def transform(self, data, outliers=None):
"""
Applies the learned Gaussian mixture model and
returns a labelled array.
Args:
data (NumPy Array): Material properties array of shape (height, width, n_properties)
outliers (NumPy Array): Binary array indicating outliers of shape (height, width)
Returns:
labels (NumPy Array): The segmented array of shape (height, width). Each pixel receives
a label corresponding to its segment.
"""
if self.gmm is None:
logger.warning("Attempting to transform prior to fitting. You must call .fit() first.")
return None
h, w, c = data.shape
data = self.get_pca_components(data)
labels = self.gmm.predict(data)
labels = np.reshape(labels, (h, w))
labels += 1 # all labels move up one
if outliers is not None:
labels *= 1 - outliers # outliers map to label 0
labels = slu.relabel(labels)
return labels
def transform(self,
data,
outliers=None,
pers_thresh=DEF_THRESH
): # NOTE need data as input to use as GMM Segmenter
"""
Applies threshold to the watershed graph and returns a labelled array.
Args:
data (NumPy Array): Material property array of shape (height, width).
outliers (NumPy Array): Binary array indicating outliers of shape (height, width)
pers_thresh (float): merging threshold
Returns:
(NumPy Array): The segmented array of shape (height, width). Each pixel receives
a label corresponding to its segment.
"""
if self.pws is None:
logger.warning(
"Attempting to transform prior to fitting. You must call .fit() first."
)
return None
if self.normalize:
data = SegmenterWatershed.normalize_data(data)
if self.smooth:
data = SegmenterWatershed.smooth_data(data)
labels = self.pws.apply_threshold(pers_thresh)
if outliers is not None:
labels *= 1 - outliers # outliers map to label 0 which are borders between grains
labels = slu.relabel(labels)
return labels
def get_grains(labels):
""" Segments classes labels into grain labels """
new_labels = measure.label(labels, connectivity=2, background=0)
new_labels = slu.relabel(new_labels)
return new_labels
def m2py_pipeline(dataframe, heightless, outlier_threshold, n_components,
padding, embedding_dim, thresh, nonlinear, normalize, zscale,
data_type, data_subtype, input_cmap):
"""
Wrapper function for m2py tools. Allows to include or exclude m2py tools in the order shown in the code.
Args:
dataframe - np.array(). 3D array of SPM data
heightless - bool. if 'True', height channel is removed, according to channel labels in config.py
outlier_threshold
n_components - int. number specifying the number of Gaussian phase distributions to be identified in dataframe
padding - int. number specifying the degree of closeness of neighbors to be included in GMM labelling
embedding_dims - int. number specifying the number of principle components to use in PCA before GMM labelling
thresh
nonlinear - bool. if 'True', nonlinear properties are generated prior to analysis. Includes x^2, x^3, abs(x),
and 1/x
normalize
zscale
data_type
data_subtype
input_cmap
Returns:
outliers - np.array(). 2D array of outliers
seg1_labels - np.array(). 2D array of GMM labels
seg2_labels - np.array(). 2D array of clustering labels
"""
fill_zeros_flag = True
remove_outliers_flag = True
input_cmap = input_cmap
# Apply frequency removal
data = pre.apply_frequency_removal(dataframe, data_type)
data_properties = config.data_info[data_type]['properties']
# Extract outliers
# outliers = pre.extract_outliers(data, data_type, prop = 'Zscale', threshold = outlier_threshold, chip_size = 256, stride = 256)
h, w, c = data.shape
outliers = np.zeros((h, w))
for prop in data_properties:
temp_outliers = pre.extract_outliers(data, data_type, prop, 2.5,
int(h / 2), int(w / 2))
pre.show_outliers(data, data_type, prop, temp_outliers)
outliers = np.logical_or(outliers, temp_outliers)
no_outliers_data = pre.smooth_outliers_from_data(data, outliers)
plt.imshow(outliers)
plt.show()
# Show a-priori distributions
pre.show_property_distributions(data, data_type, outliers)
c = data.shape[2]
num_pca_components = min(embedding_dim, c)
# Run GMM segmentation
seg1 = seg_gmm.SegmenterGMM(n_components=n_components,
embedding_dim=num_pca_components,
padding=padding,
nonlinear=nonlinear,
normalize=normalize,
zscale=zscale)
######################################## Heightless == True ##########################################
if heightless == True:
# Remove height property
zscale_id = data_properties.index("Zscale")
height_id = data_properties.index("Height")
no_height_data = np.delete(no_outliers_data, [zscale_id, height_id],
axis=2)
seg1_labels = seg1.fit_transform(no_height_data, outliers)
if fill_zeros_flag:
seg1_labels = slu.fill_out_zeros(seg1_labels, outliers)
elif remove_outliers_flag:
seg1_labels = np.ma.masked_where(outliers == 1, seg1_labels)
post.show_classification(seg1_labels, no_height_data, data_type)
# post.show_classification_correlation(seg1_labels, no_height_data, data_type)
# post.show_distributions_together(seg1_labels, no_height_data, data_type, input_cmap = 'jet')
# post.show_grain_area_distribution(seg1_labels, data_type, data_subtype)
# Overlay distributions on pixels
probs = seg1.get_probabilities(no_height_data)
post.show_overlaid_distribution(probs, no_height_data, data_type)
if embedding_dim != None:
h, w, c = no_height_data.shape
pca_components = seg1.get_pca_components(no_height_data)
pca_components = pca_components.reshape(h, w, num_pca_components)
post.show_classification_correlation(seg1_labels,
