def OnOpenImage(self, evt=None):
# 1) Get the image key
# Start with the table_id if there is one
tblNum = None
if p.table_id:
dlg = wx.TextEntryDialog(self, p.table_id + ':',
'Enter ' + p.table_id)
dlg.SetValue('0')
if dlg.ShowModal() == wx.ID_OK:
try:
tblNum = int(dlg.GetValue())
except ValueError:
errdlg = wx.MessageDialog(
self, 'Invalid value for %s!' % (p.table_id),
"Invalid value", wx.OK | wx.ICON_EXCLAMATION)
errdlg.ShowModal()
return
dlg.Destroy()
else:
dlg.Destroy()
return
# Then get the image_id
dlg = wx.TextEntryDialog(self, p.image_id + ':', 'Enter ' + p.image_id)
dlg.SetValue('')
if dlg.ShowModal() == wx.ID_OK:
try:
imgNum = int(dlg.GetValue())
except ValueError:
errdlg = wx.MessageDialog(
self, 'Invalid value for %s!' % (p.image_id),
"Invalid value", wx.OK | wx.ICON_EXCLAMATION)
errdlg.ShowModal()
return
dlg.Destroy()
else:
dlg.Destroy()
return
# Build the imkey
if p.table_id:
imkey = (tblNum, imgNum)
else:
imkey = (imgNum, )
dm = DataModel.getInstance()
if imkey not in dm.GetAllImageKeys():
errdlg = wx.MessageDialog(self, 'There is no image with that key.',
"Couldn't find image",
wx.OK | wx.ICON_EXCLAMATION)
errdlg.ShowModal()
self.Destroy()
else:
# load the image
self.img_key = imkey
self.SetImage(imagetools.FetchImage(imkey), p.image_channel_colors)
self.DoLayout()
def OnOpenImage(self, evt=None):
# 1) Get the image key
# Start with the table_id if there is one
tblNum = None
if p.table_id:
dlg = wx.TextEntryDialog(self, p.table_id + ":", "Enter " + p.table_id)
dlg.SetValue("0")
if dlg.ShowModal() == wx.ID_OK:
try:
tblNum = int(dlg.GetValue())
except ValueError:
errdlg = wx.MessageDialog(
self, "Invalid value for %s!" % (p.table_id), "Invalid value", wx.OK | wx.ICON_EXCLAMATION
)
errdlg.ShowModal()
return
dlg.Destroy()
else:
dlg.Destroy()
return
# Then get the image_id
dlg = wx.TextEntryDialog(self, p.image_id + ":", "Enter " + p.image_id)
dlg.SetValue("")
if dlg.ShowModal() == wx.ID_OK:
try:
imgNum = int(dlg.GetValue())
except ValueError:
errdlg = wx.MessageDialog(
self, "Invalid value for %s!" % (p.image_id), "Invalid value", wx.OK | wx.ICON_EXCLAMATION
)
errdlg.ShowModal()
return
dlg.Destroy()
else:
dlg.Destroy()
return
# Build the imkey
if p.table_id:
imkey = (tblNum, imgNum)
else:
imkey = (imgNum,)
dm = DataModel.getInstance()
if imkey not in dm.GetAllImageKeys():
errdlg = wx.MessageDialog(
self, "There is no image with that key.", "Couldn't find image", wx.OK | wx.ICON_EXCLAMATION
)
errdlg.ShowModal()
self.Destroy()
else:
# load the image
self.img_key = imkey
self.SetImage(imagetools.FetchImage(imkey), p.image_channel_colors)
self.DoLayout()
def get_image_keys_at_row(self, row):
'''Returns a list of image keys at the given row or None if the column
names can't be found in col_labels
'''
if self.key_indices is None or self.grouping is None:
return None
else:
if self.grouping.lower() == 'image':
return [tuple(self.data[self.row_order,:][row, self.key_indices])]
elif self.grouping.lower() == 'object':
return [tuple(self.data[self.row_order,:][row, self.key_indices[:-1]])]
else:
dm = DataModel.getInstance()
return dm.GetImagesInGroup(self.grouping, self.get_row_key(row))
def PerImageCounts(self, filter_name=None, cb=None):
# Clear the current perClassObjects storage
for bin in self.classBins:
self.perClassObjects[bin.label] = []
# Retrieve a data model instance
dm = DataModel.getInstance()
# Retrieve image keys and initialize variables
imageKeys = dm.GetAllImageKeys(filter_name)
imageAmount = float(len(imageKeys))
perImageData = []
# Process all images
for k_index, imKey in enumerate(imageKeys):
try:
# Retrieve the keys of the objects in the current image
obKeys = dm.GetObjectsFromImage(imKey)
except:
raise 'No such image: %s' % (imKey, )
return
# Calculate the amount of hits for each of the classes in the current image
classHits = {}
objectCount = [imKey[0]]
if obKeys:
classObjects = self.FilterObjectsFromClassN(keys=[imKey])
for clNum, bin in enumerate(self.classBins):
# Get the objects from the image which belong to the selected class
classHits[bin.label] = classObjects[float(clNum + 1)]
# Store the total object count of this class for the current image
nrHits = len(classHits[bin.label])
objectCount.append(nrHits)
# Store the objects for the current class and image grouped
# by class if any are found for this class in the selected image
if nrHits > 0:
self.perClassObjects[bin.label] += classHits[bin.label]
else:
# If there are objects in the image, add zeros for all bins
[objectCount.append(0) for bin in self.classBins]
# Store the results for the current image and update the callback
# function if available
perImageData.append(objectCount)
if cb:
cb(min(1, k_index / imageAmount))
return perImageData
def PerImageCounts(self, filter_name=None, cb=None):
# Clear the current perClassObjects storage
for bin in self.classBins:
self.perClassObjects[bin.label] = []
# Retrieve a data model instance
dm = DataModel.getInstance()
