassiveAggressiveClassifier()))
models.append(('GaussianProcessClassifier', GaussianProcessClassifier()))
models.append(('HistGradientBoostingClassifier', HistGradientBoostingClassifier()))
estimators = [('rf', RandomForestClassifier(n_estimators=10, random_state=42)),('svr', make_pipeline(StandardScaler(),LinearSVC(random_state=42)))]
models.append(('StackingClassifier', StackingClassifier(estimators=estimators, final_estimator=LogisticRegression())))
clf1 = LogisticRegression(multi_class='multinomial', random_state=1)
clf2 = RandomForestClassifier(n_estimators=50, random_state=1)
clf3 = GaussianNB()
models.append(('VotingClassifier', VotingClassifier(estimators=[('lr', clf1), ('rf', clf2), ('gnb', clf3)], voting='hard')))
models.append(('AdaBoostClassifier', AdaBoostClassifier()))
models.append(('GradientBoostingClassifier', GradientBoostingClassifier()))
models.append(('BaggingClassifier', BaggingClassifier()))
models.append(('ExtraTreesClassifier', ExtraTreesClassifier()))
models.append(('CategoricalNB', CategoricalNB()))
models.append(('ComplementNB', ComplementNB()))
models.append(('BernoulliNB', BernoulliNB()))
models.append(('MultinomialNB', MultinomialNB()))
models.append(('CalibratedClassifierCV', CalibratedClassifierCV()))
models.append(('LabelPropagation', LabelPropagation()))
models.append(('LabelSpreading', LabelSpreading()))
models.append(('NearestCentroid', NearestCentroid()))
models.append(('QuadraticDiscriminantAnalysis', QuadraticDiscriminantAnalysis()))
models.append(('GaussianMixture', GaussianMixture()))
models.append(('BayesianGaussianMixture', BayesianGaussianMixture()))
test_accuracy= []
names = []
for name, model in models:
try:
m = model
m.fit(X_train, y_train)
y_pred = m.predict(X_test)
train_acc = round(m.score(X_train, y_train) * 100, 2)
test_acc = metrics.accuracy_score(y_test,y_pred) *100
c_report.append(classification_report(y_test, y_pred))
test_accuracy.append(test_acc)
names.append(name)
metrix.append([name, train_acc, test_acc])
except:
print("Exception Occurred :",name)
return metrix,test_accuracy,names
## Test BGM for one appliance appl = 'WHE' # Create vector with P and Q values and plot them P = d[appl].P[init:end].values Q = d[appl].Q[init:end].values X = np.transpose([P, Q]) plt.plot(d[appl].P[init:end], d[appl].Q[init:end],'o', alpha=0.1) # Normalize X sscl = StandardScaler().fit(X) X = sscl.transform(X) # Apply clusterer bgm = BayesianGaussianMixture(n_components=33, covariance_type='full', weight_concentration_prior_type='dirichlet_distribution', random_state=42).fit(X) y_pred = bgm.predict(X) # Plot clusters with X unnormalized X = sscl.inverse_transform(X) plt.figure() plt.scatter(X[:,0],X[:,1], color=colors[y_pred]) means = sscl.inverse_transform(bgm.means_) medians = get_medians(X, y_pred) # plt.plot(means[:,0],means[:,1],'kx') plt.plot(medians[:,0],medians[:,1],'kx') # TODO: Compare mean with ground truth plt.figure() plt.plot(P)
def em_stereo(self,n_component=1,dp=True,thresh_hold=0.4):
self.num_params = 0
#The range of len(params)
_step = 0
for var_idx in tqdm(range(len(self.merge_var[0]))):
for x_v in range(len(self.merge_var[0][var_idx])):
print('Step %d'%_step,end='\r')
_step += 1
try:
for y_v in range(len(self.merge_var[0][var_idx][x_v])):
#print('cluster weights ....%d'%var_idx)
dist = []
for task_idx in range(len(self.merge_var)):
nor = np.random.normal(self.merge_var[task_idx][var_idx][x_v][y_v],np.log(1.0+np.exp(self.merge_uncertainty[task_idx][var_idx][x_v][y_v])),200)
dist.append(nor)
dist = np.array(np.asmatrix(np.concatenate(dist)).T)
if dp:
print('Initializing DPGMM%d ... '%_step,end='\r')
gmm = DPGMM( max_iter=1000, n_components=n_component, covariance_type='spherical')
else:
gmm = GMM( max_iter=200, n_components=n_component, covariance_type='spherical')
gmm.fit(dist)
new_idx_list = []
for task_idx in range(len(self.merge_var)):
#if dp:
#Strategy 1. Set threshold
predict_probability = gmm.predict_proba(np.array(self.merge_var[task_idx][var_idx][x_v][y_v]).reshape(-1,1))
f_ = True
while f_:
#if gmm.weights_[np.argmax(predict_probability)] > ( 1 / len(self.merge_var)):
if gmm.weights_[np.argmax(predict_probability)] > thresh_hold:
new_idx = np.argmax(predict_probability)
f_ = False
else:
predict_probability[0][np.argmax(predict_probability)] = 0.0
self.num_params += 1
#else:
# new_idx = gmm.predict(np.array(self.merge_var[task_idx][var_idx][x_v][y_v]).reshape(-1,1))
# if new_idx in new_idx_list:
self.num_params += 1
new_idx_list.append(new_idx)
self.merge_var[task_idx][var_idx][x_v][y_v] = gmm.means_[new_idx]
self.merge_uncertainty[task_idx][var_idx][x_v][y_v] = np.log(np.exp(gmm.covariances_[new_idx]) - 1.0)
except TypeError:
dist = []
for task_idx in range(len(self.merge_var)):
nor = np.random.normal(self.merge_var[task_idx][var_idx][x_v],np.log(1.0+np.exp(self.merge_uncertainty[task_idx][var_idx][x_v])),200)
dist.append(nor)
dist = np.array(np.asmatrix(np.concatenate(dist)).T)
if dp:
print('Initializing DPGMM%d ... '%_step,end='\r')
gmm = DPGMM( max_iter=200, n_components=n_component, covariance_type='spherical')
else:
gmm = GMM( max_iter=200, n_components=n_component, covariance_type='spherical')
gmm.fit(dist)