pca_components,
data_type,
title_flag=False)
# post.show_classification_distributions(seg1_labels, pca_components, data_type, title_flag=False)
else:
pass
## Persistence Watershed Segmentation clustering
if thresh != None:
comp_labels = list(np.unique(seg1_labels))
if 0 in comp_labels: # Avoid outlier components / class
comp_labels.remove(0)
watershed_id = data_properties.index(
"Adhesion") #Perhaps I should try Stiffness instead
seg2 = seg_water.SegmenterWatershed()
thresh = thresh
summed_labels = np.zeros_like(data[:, :, 0], dtype=np.int64)
for l in comp_labels:
watershed_data = no_outliers_data[:, :, watershed_id] * (
seg1_labels == l)
temp_labels = seg2.fit_transform(watershed_data,
outliers,
pers_thresh=thresh)
# NOTE: no need to fill out zeros in this case
# Instance (grains) segmentation of properties
print(f"Watershed Segmentation of GMM component {l}")
post.show_classification(temp_labels, no_outliers_data,
data_type)
# Add results from different components
temp_labels += np.max(
summed_labels
) # To distinguish labels from different components
temp_labels *= (seg1_labels == l)
summed_labels += temp_labels
seg2_labels = slu.relabel(summed_labels)
# Instance (grains) segmentation of properties
print("Watershed Segmentation of combined GMM components")
post.show_classification(seg2_labels, no_outliers_data, data_type)
## Conected-components clustering
else:
post_labels = seg1.get_grains(seg1_labels)
seg2_labels = slu.get_significant_labels(post_labels,
bg_contrast_flag=False)
post.show_classification(seg2_labels,
no_height_data,
data_type,
input_cmap=input_cmap)
# post.show_grain_area_distribution(seg2_labels, data_type, data_subtype)
######################################## Heightless == False ##########################################
elif heightless == False:
seg1_labels = seg1.fit_transform(no_outliers_data, outliers)
if fill_zeros_flag:
seg1_labels = slu.fill_out_zeros(seg1_labels, outliers)
elif remove_outliers_flag:
seg1_labels = np.ma.masked_where(outliers == 1, seg1_labels)
post.show_classification(seg1_labels, no_outliers_data, data_type)
post.show_distributions_together(seg1_labels,
no_outliers_data,
data_type,
input_cmap=input_cmap)
post.show_classification(seg1_labels, no_outliers_data, data_type)
post.show_classification_correlation(seg1_labels, no_outliers_data,
data_type)
post.show_distributions_together(seg1_labels,
no_outliers_data,
data_type,
input_cmap=input_cmap)
# Overlay distributions on pixels
probs = seg1.get_probabilities(no_outliers_data)
post.show_overlaid_distribution(probs, no_outliers_data, data_type)
post.show_grain_area_distribution(seg1_labels, data_type, data_subtype)
if embedding_dim != None:
h, w, c = no_outliers_data.shape
pca_components = seg1.get_pca_components(no_outliers_data)
pca_components = pca_components.reshape(h, w, num_pca_components)
post.show_classification_correlation(seg1_labels,
pca_components,
data_type,
title_flag=False)
post.show_classification_distributions(seg1_labels,
pca_components,
outliers,
data_type,
title_flag=False)
else:
pass
## Persistence Watershed Segmentation clustering
if thresh != None:
comp_labels = list(np.unique(seg1_labels))
if 0 in comp_labels: # Avoid outlier components / class
comp_labels.remove(0)
watershed_id = data_properties.index("Adhesion")
seg2 = seg_water.SegmenterWatershed()
thresh = thresh
summed_labels = np.zeros_like(data[:, :, 0], dtype=np.int64)
for l in comp_labels:
watershed_data = no_outliers_data[:, :, watershed_id] * (
seg1_labels == l) # Why the '*'? Can this be heightless??
temp_labels = seg2.fit_transform(watershed_data,
outliers,
pers_thresh=thresh)
# NOTE: no need to fill out zeros in this case
# Instance (grains) segmentation of properties
print(f"Watershed Segmentation of GMM component {l}")
post.show_classification(temp_labels, no_outliers_data,
data_type)
# Add results from different components
temp_labels += np.max(
summed_labels
) # To distinguish labels from different components
temp_labels *= (seg1_labels == l)
summed_labels += temp_labels
seg2_labels = slu.relabel(summed_labels)
# Instance (grains) segmentation of properties
print("Watershed Segmentation of combined GMM components")
post.show_classification(seg2_labels, no_outliers_data, data_type)
## Conected-components clustering
else:
post_labels = seg1.get_grains(seg1_labels)
seg2_labels = slu.get_significant_labels(post_labels,
bg_contrast_flag=True)
post.show_classification(seg2_labels,
no_outliers_data,
data_type,
input_cmap=input_cmap)
post.show_grain_area_distribution(seg2_labels, data_type, data_subtype)
else:
print('Error: Heightless flag incorrect')
return outliers, seg1_labels, seg2_labels
seg1_dict = {}
seg2_dict = {}
for fl in files:
if fl[-5] == '1': #Seg1 files
seg1_fl_list.append(fl)
elif fl[-5] == '2': #Seg2 files
seg2_fl_list.append(fl)
elif fl[-1] == 'v': #Morphology descriptor .csv files
pass
else:
print(fl[:-4], ' is messed up')
for k, fl in enumerate(seg1_fl_list):
seg1_dict[k] = np.load(map_file_path+fl)
for k, fl in enumerate(seg2_fl_list):
seg2_dict[k] = np.load(map_file_path+fl)
for i in range(len(seg1_dict)):
domain_labels = slu.relabel(seg2_dict[i])
phase_labels = slu.relabel(seg1_dict[i])
domain_props = slu.all_domain_properties(phase_labels, domain_labels)
outfile = map_file_path+seg1_fl_list[i][:-8]+'domain_metrics.csv'
domain_props.to_csv(outfile)
print ('finished file # ', i)