# Retrieve image keys and initialize variables
imageKeys = dm.GetAllImageKeys(filter_name)
imageAmount = float(len(imageKeys))
perImageData = []
# Process all images
for k_index, imKey in enumerate(imageKeys):
try:
# Retrieve the keys of the objects in the current image
obKeys = dm.GetObjectsFromImage(imKey)
except:
raise ValueError('No such image: %s' % (imKey,))
# Calculate the amount of hits for each of the classes in the current image
classHits = {}
objectCount = [imKey[0]]
if obKeys:
classObjects = self.FilterObjectsFromClassN(keys = [imKey])
for clNum, bin in enumerate(self.classBins):
# Get the objects from the image which belong to the selected class
classHits[bin.label] = classObjects[float(clNum+1)]
# Store the total object count of this class for the current image
nrHits = len(classHits[bin.label])
objectCount.append(nrHits)
# Store the objects for the current class and image grouped
# by class if any are found for this class in the selected image
if nrHits > 0:
self.perClassObjects[bin.label] += classHits[bin.label]
else:
# If there are objects in the image, add zeros for all bins
[objectCount.append(0) for bin in self.classBins]
# Store the results for the current image and update the callback
# function if available
perImageData.append(objectCount)
if cb:
cb(min(1, k_index/imageAmount))
return perImageData
def get_object_keys_at_row(self, row):
'''Returns a list of object keys at the given row or None if the column
names can't be found in col_labels
'''
if self.key_indices is None or self.grouping is None:
return None
else:
dm = DataModel.getInstance()
# If the key index for the row is an object key, just return that key
if self.grouping.lower() == 'object':
return [tuple(self.data[self.row_order,:][row, self.key_indices])]
else: # Otherwise, return all object keys in the image
imkeys = self.get_image_keys_at_row(row)
obkeys = []
for imkey in imkeys:
obs = dm.GetObjectCountFromImage(imkey)
obkeys += [tuple(list(imkey)+[i]) for i in range(1,obs+1)]
return obkeys
def on_dclick_label(self, evt):
'''Handle display of images and objects'''
if evt.Row >= 0:
obkeys = self.grid.Table.get_object_keys_at_row(evt.Row)
if self.grid.Table.grouping is None:
# We need to know how the table is grouped to know what to do
logging.warn(
'CPA does not know how to link this table to your images. Can\'t launch ImageViewer.'
)
return
elif self.grid.Table.grouping.lower() == 'object':
# For per-object grouping, show the objects in the image
imview = imagetools.ShowImage(obkeys[0][:-1],
p.image_channel_colors,
parent=self.Parent)
if obkeys is not None:
for obkey in obkeys:
imview.SelectObject(obkey)
elif self.grid.Table.grouping.lower() == 'image':
# For per-image grouping just show the images.
# If there is only one object, then highlight it
if obkeys is not None and len(obkeys) == 1:
imview = imagetools.ShowImage(obkeys[0][:-1],
p.image_channel_colors,
parent=self.Parent)
imview.SelectObject(obkeys[0])
else:
imkeys = self.grid.Table.get_image_keys_at_row(evt.Row)
if imkeys:
#XXX: warn if there are a lot
for imkey in imkeys:
imagetools.ShowImage(imkey,
p.image_channel_colors,
parent=self.Parent)
else:
key_cols = self.grid.Table.get_row_key(evt.Row)
if key_cols:
dm = DataModel.getInstance()
imkeys = dm.GetImagesInGroup(self.grid.Table.grouping,
key_cols)
for imkey in imkeys:
imagetools.ShowImage(imkey,
p.image_channel_colors,
parent=self.Parent)
def get_image_keys_at_row(self, row):
'''Returns a list of image keys at the given row or None if the column
names can't be found in col_labels
'''
if self.key_indices is None or self.grouping is None:
return None
else:
if self.grouping.lower() == 'image':
return [
tuple(self.data[self.row_order, :][row, self.key_indices])
]
elif self.grouping.lower() == 'object':
return [
tuple(self.data[self.row_order, :][row,
self.key_indices[:-1]])
]
else:
dm = DataModel.getInstance()
return dm.GetImagesInGroup(self.grouping,
self.get_row_key(row))
def get_object_keys_at_row(self, row):
'''Returns a list of object keys at the given row or None if the column
names can't be found in col_labels
'''
if self.key_indices is None or self.grouping is None:
return None
else:
dm = DataModel.getInstance()
# If the key index for the row is an object key, just return that key
if self.grouping.lower() == 'object':
return [
tuple(self.data[self.row_order, :][row, self.key_indices])
]
else: # Otherwise, return all object keys in the image
imkeys = self.get_image_keys_at_row(row)
obkeys = []
for imkey in imkeys:
obs = dm.GetObjectCountFromImage(imkey)
obkeys += [
tuple(list(imkey) + [i]) for i in range(1, obs + 1)
]
return obkeys
def on_dclick_label(self, evt):
'''Handle display of images and objects'''
if evt.Row >= 0:
obkeys = self.grid.Table.get_object_keys_at_row(evt.Row)
if self.grid.Table.grouping is None:
# We need to know how the table is grouped to know what to do
logging.warn('CPA does not know how to link this table to your images. Can\'t launch ImageViewer.')