new_idx_list = []
for task_idx in range(len(self.merge_var)):
#if dp:
#Strategy 1. Set threshold
predict_probability = gmm.predict_proba(np.array(self.merge_var[task_idx][var_idx][x_v]).reshape(-1,1))
f_ = True
while f_:
#if gmm.weights_[np.argmax(predict_probability)] > ( 1 / len(self.merge_var)):
if gmm.weights_[np.argmax(predict_probability)] > thresh_hold:
new_idx = np.argmax(predict_probability)
f_ = False
else:
predict_probability[0][np.argmax(predict_probability)] = 0.0
self.num_params += 1
#else:
# new_idx = gmm.predict(np.array(self.merge_var[task_idx][var_idx][x_v]).reshape(-1,1))
# if new_idx in new_idx_list:
# self.num_params += 1
new_idx_list.append(new_idx)
self.merge_var[task_idx][var_idx][x_v] = gmm.means_[new_idx]
self.merge_uncertainty[task_idx][var_idx][x_v] = np.log(np.exp(gmm.covariances_[new_idx]) - 1.0)
def _make_bgmm(self, label, X, Y):
indices = np.where(Y == label)
bgm = BayesianGaussianMixture(n_components=3, max_iter=200, tol=1e-3)
bgm.fit(X[indices])
return bgm
from sklearn.mixture import GaussianMixture
"""
GMM 高斯混合模型
n_components 高斯分布的个数
covariance_type 各个高斯分布的方差关系，
full -> 各个分布之间没有关系
"""
data = []
g = GaussianMixture(n_components=2,
covariance_type='full',
tol=1e-6,
max_iter=1000)
g.fit(data)
print('类别概率:\t', g.weights_[0])
print('均值:\n', g.means_, '\n')
print('方差:\n', g.covariances_, '\n')
# DPGMM 可以解决自动调整 n_components 的问题
from sklearn.mixture import BayesianGaussianMixture
dpgmm = BayesianGaussianMixture(
n_components=3,
covariance_type='full',
max_iter=1000,
n_init=5,
weight_concentration_prior_type='dirichlet_process',
weight_concentration_prior=10)
from sklearn.mixture import BayesianGaussianMixture
from sklearn.linear_model import LogisticRegression
dataset = loaddataset(train_file)
testset = loaddataset(test_file)
names = [
"Nearest Neighbors",
"RBF SVM",
"Decision Tree",
"Random Forest",
"AdaBoost",
]
ab = AdaBoostClassifier(random_state=1)
bgm = BayesianGaussianMixture(random_state=1)
dt = DecisionTreeClassifier(random_state=1)
gb = GradientBoostingClassifier(random_state=1)
lr = LogisticRegression(random_state=1)
rf = RandomForestClassifier(random_state=1)
classifiers = [
KNeighborsClassifier(3),
GaussianProcessClassifier(1.0 * RBF(1.0), warm_start=True),
RandomForestClassifier(random_state=1),
GaussianNB(),
QuadraticDiscriminantAnalysis()
]
svcl = LinearSVC(random_state=1)
svcg = SVC(random_state=1)
alpha=0.5,
clip_box=ax.bbox)
ax.add_artist(e)
ax1_min, ax1_max, ax2_min, ax2_max = plt.axis()
plt.xlim((x1_min, x1_max))
plt.ylim((x2_min, x2_max))
plt.title("GMM", fontsize=18)
plt.grid(True)
#DPGMM
dpgmm = BayesianGaussianMixture(
n_components=n_compenents,
covariance_type="full",
max_iter=100,
n_init=5,
weight_concentration_prior_type="dirichlet_process",
weight_concentration_prior=100)
dpgmm.fit(x)
centers = dpgmm.means_
covs = dpgmm.covariances_
print("DPGMM均值=：\n", centers)
print("DPGMM方差=：\n", covs)
y_hat = dpgmm.predict(x)
ax = plt.subplot(212)
grid_hat = dpgmm.predict(grid_test)
grid_hat = grid_hat.reshape(x1.shape)
plt.pcolormesh(x1, x2, grid_hat, cmap=cm)
plt.scatter(x[:, 0], x[:, 1], c=y, s=30, cmap=cm, marker='o')
for k, e in enumerate(ells):
a.add_artist(e)
e.set_facecolor((1 - w[k], 1 - w[k], 1 - w[k]))
a.set_xlim(np.min(x1) - np.min(x1) / 10, np.max(x1) + np.min(x1) / 10)
a.set_ylim(np.min(x2) - np.min(x2) / 10, np.max(x2) + np.min(x2) / 10)
plt.scatter(x1, x2)
plt.savefig('plot_{}.png'.format(iteration))
plt.show()
plt.close()
estimators = [
("Infinite mixture with a Dirichlet process\n prior and" r"$\gamma_0=$",
BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_process",
n_components=2 * 5, reg_covar=0, init_params='random',
max_iter=10, mean_precision_prior=.8,
random_state=3), [1], 10),
("Infinite mixture with a Dirichlet process\n prior and" r"$\gamma_0=$",
BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_process",
n_components=2 * 5, reg_covar=0, init_params='random',
max_iter=15, mean_precision_prior=.8,
random_state=3), [1], 15),
("Infinite mixture with a Dirichlet process\n prior and" r"$\gamma_0=$",
BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_process",
n_components=2 * 5, reg_covar=0, init_params='random',
max_iter=20, mean_precision_prior=.8,
random_state=3), [1], 20),
("Infinite mixture with a Dirichlet process\n prior and" r"$\gamma_0=$",
x_train_unlabeled = train_unlabeled
#Switch to numpy
# Preprocessing X
x_train = []
x_train_labeled = np.array(x_train_labeled)
x_train_unlabeled = np.array(x_train_unlabeled)
x_train.extend(x_train_labeled)
x_train.extend(x_train_unlabeled)
x_test = np.array(test)
# Preprocessing y
y_train_labeled = np.array(y_train_labeled)
ones = -1 * np.ones(21000)
ones = np.array(ones)
y_train = np.concatenate((y_train_labeled, ones)).astype(int)
# # GMM
# GMM = BayesianGaussianMixture(n_components=10, random_state=0)
# GMM.fit(x_train, y_train)
# y_pred = GMM.predict(x_test)
# Bayesian GMM
BGMM = BayesianGaussianMixture(n_components=10, random_state=0)
BGMM.fit(x_train, y_train)
y_pred = BGMM.predict(x_test)
# output results
d = {'Id': test.index, 'y': y_pred}
output = pd.DataFrame(d)
output.to_csv('output5.csv', index=False)
def test_compare_covar_type():