return
elif self.grid.Table.grouping.lower() == 'object':
# For per-object grouping, show the objects in the image
imview = imagetools.ShowImage(obkeys[0][:-1],
p.image_channel_colors,
parent=self.Parent)
if obkeys is not None:
for obkey in obkeys:
imview.SelectObject(obkey)
elif self.grid.Table.grouping.lower() == 'image':
# For per-image grouping just show the images.
# If there is only one object, then highlight it
if obkeys is not None and len(obkeys) == 1:
imview = imagetools.ShowImage(obkeys[0][:-1],
p.image_channel_colors,
parent=self.Parent)
imview.SelectObject(obkeys[0])
else:
imkeys = self.grid.Table.get_image_keys_at_row(evt.Row)
if imkeys:
#XXX: warn if there are a lot
for imkey in imkeys:
imagetools.ShowImage(imkey, p.image_channel_colors,
parent=self.Parent)
else:
key_cols = self.grid.Table.get_row_key(evt.Row)
if key_cols:
dm = DataModel.getInstance()
imkeys = dm.GetImagesInGroup(self.grid.Table.grouping, key_cols)
for imkey in imkeys:
imagetools.ShowImage(imkey, p.image_channel_colors,
parent=self.Parent)
def FilterObjectsFromClassN(self, classN=None, keys=None):
'''
Filter the input objects to output the keys of those in classN,
using a defined SVM model classifier.
'''
# Retrieve instance of the database connection
db = dbconnect.DBConnect.getInstance()
object_data = {}
if isinstance(keys, str):
object_data[0] = db.GetCellDataForClassifier(keys)
elif keys != []:
if len(keys) == len(dbconnect.image_key_columns()):
# Retrieve instance of the data model and retrieve objects in the requested image
dm = DataModel.getInstance()
obKeys = dm.GetObjectsFromImage(keys[0])
else:
obKeys = keys
for key in obKeys:
object_data[key] = db.GetCellDataForClassifier(key)
sorted_keys = sorted(object_data.keys())
values_array = np.array([object_data[key] for key in sorted_keys])
scaled_values = self.ScaleData(values_array)
pred_labels = self.model.predict(scaled_values)
# Group the object keys per class
classObjects = {}
for index in range(1, len(self.classBins) + 1):
classObjects[float(index)] = []
for index, label in enumerate(pred_labels):
classObjects[np.int(label) + 1].append(sorted_keys[index])
# Return either a summary of all classes and their corresponding objects
# or just the objects for a specific class
if classN is None:
return classObjects
else:
return classObjects[classN]
def FilterObjectsFromClassN(self, classN = None, keys = None):
'''
Filter the input objects to output the keys of those in classN,
using a defined SVM model classifier.
'''
# Retrieve instance of the database connection
db = dbconnect.DBConnect.getInstance()
object_data = {}
if isinstance(keys, str):
object_data[0] = db.GetCellDataForClassifier(keys)
elif keys != []:
if len(keys) == len(dbconnect.image_key_columns()):
# Retrieve instance of the data model and retrieve objects in the requested image
dm = DataModel.getInstance()
obKeys = dm.GetObjectsFromImage(keys[0])
else:
obKeys = keys
for key in obKeys:
object_data[key] = db.GetCellDataForClassifier(key)
sorted_keys = sorted(object_data.keys())
values_array = np.array([object_data[key] for key in sorted_keys])
scaled_values = self.ScaleData(values_array)
pred_labels = self.model.predict(scaled_values)
# Group the object keys per class
classObjects = {}
for index in range(1, len(self.classBins)+1):
classObjects[float(index)] = []
for index, label in enumerate(pred_labels):
classObjects[np.int(label)+1].append(sorted_keys[index])
# Return either a summary of all classes and their corresponding objects
# or just the objects for a specific class
if classN is None:
return classObjects
else:
return classObjects[classN]
def get_object_keys_at_row(self, row):
# XXX: needs to be updated to work for per_well data
if self.table_name == p.image_table:
# return all objects in this image
key = self.get_row_key(row)
if key is None:
return None
dm = DataModel.getInstance()
n_objects = dm.GetObjectCountFromImage(key)
return [tuple(list(key) + [i]) for i in range(n_objects)]
elif self.table_name == p.object_table:
key = self.get_row_key(row)
if key is None:
return None
return [key]
else:
key = []
for col in dbconnect.object_key_columns():
if col not in self.col_labels:
return None
else:
col_index = self.col_labels.tolist().index(col)
key += [self.GetValue(row, col_index)]
return [tuple(key)]
def get_object_keys_at_row(self, row):
# XXX: needs to be updated to work for per_well data
if self.table_name == p.image_table:
# return all objects in this image
key = self.get_row_key(row)
if key is None:
return None
dm = DataModel.getInstance()
n_objects = dm.GetObjectCountFromImage(key)
return [tuple(list(key) + [i]) for i in range(n_objects)]
elif self.table_name == p.object_table:
key = self.get_row_key(row)
if key is None:
return None
return [key]
else:
key = []
for col in dbconnect.object_key_columns():
if col not in self.col_labels:
return None
else:
col_index = self.col_labels.tolist().index(col)
key += [self.GetValue(row, col_index)]
return [tuple(key)]
def score_objects(properties, ts, gt, nRules, filter_name=None, group='Image',
show_results=False, results_table=None, overwrite=False):
'''
Trains a Classifier on a training set and scores the experiment
returns the table of scores as a numpy array.
properties -- Properties instance
ts -- TrainingSet instance
gt -- Ground Truth instance
nRules -- number of rules to use
filter_name -- name of a filter to use from the properties file
group -- name of a group to use from the properties file
show_results -- whether or not to show the results in TableViewer
results_table -- table name to save results to or None.
'''
p = properties
#db = DBConnect.getInstance() ## Removed writing to db. Results_table should be 'None' anyway
dm = DataModel.getInstance()
#if group == None:
#group = 'Image'
if results_table:
if db.table_exists(results_table) and not overwrite:
print 'Table "%s" already exists. Delete this table before running scoreall.'%(results_table)
return None
print ''
print 'properties: ', properties
print 'initial training set: ', ts
print 'ground truth training set: ', gt
print '# rules: ', nRules
print 'filter: ', filter_name
print 'grouping by: ', group
print 'show results: ', show_results
print 'results table: ', results_table
print 'overwrite: ', overwrite
print ''
nClasses = len(ts.labels)
nKeyCols = len(image_key_columns())
assert 200 > nRules > 0, '# of rules must be between 1 and 200. Value was %s'%(nRules,)
assert filter_name in p._filters.keys()+[None], 'Filter %s not found in properties file. Valid filters are: %s'%(filter_name, ','.join(p._filters.keys()),)
assert group in p._groups.keys()+['Image', 'None'], 'Group %s not found in properties file. Valid groups are: %s'%(group, ','.join(p._groups.keys()),)
output = StringIO()
logging.info('Training classifier with %s rules...'%nRules)
t0 = time()
weaklearners = fastgentleboostingmulticlass.train(ts.colnames,
nRules, ts.label_matrix,
ts.values, output)
logging.info('Training done in %f seconds'%(time()-t0))
t0 = time()
#def update(frac):
#logging.info('%d%% '%(frac*100.,))
## Score Ground Truth using established classifier
gt_predicted_scores = per_cell_scores(weaklearners, gt.values, gt.colnames)
#plt.hist(gt_predicted_scores)
#plt.show()
gt_predicted_signs = np.sign(gt_predicted_scores)
## Compare Ground Truth score signs with the actual ground truth values
numclasses = ts.labels.size
gt_actual_signs = gt.label_matrix[:,0]
cm_unrotated = metrics.confusion_matrix(gt_actual_signs,gt_predicted_signs)
## sklearn.metrics.confusion_matrix -- 2D confusion matrix is inverted from convention.
## https://github.com/scikit-learn/scikit-learn/issues/1664
cm = np.rot90(np.rot90(cm_unrotated))
fpr, sens, thresholds = metrics.roc_curve(gt_actual_signs,gt_predicted_signs)
spec = 1-fpr
s = np.sum(cm,axis=1)
percent = [100*cm[i,i]/float(s[i]) for i in range(len(s))]
avg = np.mean(percent)
avgTotal = 100 * np.trace(cm) / float(np.sum(cm))
print 'accuracy = %f' % avgTotal
print 'Confusion Matrix = ... '
print cm
my_sens = cm[0,0] / float(cm[0,0] + cm[0,1]) #TP/(TP+FN)
my_spec = cm[1,1] / float(cm[1,1] + cm[1,0]) #TN/(TN+FP)
print 'My_Sensitivity = %f' % my_sens
print 'My_Specificity = %f' % my_spec
print 'Sensitivity = ...'
print sens
print 'Specificity = ...'