# We can compare the 'full' precision with the other cov_type if we apply
# 1 iter of the M-step (done during _initialize_parameters).
rng = np.random.RandomState(0)
rand_data = RandomData(rng, scale=7)
X = rand_data.X['full']
n_components = rand_data.n_components
for prior_type in PRIOR_TYPE:
# Computation of the full_covariance
bgmm = BayesianGaussianMixture(
weight_concentration_prior_type=prior_type,
n_components=2 * n_components,
covariance_type='full',
max_iter=1,
random_state=0,
tol=1e-7)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X, np.random.RandomState(0))
full_covariances = (
bgmm.covariances_ *
bgmm.degrees_of_freedom_[:, np.newaxis, np.newaxis])
# Check tied_covariance = mean(full_covariances, 0)
bgmm = BayesianGaussianMixture(
weight_concentration_prior_type=prior_type,
n_components=2 * n_components,
covariance_type='tied',
max_iter=1,
random_state=0,
tol=1e-7)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X, np.random.RandomState(0))
tied_covariance = bgmm.covariances_ * bgmm.degrees_of_freedom_
assert_almost_equal(tied_covariance, np.mean(full_covariances, 0))
# Check diag_covariance = diag(full_covariances)
bgmm = BayesianGaussianMixture(
weight_concentration_prior_type=prior_type,
n_components=2 * n_components,
covariance_type='diag',
max_iter=1,
random_state=0,
tol=1e-7)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X, np.random.RandomState(0))
diag_covariances = (bgmm.covariances_ *
bgmm.degrees_of_freedom_[:, np.newaxis])
assert_almost_equal(
diag_covariances,
np.array([np.diag(cov) for cov in full_covariances]))
# Check spherical_covariance = np.mean(diag_covariances, 0)
bgmm = BayesianGaussianMixture(
weight_concentration_prior_type=prior_type,
n_components=2 * n_components,
covariance_type='spherical',
max_iter=1,
random_state=0,
tol=1e-7)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X, np.random.RandomState(0))
spherical_covariances = bgmm.covariances_ * bgmm.degrees_of_freedom_
assert_almost_equal(spherical_covariances,
np.mean(diag_covariances, 1))
# Plot the latent space
latent_vectors = chain_call(m1.latent_sample, X_test, 1000)
#%%
# verify sklearn gaussian mixture?
from sklearn.cluster import KMeans
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
from matplotlib import pyplot as plt
from sklearn.mixture import BayesianGaussianMixture
pca = PCA(2)
# pca = TSNE(2)
X_pca = pca.fit_transform(latent_vectors)
kmeans = BayesianGaussianMixture(10, tol=1e-6, max_iter=1000)
pred = kmeans.fit_predict(X_pca)
print(purity_score(y_test, pred))
#%%
df_latent = pd.DataFrame({
"x1": X_pca[:, 0],
"x2": X_pca[:, 1],
"cat": ["pred_{}".format(i) for i in y_test],
"kmeans": ["pred_{}".format(i) for i in pred],
})
plt.figure(figsize=(10, 10))
sns.scatterplot(data=df_latent, x="x1", y="x2", hue="cat")
plt.figure(figsize=(10, 10))
sns.scatterplot(data=df_latent, x="x1", y="x2", hue="kmeans")
def gets_best_model(self, X, target):
best_classifiers = []
outer_cv = StratifiedKFold(n_splits=self.num_folds,
shuffle=True,
random_state=1)
model_factory = [
AdaBoostClassifier(),
BaggingClassifier(),
BayesianGaussianMixture(),
BernoulliNB(),
CalibratedClassifierCV(),
CatBoostClassifier(verbose=False),
DecisionTreeClassifier(),
ExtraTreesClassifier(),
GaussianMixture(),
GaussianNB(),
GradientBoostingClassifier(),
KNeighborsClassifier(),
LinearDiscriminantAnalysis(),
LogisticRegression(max_iter=1000),
LogisticRegressionCV(max_iter=1000),
MLPClassifier(),
QuadraticDiscriminantAnalysis(),
RandomForestClassifier(),
SGDClassifier()
]
logging.basicConfig(filename="ml_dl_toolbox_logfilename.log",
level=logging.INFO,
format='%(asctime)s %(levelname)-8s %(message)s',
datefmt='%Y-%m-%d %H:%M:%S')
scoring = ('accuracy', 'neg_mean_squared_error')
try:
for el in model_factory:
el.seed = self.seed
scores = cross_validate(el,
X.drop(target, axis=1),
X[target],
cv=outer_cv,
n_jobs=-1,
scoring=scoring)
scores = abs(
np.sqrt(
np.mean(scores['test_neg_mean_squared_error']) *
-1)) / np.mean(scores['test_accuracy'])
score_description = [
el, '{el}'.format(el=el.__class__.__name__),
"%0.5f" % scores
]
best_classifiers.append(score_description)
best_model = pd.DataFrame(
best_classifiers,
columns=["Algorithm", "Model",
"RMSE/Accuracy"]).sort_values("RMSE/Accuracy",
axis=0,
ascending=True)
best_model = best_model.reset_index()
except OSError:
logging.error('Check data structure')
else:
logging.info('Best fitting algorithm: ' + best_model["Model"][0] +
" RMSE/Accuracy: " + best_model["RMSE/Accuracy"][0])
return best_model["Algorithm"][0]
np.random.seed(1)
X = np.c_[data_thr.orbit, data_thr.rate, data_thr.rateA, data_thr.rateB,
data_thr.rateC, data_thr.rateCA]
scaler = StandardScaler()
X = scaler.fit_transform(X)
# 1 corresponds to data_thr.rate and 4=5-1 to data_thr.rateC
w = w / np.sqrt(scaler.var_[1:])
# w = np.exp(-np.exp(3 * w.mean(axis=1)))
w = 1. / w.mean(axis=1) ** 2
Html_file = open("gmm_sklearn_files/gmm3_sklearn.html", "w")
gmm = BayesianGaussianMixture(n_components=3, alpha_prior=0.1, beta_prior=1,
n_init=5)
gmm.fit(X) # , weights=w) not implemented in sklearn yet
preds = gmm.predict(X)
probs = gmm.predict_proba(X)
data_thr['preds'] = pd.Series(preds).astype("category")
color_key = ["red", "blue", "yellow", "grey", "black", "purple", "pink",
"brown", "green", "orange"] # Spectral9
color_key = color_key[:len(set(preds))+1]
covs = gmm.covariances_
means = gmm.means_
# transform cov for non-standardizeed data:
covs = np.array([np.dot(np.diag(np.sqrt(scaler.var_)),
def main_process(img_np: np.ndarray, ccs: List[np.ndarray],
text_lines: List[Tuple[int, int, int,
int]], cc2textline_assignment):
if len(ccs) == 0:
return
final_mask = np.zeros_like(ccs[0])
dpgmm = BayesianGaussianMixture(n_components=5, covariance_type='diag')
kern = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
text_line_colors = defaultdict(list)
#print(cc2textline_assignment)
for i, cc in enumerate(tqdm(ccs)):
if np.sum(cv2.bitwise_and(cc, final_mask)) > 0:
final_mask = cv2.bitwise_or(final_mask, cc)
continue
pixels = img_np[cv2.erode(cc, kern) > 127] #img_np[cc > 127]#
if len(pixels) < 5:
final_mask = cv2.bitwise_or(final_mask, cc)
continue
cls1 = dpgmm.fit(pixels)
# print(cls1.means_)
# print(cls1.weights_)
# print(np.sqrt(cls1.covariances_))
cls1_top2_mean, cls1_top2_stddev, cls1_k = find_top_k_dpgmm(cls1, 2)
# cls1_top2_mean[0][0] = 4
# cls1_top2_mean[0][1] = 4
# cls1_top2_mean[0][2] = 4
# cls1_top2_mean[1][0] = 253
# cls1_top2_mean[1][1] = 253
# cls1_top2_mean[1][2] = 253
# cls1_top2_stddev[0][0] = 5
# cls1_top2_stddev[0][1] = 5
# cls1_top2_stddev[0][2] = 5
# cls1_top2_stddev[1][0] = 5
# cls1_top2_stddev[1][1] = 5
# cls1_top2_stddev[1][2] = 5
# if cls1_top2_mean[0][0] > 100 :
# breakpoint()
cls1_top2_stddev_ext = np.round(cls1_top2_stddev * COLOR_RANGE_SIGMA)
# if i == 2 :
# breakpoint()
top1_mask = extend_cc_region(img_np, cc, cls1_top2_mean[0],
cls1_top2_stddev_ext[0])
top2_mask = extend_cc_region(img_np, cc, cls1_top2_mean[1],
cls1_top2_stddev_ext[1])
# area1 = int(top1_mask.sum())
# area2 = int(top2_mask.sum())
# if area1 == 0 or area2 == 0 :
# breakpoint()
# continue
# area_cc = int(cc.sum())
# if abs(area1 - area_cc) < abs(area2 - area_cc) :
# D = top1_mask
# selected_idx = 0
# else :
# D = top2_mask
# selected_idx = 1
# intersect_area1 = int(cv2.bitwise_and(cc, top1_mask).sum())
# intersect_area2 = int(cv2.bitwise_and(cc, top2_mask).sum())
iou1 = cv2.bitwise_and(cc, top1_mask).sum() / cv2.bitwise_or(
cc, top1_mask).sum()
iou2 = cv2.bitwise_and(cc, top2_mask).sum() / cv2.bitwise_or(
cc, top2_mask).sum()
if iou1 > iou2:
D = top1_mask
selected_idx = 0
if iou1 < 1e-1:
#print(iou1)
D = cc
selected_idx = -1
else:
D = top2_mask
selected_idx = 1
if iou2 < 1e-1:
#print(iou2)
D = cc
selected_idx = -1
# print(selected_idx, iou1, iou2)
# save_rgb('text_mask_utils_tmp.png', cc)
# save_rgb('text_mask_utils_tmp_color.png', img_np*(cc>0)[:,:,None])
# save_rgb('text_mask_utils_tmp_top2_mask.png', top2_mask)
# save_rgb('text_mask_utils_tmp_top1_mask.png', top1_mask)
# input('x')
# if cc2textline_assignment[i] == 12 :
# breakpoint()
D = cv2.bitwise_or(cc, D)
D = cv2.dilate(D, kern)
# if cls1_top2_mean[selected_idx][0] < 100 :
# breakpoint()
final_mask = cv2.bitwise_or(final_mask, D)
# seed_point_candidates_mask = cv2.inRange(cc_region, cls1_top2_mean[0] - cls1_top2_stddev_ext[0], cls1_top2_mean[0] + cls1_top2_stddev_ext[0])
# seed_point_candidates_y, seed_point_candidates_x = np.where(seed_point_candidates_mask > 127)
# seed_point = (seed_point_candidates_x[0] + x, seed_point_candidates_y[0] + y)
# D = np.zeros((cc.shape[0] + 2, cc.shape[1] + 2), dtype = np.uint8)
# cv2.floodFill(img_np, D, seed_point, (255), cls1_top2_stddev_ext[0].tolist(), cls1_top2_stddev_ext[0].tolist(), cv2.FLOODFILL_MASK_ONLY)
# D = D[1: -1, 1: -1] * 255
# final_mask = cv2.bitwise_or(final_mask, D)
# now we find text color
if selected_idx == -1:
continue # skip
text_color_value = cls1_top2_mean[selected_idx]
text_color_stddev = cls1_top2_stddev[selected_idx]
#print('color=', text_color_value, text_color_stddev)
text_line_colors[cc2textline_assignment[i]].append(text_color_value)
textline_img = np.copy(img_np)
def get_textline_color(j, visited_text_lines: List[int]):
visited_text_lines.append(j)
(x, y, w, h) = text_lines[j]
colors = text_line_colors[j]
if not colors:
#print(f'[!!!!!!!!!!!!] Textline {j} has no color assigned, seraching for closest')
min_dist = 10000000000000000000
min_dist_k = -1
for k, (x2, y2, w2, h2) in enumerate(text_lines):
d = rect_distance(x, y, x + w, y + h, x2, y2, x2 + w2, y2 + h2)
if d < min_dist and k not in visited_text_lines:
min_dist = d
min_dist_k = k
if min_dist_k == -1:
print(
f'[!!!!!!!!!!!!] Textline {j} has no color assigned and unable to find closest rectangle, defaulting to black'
)
return np.zeros((3, ), dtype=np.uint8)
return get_textline_color(min_dist_k, visited_text_lines)
else:
clr = np.round(np.mean(np.array(colors), axis=0))
return clr
textline_colors = []
for j, (x, y, w, h) in enumerate(text_lines):
# colors = text_line_colors[j]
# if not colors :
# print(f'[!!!!!!!!!!!!] Textline {j} has no color assigned')
# cv2.rectangle(textline_img, (x, y), (x + w, y + h), (0, 255, 0), 2)
# continue
clr = get_textline_color(j, []).tolist()
textline_colors.append(clr)
cv2.rectangle(textline_img, (x, y), (x + w, y + h), clr, 2)
#save_rgb('text_mask_utils_final_textlines.png', textline_img)
#save_rgb('text_mask_utils_final_mask_masked.png', (final_mask > 127)[:,:,None] * img_np)
#cv2.imwrite('text_mask_utils_final_mask.png', final_mask)
return final_mask, textline_colors
'energy': np.double,
'liveness': np.double,
'loudness': np.double,
'mode': np.double,
'speechiness': np.double,
'tempo': np.double,
'valence': np.double,
'instrumentalness': np.double,
}
df = df.astype(types)
X = df.drop(labels=['id ', 'artist', 'name', 'mode', 'tempo', 'loudness'],
axis=1)
# Initial model to get appropriate number of clusters
bgmm = BayesianGaussianMixture(
n_components=20,
n_init=20,
)
bgmm.fit(X)
num_clusters = len([w for w in bgmm.weights_ if w > 0.05])
bgmm_2 = BayesianGaussianMixture(n_components=num_clusters, n_init=20)
probs = pd.DataFrame(bgmm_2.fit(X).predict_proba(X))
results = pd.concat([df, probs], axis=1, sort=False)