print spec
print 'Done calculating'
############
## Confusion Matrix code from here: http://stackoverflow.com/questions/5821125/how-to-plot-confusion-matrix-with-string-axis-rather-than-integer-in-python
conf_arr = cm
norm_conf = []
## This normalizes each *row* to the color map, but I chose to normalize the whole confusion matrix to the same scale
##for i in conf_arr:
##a = 0
##tmp_arr = []
##a = sum(i, 0)
##for j in i:
##tmp_arr.append(float(j)/float(a))
##norm_conf.append(tmp_arr)
norm_conf = conf_arr / float(np.max(conf_arr))
if DISPLAY_CONFUSION_MATRIX:
fig = plt.figure()
plt.clf()
ax = fig.add_subplot(111)
ax.set_aspect(1)
res = ax.imshow(np.array(norm_conf), cmap=plt.cm.jet,
interpolation='nearest')
width = len(conf_arr)
height = len(conf_arr[0])
for x in xrange(width):
for y in xrange(height):
ax.annotate(str(conf_arr[x][y]), xy=(y, x),
horizontalalignment='center',
verticalalignment='center')
cb = fig.colorbar(res)
#cb.set_cmap = [0,1]
if width == 2 and height == 2:
plt.xticks([0,1],['FP','TN'])
plt.yticks([0,1],['TP','FP'])
else:
alphabet = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
plt.xticks(range(width), alphabet[:width])
plt.yticks(range(height), alphabet[:height])
plt.show()
print 'Done'
from sys import stderr
from tempfile import gettempdir
from time import ctime, time
#from wx.lib.embeddedimage import PyEmbeddedImage
import dbconnect
import imagetools
import csv
import logging
import numpy as np
import os
import sys
import weakref
import wx
import wx.grid
dm = DataModel.getInstance()
db = dbconnect.DBConnect.getInstance()
p = Properties.getInstance()
ID_LOAD_CSV = wx.NewId()
ID_SAVE_CSV = wx.NewId()
ID_EXIT = wx.NewId()
DO_NOT_LINK_TO_IMAGES = 'Do not link to images'
ROW_LABEL_SIZE = 30
# Icon to be used for row headers (difficult to implement)
#img_icon = PyEmbeddedImage('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')
class HugeTable(wx.grid.PyGridTableBase):
'''
def score(properties,
ts,
nRules,
filter_name=None,
group='Image',
show_results=False,
results_table=None,
overwrite=False):
'''
Trains a Classifier on a training set and scores the experiment
returns the table of scores as a numpy array.
properties -- Properties instance
ts -- TrainingSet instance
nRules -- number of rules to use
filter_name -- name of a filter to use from the properties file
group -- name of a group to use from the properties file
show_results -- whether or not to show the results in TableViewer
results_table -- table name to save results to or None.
'''
p = properties
db = DBConnect.getInstance()
dm = DataModel.getInstance()
if group == None:
group = 'Image'
if results_table:
if db.table_exists(results_table) and not overwrite:
print 'Table "%s" already exists. Delete this table before running scoreall.' % (
results_table)
return None
print ''
print 'properties: ', properties
print 'training set: ', ts
print '# rules: ', nRules
print 'filter: ', filter_name
print 'grouping by: ', group
print 'show results: ', show_results
print 'results table: ', results_table
print 'overwrite: ', overwrite
print ''
nClasses = len(ts.labels)
nKeyCols = len(image_key_columns())
assert 200 > nRules > 0, '# of rules must be between 1 and 200. Value was %s' % (
nRules, )
assert filter_name in p._filters.keys() + [
None
], 'Filter %s not found in properties file. Valid filters are: %s' % (
filter_name,
','.join(p._filters.keys()),
)
assert group in p._groups.keys() + [
'Image'
], 'Group %s not found in properties file. Valid groups are: %s' % (
group,
','.join(p._groups.keys()),
)
output = StringIO()
logging.info('Training classifier with %s rules...' % nRules)
t0 = time()
weaklearners = fastgentleboostingmulticlass.train(ts.colnames, nRules,
ts.label_matrix,
ts.values, output)
logging.info('Training done in %f seconds' % (time() - t0))
logging.info('Computing per-image class counts...')
t0 = time()
def update(frac):
logging.info('%d%% ' % (frac * 100., ))
keysAndCounts = multiclasssql.PerImageCounts(weaklearners,
filter_name=(filter_name
or None),
cb=update)
keysAndCounts.sort()
logging.info('Counts found in %f seconds' % (time() - t0))
if not keysAndCounts:
logging.error(
'No images are in filter "%s". Please check the filter definition in your properties file.'
% (filter_name))
raise Exception(
'No images are in filter "%s". Please check the filter definition in your properties file.'