# inital results - fluctuate a lot and also the weightings seem to do bugger all.
# TODO:
# - REAAALLLY have a look at what features are actually important. instrumentallness is not one of them...
# - DBSCAN is probs a pretty good way of getting the clusters actually - do a minimum value of like 10 songs
def cluster(idx_arr,
h_arr,
w_arr,
data,
cw,
samplerate,
minp,
percentile,
npc,
ngaus,
plot_steps=False):
l = np.ones(h_arr.shape)
if ((np.max(h_arr) - np.min(h_arr)) / np.max(h_arr)) > 0.25:
# classify by height
bgm = BayesianGaussianMixture(ngaus, max_iter=100, n_init=10)
l = bgm.fit_predict(h_arr.reshape(-1, 1))
# if any of the clusters merged have very little height difference, merge them.
if len(np.unique(l)) > 1:
for ll in np.unique(l):
if ((np.max(h_arr[l != ll]) - np.min(h_arr[l != ll])) /
np.max(h_arr[l != ll])) < 0.25:
l[l != ll] = np.max(l) + 1
if plot_steps == True:
mean_eods, eod_times, _ = find_window(data,
idx_arr,
l,
h_arr,
rm=False)
print('clustering based on hight')
plot_all(data, ts, samplerate, ms, vs)
# now cluster based on waveform
al = np.ones(len(l)) * -1
# extract snippets
snippets = np.stack([
data[int(idx - cw * samplerate / 2):int(idx + cw * samplerate / 2)]
for idx in idx_arr
])
# keep track of the labels so that no labels are overwritten
maxlab = 0
for hl in np.unique(l):
if len(l[l == hl]) > minp:
# extract snippets, idxs and hs for this hight cluster
csnippets = StandardScaler().fit_transform(snippets[l == hl])
cidx_arr = idx_arr[l == hl]
ch_arr = h_arr[l == hl]
# extract relevant snippet features
pca = PCA(npc).fit(csnippets).transform(csnippets)
# determine good epsilon
knn = np.sort(pairwise_distances(pca, pca))[:, minp]
eps = np.percentile(knn, percentile)
# cluster by EOD shape
c = DBSCAN(eps=eps, min_samples=minp).fit(pca).labels_
if plot_steps == True:
mean_eods, eod_times, _ = find_window(data,
cidx_arr,
c,
ch_arr,
rm=False)
print('clustering on scaled eods')
plot_all(data, ts, samplerate, ms, vs)
# cluster again without scaling (sometimes this works better wrt scaling)
csnippets_ns = snippets[l == hl]
pca = PCA(npc).fit(csnippets_ns).transform(csnippets_ns)
knn = np.sort(pairwise_distances(pca, pca))[:, minp]
eps = np.percentile(knn, percentile)
c_ns = DBSCAN(eps=eps, min_samples=minp).fit(pca).labels_
if plot_steps == True:
mean_eods, eod_times = find_window(data,
cidx_arr,
c_ns,
ch_arr,
rm=False)
print('clustering on non-scaled eods')
plot_all(data, ts, samplerate, ms, vs)
# merge results for scaling and without scaling
_, _, _, c = merge_clusters(c, c_ns, cidx_arr, cidx_arr, ch_arr,
data, samplerate)
if plot_steps == True:
mean_eods, eod_times = find_window(data,
cidx_arr,
c,
ch_arr,
rm=False)
print('merged scale and non-scaled')
plot_all(data, ts, samplerate, ms, vs)
# update maxlab so that no clusters are overwritten
c[c == -1] = -maxlab - 1
al[l == hl] = c + maxlab
maxlab = np.max(al) + 1
# return the overall clusters (al) and the clusters based on hight (l)
return al, l
value, vector = sp.linalg.eigh(cov)
width, height = value[0], value[1]
v = vector[0] / sp.linalg.norm(vector[0])
angle = 180* np.arctan(v[1] / v[0]) / np.pi
e = Ellipse(xy=center, width=width, height=height,
angle=angle, color=clrs[i], alpha=0.5, clip_box = ax.bbox)
ax.add_artist(e)
ax1_min, ax1_max, ax2_min, ax2_max = plt.axis()
plt.xlim((x1_min, x1_max))
plt.ylim((x2_min, x2_max))
plt.title('GMM', fontsize=15)
plt.grid(b=True, ls=':', color='#606060')
# DPGMM
dpgmm = BayesianGaussianMixture(n_components=n_components, covariance_type='full', max_iter=1000, n_init=5,
weight_concentration_prior_type='dirichlet_process', weight_concentration_prior=0.1)
dpgmm.fit(x)
centers = dpgmm.means_
covs = dpgmm.covariances_
print('DPGMM均值 = \n', centers)
print('DPGMM方差 = \n', covs)
y_hat = dpgmm.predict(x)
print(y_hat)
ax = plt.subplot(212)
grid_hat = dpgmm.predict(grid_test)
grid_hat = grid_hat.reshape(x1.shape)
plt.pcolormesh(x1, x2, grid_hat, cmap=cm)
plt.scatter(x[:, 0], x[:, 1], s=20, c=y, cmap=cm, marker='o', edgecolors='#202020')
for i, cc in enumerate(zip(centers, covs)):
# Parameters of the dataset
random_state, n_components, n_features = 2, 3, 2
colors = np.array(['#0072B2', '#F0E442', '#D55E00'])
covars = np.array([[[.7, .0], [.0, .1]], [[.5, .0], [.0, .1]],
[[.5, .0], [.0, .1]]])
samples = np.array([200, 500, 200])
means = np.array([[.0, -.70], [.0, .0], [.0, .70]])
# mean_precision_prior= 0.8 to minimize the influence of the prior
estimators = [("Finite mixture with a Dirichlet distribution\nprior and "
r"$\gamma_0=$",
BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_distribution",
n_components=2 * n_components,
reg_covar=0,
init_params='random',
max_iter=1500,
mean_precision_prior=.8,
random_state=random_state), [0.001, 1, 1000]),
("Infinite mixture with a Dirichlet process\n prior and"
r"$\gamma_0=$",
BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_process",
n_components=2 * n_components,
reg_covar=0,
init_params='random',
max_iter=1500,
mean_precision_prior=.8,
random_state=random_state), [1, 1000, 100000])]
# Generate data
def test_bayesian_mixture_precisions_prior_initialisation():
rng = np.random.RandomState(0)
n_samples, n_features = 10, 2
X = rng.rand(n_samples, n_features)
# Check raise message for a bad value of degrees_of_freedom_prior
bad_degrees_of_freedom_prior_ = n_features - 1.0
bgmm = BayesianGaussianMixture(degrees_of_freedom_prior=bad_degrees_of_freedom_prior_, random_state=rng)
assert_raise_message(
ValueError,
"The parameter 'degrees_of_freedom_prior' should be "
"greater than %d, but got %.3f." % (n_features - 1, bad_degrees_of_freedom_prior_),
bgmm.fit,
X,
)
# Check correct init for a given value of degrees_of_freedom_prior
degrees_of_freedom_prior = rng.rand() + n_features - 1.0
bgmm = BayesianGaussianMixture(degrees_of_freedom_prior=degrees_of_freedom_prior, random_state=rng).fit(X)
assert_almost_equal(degrees_of_freedom_prior, bgmm.degrees_of_freedom_prior_)
# Check correct init for the default value of degrees_of_freedom_prior
degrees_of_freedom_prior_default = n_features
bgmm = BayesianGaussianMixture(degrees_of_freedom_prior=degrees_of_freedom_prior_default, random_state=rng).fit(X)
assert_almost_equal(degrees_of_freedom_prior_default, bgmm.degrees_of_freedom_prior_)
# Check correct init for a given value of covariance_prior
covariance_prior = {
"full": np.cov(X.T, bias=1) + 10,
"tied": np.cov(X.T, bias=1) + 5,
"diag": np.diag(np.atleast_2d(np.cov(X.T, bias=1))) + 3,
"spherical": rng.rand(),
}
bgmm = BayesianGaussianMixture(random_state=rng)
for cov_type in ["full", "tied", "diag", "spherical"]:
bgmm.covariance_type = cov_type
bgmm.covariance_prior = covariance_prior[cov_type]
bgmm.fit(X)
assert_almost_equal(covariance_prior[cov_type], bgmm.covariance_prior_)
# Check raise message for a bad spherical value of covariance_prior
bad_covariance_prior_ = -1.0
bgmm = BayesianGaussianMixture(
covariance_type="spherical", covariance_prior=bad_covariance_prior_, random_state=rng
)
assert_raise_message(
ValueError,
"The parameter 'spherical covariance_prior' "
"should be greater than 0., but got %.3f." % bad_covariance_prior_,
bgmm.fit,
X,
)
# Check correct init for the default value of covariance_prior
covariance_prior_default = {
"full": np.atleast_2d(np.cov(X.T)),
"tied": np.atleast_2d(np.cov(X.T)),
"diag": np.var(X, axis=0, ddof=1),
"spherical": np.var(X, axis=0, ddof=1).mean(),
}
bgmm = BayesianGaussianMixture(random_state=0)
for cov_type in ["full", "tied", "diag", "spherical"]:
bgmm.covariance_type = cov_type
bgmm.fit(X)
assert_almost_equal(covariance_prior_default[cov_type], bgmm.covariance_prior_)
def BGMreport(path, visualize=1, cut_n=6):
t2 = 15
t3 = 0.07
n_components = 3
denses, _ = finddensefromcut(path, cut_n)
maxd = []
for dense in denses[(cut_n - 5):]:
maxd.append(max(dense))
lofd = len(denses[0])
samples = list()
for i in range((cut_n - 5), cut_n): #sampling for BGM
samples.append(np.array(tosample(denses[i])).reshape(-1, 1))
allmeans = []
allcovs = []
allweights = []
BGM45 = np.zeros((45))
for i in range(5):
BGM = BayesianGaussianMixture(
n_components=n_components,
covariance_type='spherical',
weight_concentration_prior=0.000000000001,
max_iter=500)
BGM.fit(samples[i])
means = np.reshape(BGM.means_, (-1, ))
permu = np.argsort(means)
means = means[permu]
BGM45[i * 9 + 3:i * 9 + 6] = means
allmeans.append(means)
covs = BGM.covariances_
covs = covs[permu]
BGM45[i * 9 + 6:i * 9 + 9] = covs
allcovs.append(covs)
weights = BGM.weights_
weights = weights[permu]
BGM45[i * 9:i * 9 + 3] = weights * len(samples[i])
allweights.append(weights)
if visualize == 1:
l = 0
for i in range(cut_n - 5, cut_n): #visualization
l += 1
plt.subplot(2, n_components, l), plt.plot(denses[i])
X = np.linspace(0, lofd, num=200, endpoint=False)
Ys = toGM(X, n_components, allmeans[l - 1], allcovs[l - 1],
allweights[l - 1])
for j in range(n_components):
#plt.subplot(1,5,l),plt.plot([allmeans[l-1][j],allmeans[l-1][j]],[0,255])
plt.subplot(2, n_components,
l), plt.plot(X,
len(samples[l - 1]) * Ys[j])
#plt.subplot(2,n_components,l),plt.plot(X,Ys[j])
plt.ylim(0, 255)
plt.show()
ans = np.zeros((12, ))
pre = np.zeros((5, n_components))
for i in range(
5
): ###preprocessing the data to avoid peak overlapping(far overlap and near overlap) influence: identify far/near overlap cases and suppress far overlap peaks, amplify near overlap peaks
###如果很理想的情况应该能把两个far overlap的peak合并成一个在中间mean的，但是现在可以先直接把两个抑制掉，毕竟就不太可能是单克隆峰了。far overlap也就是两个峰实际上在图里面是同一个，BGM将其拆分从而更好的拟合高斯模型，我们这里将其抑制因为能够拆分为两个峰的基本上cov都比较大，不尖。
for j in range(n_components):
for l in range(n_components):
if j < l:
if allweights[i][j] / allweights[i][l] > 3 or allweights[
i][j] / allweights[i][
l] < 0.3333: #ignore when weight difference is too large
continue
if allcovs[i][j] / allweights[i][j] / allcovs[i][
l] * allweights[i][l] / abs(
allmeans[i][j] - allmeans[i][l]
) * mean(
np.sqrt(allcovs[i][j]), np.sqrt(allcovs[i][l])
) > 2 or allcovs[i][l] / allweights[i][
l] / allcovs[i][j] * allweights[i][j] / abs(
allmeans[i][j] - allmeans[i][l]
) * mean(
np.