% (filter_name))
# AGGREGATE PER_IMAGE COUNTS TO GROUPS IF NOT GROUPING BY IMAGE
if group != 'Image':
logging.info('Grouping %s counts by %s...' % (p.object_name[0], group))
t0 = time()
imData = {}
for row in keysAndCounts:
key = tuple(row[:nKeyCols])
imData[key] = np.array([float(v) for v in row[nKeyCols:]])
groupedKeysAndCounts = np.array([
list(k) + vals.tolist()
for k, vals in dm.SumToGroup(imData, group).items()
],
dtype=object)
nKeyCols = len(dm.GetGroupColumnNames(group))
logging.info('Grouping done in %f seconds' % (time() - t0))
else:
groupedKeysAndCounts = np.array(keysAndCounts, dtype=object)
# FIT THE BETA BINOMIAL
logging.info('Fitting beta binomial distribution to data...')
counts = groupedKeysAndCounts[:, -nClasses:]
alpha, converged = polyafit.fit_betabinom_minka_alternating(counts)
logging.info(' alpha = %s converged = %s' % (alpha, converged))
logging.info(' alpha/Sum(alpha) = %s' % ([a / sum(alpha)
for a in alpha]))
# CONSTRUCT ARRAY OF TABLE DATA
logging.info('Computing enrichment scores for each group...')
t0 = time()
tableData = []
for i, row in enumerate(groupedKeysAndCounts):
# Start this row with the group key:
tableRow = list(row[:nKeyCols])
if group != 'Image':
tableRow += [
len(dm.GetImagesInGroup(group, tuple(row[:nKeyCols])))
]
# Append the counts:
countsRow = [int(v) for v in row[nKeyCols:nKeyCols + nClasses]]
tableRow += [sum(countsRow)]
tableRow += countsRow
if p.area_scoring_column is not None:
# Append the areas
countsRow = [int(v) for v in row[-nClasses:]]
tableRow += [sum(countsRow)]
tableRow += countsRow
# Append the scores:
# compute enrichment probabilities of each class for this image OR group
scores = np.array(dirichletintegrate.score(alpha, np.array(countsRow)))
# clamp to [0,1] to
scores[scores > 1.] = 1.
scores[scores < 0.] = 0.
tableRow += scores.tolist()
# Append the logit scores:
# Special case: only calculate logit of "positives" for 2-classes
if nClasses == 2:
tableRow += [np.log10(scores[0]) - (np.log10(1 - scores[0]))
] # compute logit of each probability
else:
tableRow += [
np.log10(score) - (np.log10(1 - score)) for score in scores
] # compute logit of each probability
tableData.append(tableRow)
tableData = np.array(tableData, dtype=object)
logging.info('Enrichments computed in %f seconds' % (time() - t0))
# CREATE COLUMN LABELS LIST
# if grouping isn't per-image, then get the group key column names.
if group != 'Image':
colnames = dm.GetGroupColumnNames(group)
else:
colnames = list(image_key_columns())
# record the column indices for the keys
key_col_indices = [i for i in range(len(colnames))]
if group != 'Image':
colnames += ['Number_of_Images']
colnames += ['Total_%s_Count' % (p.object_name[0].capitalize())]
for i in xrange(nClasses):
colnames += [
'%s_%s_Count' %
(ts.labels[i].capitalize(), p.object_name[0].capitalize())
]
if p.area_scoring_column is not None:
colnames += ['Total_%s_Area' % (p.object_name[0].capitalize())]
for i in xrange(nClasses):
colnames += [
'%s_%s_Area' %
(ts.labels[i].capitalize(), p.object_name[0].capitalize())
]
for i in xrange(nClasses):
colnames += ['pEnriched_%s' % (ts.labels[i])]
if nClasses == 2:
colnames += ['Enriched_Score_%s' % (ts.labels[0])]
else:
for i in xrange(nClasses):
colnames += ['Enriched_Score_%s' % (ts.labels[i])]
title = results_table or "Enrichments_per_%s" % (group, )
if filter_name:
title += "_filtered_by_%s" % (filter_name, )
title += ' (%s)' % (os.path.split(p._filename)[1])
if results_table:
print 'Creating table %s' % (results_table)
success = db.CreateTableFromData(tableData,
colnames,
results_table,
temporary=False)
if not success:
print 'Failed to create results table :('
if show_results:
import tableviewer
tableview = tableviewer.TableViewer(None, title=title)
if results_table and overwrite:
tableview.load_db_table(results_table)
else:
tableview.table_from_array(tableData, colnames, group,
key_col_indices)
tableview.set_fitted_col_widths()
tableview.Show()
return tableData
def __init__(self, properties=None, parent=None, id=ID_IMAGE_GALLERY, **kwargs):
if properties is not None:
global p
p = properties
global db
db = dbconnect.DBConnect.getInstance()
wx.Frame.__init__(self, parent, id=id, title='CPA/ImageGallery - %s' % \
(os.path.basename(p._filename)), size=(800, 600), **kwargs)
if parent is None and not sys.platform.startswith('win'):
self.tbicon = wx.TaskBarIcon()
self.tbicon.SetIcon(icons.get_cpa_icon(), 'CPA/ImageGallery')
else:
self.SetIcon(icons.get_cpa_icon())
self.SetName('ImageGallery')
db.register_gui_parent(self)
global dm
dm = DataModel.getInstance()
if not p.is_initialized():
logging.critical('ImageGallery requires a properties file. Exiting.')
raise Exception('ImageGallery requires a properties file. Exiting.')
self.pmb = None
self.worker = None
self.trainingSet = None
self.classBins = []
self.binsCreated = 0
self.chMap = p.image_channel_colors[:]
self.toggleChMap = p.image_channel_colors[
:] # used to store previous color mappings when toggling colors on/off with ctrl+1,2,3...
self.brightness = 1.0
self.scale = 1.0
self.contrast = 'Linear'
self.defaultTSFileName = None
self.defaultModelFileName = None
self.lastScoringFilter = None
self.menuBar = wx.MenuBar()
self.SetMenuBar(self.menuBar)
self.CreateMenus()
self.CreateStatusBar()
#### Create GUI elements
# Top level - three split windows
self.splitter = wx.SplitterWindow(self, style=wx.NO_BORDER | wx.SP_3DSASH)
self.fetch_and_rules_panel = wx.Panel(self.splitter)
self.bins_splitter = wx.SplitterWindow(self.splitter, style=wx.NO_BORDER | wx.SP_3DSASH)
# fetch & rules
self.fetch_panel = wx.Panel(self.fetch_and_rules_panel)
self.find_rules_panel = wx.Panel(self.fetch_and_rules_panel)
# sorting bins
self.gallery_panel = wx.Panel(self.bins_splitter)
o_label = p.object_name[0] if p.classification_type == 'image' else '' + ' image gallery'
self.gallery_box = wx.StaticBox(self.gallery_panel, label=o_label)
self.gallery_sizer = wx.StaticBoxSizer(self.gallery_box, wx.VERTICAL)
self.galleryBin = sortbin.SortBin(parent=self.gallery_panel,
classifier=self,
label='image gallery',
parentSizer=self.gallery_sizer)
self.gallery_sizer.Add(self.galleryBin, proportion=1, flag=wx.EXPAND)
self.gallery_panel.SetSizer(self.gallery_sizer)
self.objects_bin_panel = wx.Panel(self.bins_splitter)
# fetch objects interface
self.startId = wx.TextCtrl(self.fetch_panel, id=-1, value='1', size=(60, -1), style=wx.TE_PROCESS_ENTER)
self.endId = wx.TextCtrl(self.fetch_panel, id=-1, value='100', size=(60, -1), style=wx.TE_PROCESS_ENTER)
self.fetchChoice = wx.Choice(self.fetch_panel, id=-1, choices=['range','all','individual'])
self.fetchChoice.SetSelection(0)
self.filterChoice = wx.Choice(self.fetch_panel, id=-1,
choices=['experiment'] + p._filters_ordered + p._groups_ordered + [
CREATE_NEW_FILTER])
self.fetchFromGroupSizer = wx.BoxSizer(wx.HORIZONTAL)
self.fetchBtn = wx.Button(self.fetch_panel, -1, 'Fetch!')
#### Create Sizers
self.fetchSizer = wx.BoxSizer(wx.HORIZONTAL)
self.find_rules_sizer = wx.BoxSizer(wx.HORIZONTAL)
self.fetch_and_rules_sizer = wx.BoxSizer(wx.VERTICAL)
self.classified_bins_sizer = wx.BoxSizer(wx.HORIZONTAL)
#### Add elements to sizers and splitters
# fetch panel
self.fetchSizer.AddStretchSpacer()
self.fetchSizer.Add(wx.StaticText(self.fetch_panel, -1, 'Fetch '), flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(self.fetchChoice, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchTxt = wx.StaticText(self.fetch_panel, -1, label='of image IDs:')
self.fetchSizer.Add(self.fetchTxt, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(self.startId, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchTxt2 = wx.StaticText(self.fetch_panel, -1, label='to')
self.fetchSizer.Add(self.fetchTxt2, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(self.endId, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
#self.fetchSizer.Add(self.obClassChoice, flag=wx.ALIGN_CENTER_VERTICAL)
#self.fetchSizer.AddSpacer((5, 20))
self.fetchTxt3 = wx.StaticText(self.fetch_panel, -1, label='images')
self.fetchSizer.Add(self.fetchTxt3, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(wx.StaticText(self.fetch_panel, -1, 'from'), flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(self.filterChoice, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((10, 20))
self.fetchSizer.Add(self.fetchFromGroupSizer, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddSpacer((5, 20))
self.fetchSizer.Add(self.fetchBtn, flag=wx.ALIGN_CENTER_VERTICAL)
self.fetchSizer.AddStretchSpacer()
self.fetch_panel.SetSizerAndFit(self.fetchSizer)
# fetch and rules panel
self.fetch_and_rules_sizer.Add((5, 5))
self.fetch_and_rules_sizer.Add(self.fetch_panel, flag=wx.EXPAND)
self.fetch_and_rules_sizer.Add((5, 5))
self.fetch_and_rules_panel.SetSizerAndFit(self.fetch_and_rules_sizer)
# classified bins panel
self.objects_bin_panel.SetSizer(self.classified_bins_sizer)
# splitter windows
self.splitter.SplitHorizontally(self.fetch_and_rules_panel, self.bins_splitter,
self.fetch_and_rules_panel.GetMinSize()[1])
self.bins_splitter.SplitHorizontally(self.gallery_panel, self.objects_bin_panel)
self.splitter.SetSashGravity(0.0)
self.bins_splitter.SetSashGravity(0.5)