sqrt(allcovs[i][j]), np.sqrt(allcovs[i][l])
) > 2: #if the cov difference is large than it will be ignored from far overlap because there should be two peaks in the original density plot
#near overlap situation is when a sharp peak is on a mild one. it happens when monoclonal peak has a background polyclonal peak. here we amplify the sharp peaks' weight when their cov difference is large enough or their distance is close enough so that it will be detected as abnormal in the classification step
if abs(allmeans[i][j] -
allmeans[i][l]) < 3.5 * np.sqrt(
max(allcovs[i][j], allcovs[i][l])):
neww = allweights[i][j] + allweights[i][l]
if allcovs[i][l] / allweights[i][l] / allcovs[i][
j] * allweights[i][j] > 1 and allweights[
i][j] > 0.15:
if allcovs[i][j] < 400:
allweights[i][j] = neww
else:
if allcovs[i][l] < 400:
allweights[i][l] = neww
continue
if allcovs[i][j] / allweights[i][j] / len(
samples[i]
) < t3 / 2.5 or allcovs[i][l] / allweights[i][l] / len(
samples[i]
) < t3 / 2.5: #if one of the considered peak has very small variance, then it should not be far overlap situation where the original peak is mild
continue
if allcovs[i][j] < 70 or allcovs[i][l] < 70:
continue
elif abs(allmeans[i][j] - allmeans[i][l]) < 3.5 * np.sqrt(
max(allcovs[i][j], allcovs[i][l])
): #far overlap situation where there is only a mild peak in the original density plot, and GMM model break it down to two sharper peaks to fit the guassian curves more accurately. here we just suppress the peaks and thus we cannot determine the column is abnormal because of the two considered components
pre[i][j] = pre[i][l] = 1
for i in [0, 1, 2]:
for j in [3, 4]:
if maxd[i] < 50 or maxd[j] < 50:
continue
else:
for k in range(len(allmeans[i])):
for l in range(len(allmeans[j])):
if pre[i][k] == 1 or pre[j][l] == 1:
continue
if abs(allmeans[i][k] - allmeans[j][l]) > lofd / t2:
continue
else:
if allweights[i][k] < 0.1 or allweights[j][l] < 0.1:
continue
else:
if allcovs[i][k] / allweights[i][k] / len(
samples[i]
) > t3 or allcovs[j][l] / allweights[j][l] / len(
samples[j]
) > t3: ###the t figure, represents the sharpness of the peak. just variance is not enough, we need to consider n_samples and weights too.
continue
else:
ans[i * 2 + j - 2] = 1
ans[7 + i] = 1
ans[7 + j] = 1
ans[0] = 1
for i in range(5):
for j in range(n_components):
if pre[i][j] == 1:
continue
if maxd[i] < 80:
continue
elif allweights[i][j] < 0.05:
continue
if allcovs[i][j] / allweights[i][j] / len(
samples[i]) > t3: ###t-figure
continue
else:
ans[7 + i] = 1
ans[0] = 1
return ans, BGM45
def test_compare_covar_type():
# We can compare the 'full' precision with the other cov_type if we apply
# 1 iter of the M-step (done during _initialize_parameters).
rng = np.random.RandomState(0)
rand_data = RandomData(rng, scale=7)
X = rand_data.X["full"]
n_components = rand_data.n_components
# Computation of the full_covariance
bgmm = BayesianGaussianMixture(
n_components=2 * n_components, covariance_type="full", max_iter=1, random_state=0, tol=1e-7
)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X)
full_covariances = bgmm.covariances_ * bgmm.degrees_of_freedom_[:, np.newaxis, np.newaxis]
# Check tied_covariance = mean(full_covariances, 0)
bgmm = BayesianGaussianMixture(
n_components=2 * n_components, covariance_type="tied", max_iter=1, random_state=0, tol=1e-7
)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X)
tied_covariance = bgmm.covariances_ * bgmm.degrees_of_freedom_
assert_almost_equal(tied_covariance, np.mean(full_covariances, 0))
# Check diag_covariance = diag(full_covariances)
bgmm = BayesianGaussianMixture(
n_components=2 * n_components, covariance_type="diag", max_iter=1, random_state=0, tol=1e-7
)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X)
diag_covariances = bgmm.covariances_ * bgmm.degrees_of_freedom_[:, np.newaxis]
assert_almost_equal(diag_covariances, np.array([np.diag(cov) for cov in full_covariances]))
# Check spherical_covariance = np.mean(diag_covariances, 0)
bgmm = BayesianGaussianMixture(
n_components=2 * n_components, covariance_type="spherical", max_iter=1, random_state=0, tol=1e-7
)
bgmm._check_initial_parameters(X)
bgmm._initialize_parameters(X)
spherical_covariances = bgmm.covariances_ * bgmm.degrees_of_freedom_
assert_almost_equal(spherical_covariances, np.mean(diag_covariances, 1))
value, vector = sp.linalg.eigh(cov)
width, height = value[0], value[1]
v = vector[0] / sp.linalg.norm(vector[0])
angle = 180* np.arctan(v[1] / v[0]) / np.pi
e = Ellipse(xy=center, width=width, height=height,
angle=angle, color=clrs[i], alpha=0.5, clip_box = ax.bbox)
ax.add_artist(e)
ax1_min, ax1_max, ax2_min, ax2_max = plt.axis()
plt.xlim((x1_min, x1_max))
plt.ylim((x2_min, x2_max))
plt.title('GMM', fontsize=15)
plt.grid(b=True, ls=':', color='#606060')
# DPGMM
dpgmm = BayesianGaussianMixture(n_components=n_components, covariance_type='full', max_iter=1000, n_init=5,
weight_concentration_prior_type='dirichlet_process', weight_concentration_prior=0.1)
dpgmm.fit(x) #假定高斯分布的参数是随机变量，且服从dirichlet_process过程，weight_concentration_prior越大越考虑到先验，越小越靠近样本
centers = dpgmm.means_
covs = dpgmm.covariances_
print u'DPGMM均值 = \n', centers
print u'DPGMM方差 = \n', covs
y_hat = dpgmm.predict(x)
print y_hat
ax = plt.subplot(212)
grid_hat = dpgmm.predict(grid_test)
grid_hat = grid_hat.reshape(x1.shape)
plt.pcolormesh(x1, x2, grid_hat, cmap=cm)
plt.scatter(x[:, 0], x[:, 1], s=20, c=y, cmap=cm, marker='o', edgecolors='#202020')
for i, cc in enumerate(zip(centers, covs)):
X_new = np.array([[-0.5, 0], [0, 0.5], [1, -0.1], [2, 1]])
knn.predict(X_new)
print(knn.predict_proba(X_new))
y_dist, y_pred_idx = knn.kneighbors(X_new, n_neighbors=1)
y_pred = dbscan.labels_[dbscan.core_sample_indices_][y_pred_idx]
y_pred[y_dist > 0.2] = -1
print(y_pred.ravel())
from sklearn.mixture import GaussianMixture
gm = GaussianMixture(n_components=3, n_init=10)
gm.fit(X)
X_new, y_new = gm.sample(6)
print("density:")
print(gm.score_samples(X))
densities = gm.score_samples(X)
density_threshold = np.percentile(densities, 4)
anomalies = X[densities < density_threshold]
print(gm.bic(X))
print(gm.aic(X))
from sklearn.mixture import BayesianGaussianMixture
bgm = BayesianGaussianMixture(n_components=10, n_init=10)
bgm.fit(X)
print("weights:")
print(np.round(bgm.weights_, 2))
from sklearn.mixture import BayesianGaussianMixture, GaussianMixture from sklearn.model_selection import train_test_split import sklearn import pandas import cv2 input_path = "/Users/Srikanth/PycharmProjects/DataMiningClass/datasets/face.jpg" img = cv2.imread(input_path) o_shape = img.shape k = 2 new_data = img.reshape(-1, 3) vgmm = BayesianGaussianMixture(n_components=k) # vgmm = GaussianMixture(n_components=k) vgmm = vgmm.fit(new_data) cluater = vgmm.predict(new_data) # Reshape the input data to the orignal shape cluater = cluater.reshape(o_shape[0], o_shape[1]) from matplotlib import pyplot pyplot.imshow(cluater) pyplot.show()
def wnn_entropy(points, k=None, weights=True, n_est=None, gmm=None):
r"""
Weighted Kozachenko-Leonenko nearest-neighbour entropy calculation.
*k* is the number of neighbours to consider, with default $k=n^{1/3}$
*n_est* is the number of points to use for estimating the entropy,
with default $n_\rm{est} = n$
*weights* is True for default weights, False for unweighted (using the
distance to the kth neighbour only), or a vector of weights of length *k*.
*gmm* is the number of gaussians to use to model the distribution using
a gaussian mixture model. Default is 0, and the points represent an
empirical distribution.
Returns entropy H in bits and its uncertainty.
Berrett, T. B., Samworth, R.J., Yuan, M., 2016. Efficient multivariate
entropy estimation via k-nearest neighbour distances.
https://arxiv.org/abs/1606.00304
"""
from sklearn.neighbors import NearestNeighbors
n, d = points.shape
# Default to the full set
if n_est is None:
n_est = n
# reduce size of draw to n_est
if n_est >= n:
x = points
else:
x = points[permutation(n)[:n_est]]
n = n_est
# Default k based on n
if k is None:
# Private communication: cube root of n is a good choice for k
# Personal observation: k should be much bigger than d
k = max(int(n**(1/3)), 3*d)
# If weights are given then use them (setting the appropriate k),
# otherwise use the default weights.
if isinstance(weights, bool):
weights = _wnn_weights(k, d, weights)
else:
k = len(weights)
#print("weights", weights, sum(weights))
# select knn algorithm
algorithm = 'auto'
#algorithm = 'kd_tree'
#algorithm = 'ball_tree'
#algorithm = 'brute'
n_components = 0 if gmm is None else gmm
# H = 1/n sum_i=1^n sum_j=1^k w_j log E_{j,i}
# E_{j,i} = e^-Psi(j) V_d (n-1) z_{j,i}^d = C z^d
# logC = -Psi(j) + log(V_d) + log(n-1)
# H = 1/n sum sum w_j logC + d/n sum sum w_j log(z)
# = sum w_j logC + d/n sum sum w_j log(z)
# = A + d/n B
# H^2 = 1/n sum
Psi = digamma(np.arange(1, k+1))
logVd = d/2*log(pi) - gammaln(1 + d/2)
logC = -Psi + logVd + log(n-1)
# TODO: standardizing points doesn't work.
# Standardize the data so that distances conform. This is equivalent to
# a u-substitution u = sigma x + mu, so the integral needs to be corrected
# for dU = det(sigma) dx. Since the standardization squishes the dimensions
# independently, sigma is a diagonal matrix, with the determinant equal to
# the product of the diagonal elements.
#x, mu, sigma = standardize(x) # Note: sigma may be zero
#detDU = np.prod(sigma)
detDU = 1.
if n_components > 0:
# Use Gaussian mixture to model the distribution
from sklearn.mixture import GaussianMixture as GMM
predictor = GMM(n_components=gmm, covariance_type='full')
predictor.fit(x)
eval_x, _ = predictor.sample(n_est)
#weight_x = predictor.score_samples(eval_x)
skip = 0
else:
# Empirical distribution
# TODO: should we use the full draw for kNN and a subset for eval points?
# Choose a subset for evaluating the entropy estimate, if desired
#print(n_est, n)
#eval_x = x if n_est >= n else x[permutation(n)[:n_est]]
eval_x = x
#weight_x = 1
skip = 1
tree = NearestNeighbors(algorithm=algorithm, n_neighbors=k+skip)
tree.fit(x)
dist, _ind = tree.kneighbors(eval_x, n_neighbors=k+skip, return_distance=True)
# Remove first column. Since test points are in x, the first column will
# be a point from x with distance 0, and can be ignored.
if skip:
dist = dist[:, skip:]
# Find log distances. This can be problematic for MCMC runs where a
# step is rejected, and therefore identical points are in the distribution.
# Ignore them by replacing these points with nan and using nanmean.
# TODO: need proper analysis of duplicated points in MCMC chain
dist[dist == 0] = nan
logdist = log(dist)
H_unweighted = logC + d*np.nanmean(logdist, axis=0)
H = np.dot(H_unweighted, weights)[0]
Hsq_k = np.nanmean((logC[-1] + d*logdist[:,-1])**2)
# TODO: abs shouldn't be needed?
if Hsq_k < H**2:
print("warning: avg(H^2) < avg(H)^2")
dH = sqrt(abs(Hsq_k - H**2)/n_est)
#print("unweighted", H_unweighted)
#print("weighted", H, Hsq_k, H**2, dH, detDU, LN2)
return H * detDU / LN2, dH * detDU / LN2
def cluster_vbgm(aligned_maps):
# sample_by_features = np.vstack([xmap.flatten() for xmap in aligned_maps])
embedding = embed(aligned_maps)
clusterer = BayesianGaussianMixture(n_components=10)
return clusterer.fit_predict(embedding)