self.splitter.SetMinimumPaneSize(max(50, self.fetch_and_rules_panel.GetMinHeight()))
self.bins_splitter.SetMinimumPaneSize(50)
self.SetMinSize((self.fetch_and_rules_panel.GetMinWidth(), 4 * 50 + self.fetch_and_rules_panel.GetMinHeight()))
# Set initial state
self.filterChoice.SetSelection(0)
# JEN - Start Add
# self.openDimensReduxBtn.Disable()
# JEN - End Add
self.fetchSizer.Hide(self.fetchFromGroupSizer)
#####################
#### GUI Section ####
#####################
# add the default classes
#for class in range(1, num_classes+1):
self.AddSortClass('objects of selected image')
#self.AddSortClass('negative')
self.Layout()
self.Center()
self.MapChannels(p.image_channel_colors[:])
self.BindMouseOverHelpText()
#self.Bind(wx.EVT_BUTTON, self.OnInspect, self.inspectBtn)
# JEN - Start Add
# self.Bind(wx.EVT_BUTTON, self.OpenDimensRedux, self.openDimensReduxBtn)
# JEN - End Add
self.Bind(wx.EVT_BUTTON, self.OnFetch, self.fetchBtn)
self.startId.Bind(wx.EVT_TEXT, self.ValidateIntegerField)
self.startId.Bind(wx.EVT_TEXT_ENTER, self.OnFetch)
self.Bind(wx.EVT_CLOSE, self.OnClose)
self.Bind(wx.EVT_CHAR, self.OnKey) # Doesn't work for windows
tilecollection.EVT_TILE_UPDATED(self, self.OnTileUpdated)
self.Bind(sortbin.EVT_QUANTITY_CHANGED, self.QuantityChanged)
self.Bind(wx.EVT_CHOICE, self.OnSelectFetchChoice, self.fetchChoice)
self.Bind(wx.EVT_CHOICE, self.OnSelectFilter, self.filterChoice)
def FormatPlateMapData(keys_and_vals, categorical=False):
'''
keys_and_vals -- a list of lists of well-keys and values
eg: [['p1', 'A01', 0.2],
['p1', 'A02', 0.9], ...]
returns a 2-tuple containing:
-an array in the shape of the plate containing the given values with
NaNs filling empty slots. If multiple sites per-well are given, then
the array will be shaped (rows, cols, sites)
-an array in the shape of the plate containing the given keys with
(UnknownPlate, UnknownWell) filling empty slots
'''
from itertools import groupby
keys_and_vals = np.array(keys_and_vals)
nkeycols = len(dbconnect.well_key_columns())
shape = list(p.plate_shape)
if p.plate_type == '5600':
well_keys = keys_and_vals[:,:-1] # first column(s) are keys
data = keys_and_vals[:,-1] # last column is data
assert data.ndim == 1
if len(data) < 5600: raise Exception(
'''The measurement you chose to plot was missing for some spots.
Because CPA doesn't know the well labelling convention used by this
microarray, we can't be sure how to plot the data. If you are
plotting an object measurement, you may have some spots with 0
objects and therefore no entry in the table.''')
assert len(data) == 5600
data = np.array(list(meander(data.reshape(shape)))).reshape(shape)
sort_indices = np.array(list(meander(np.arange(np.prod(shape)).reshape(shape)))).reshape(shape)
well_keys = np.array(list(meander(well_keys.reshape(shape + [nkeycols] )))).reshape(shape + [nkeycols])
return data, well_keys, sort_indices
# compute the number of sites-per-well as the max number of rows with the same well-key
nsites = max([len(list(grp))
for k, grp in groupby(keys_and_vals,
lambda row: tuple(row[:nkeycols]))
])
if nsites > 1:
# add a sites dimension to the array shape if there's >1 site per well
shape += [nsites]
data = np.ones(shape) * np.nan
if categorical:
data = data.astype('object')
if p.plate_id:
dummy_key = ('UnknownPlate', 'UnknownWell')
else:
dummy_key = ('UnknownWell',)
well_keys = np.array([dummy_key] * np.prod(shape),
dtype=object).reshape(shape + [nkeycols])
sort_indices = np.ones(data.shape)*np.nan
dm = DataModel.getInstance()
ind = keys_and_vals.argsort(axis=0)
for i, (k, well_grp) in enumerate(groupby(keys_and_vals[ind[:,len(dummy_key)-1],:],
lambda row: tuple(row[:len(dummy_key)]))):
(row, col) = dm.get_well_position_from_name(k[-1])
well_data = np.array(list(well_grp))[:,-1]
if len(well_data) == 1:
data[row, col] = well_data[0]
sort_indices[row,col] = ind[:,len(dummy_key)-1][i]
else:
data[row, col] = well_data
sort_indices[row,col] = ind[:,len(dummy_key)-1][i*nsites + np.array(range(nsites))]
well_keys[row, col] = k
return data, well_keys, sort_indices
def score(properties, ts, nRules, filter_name=None, group='Image',
show_results=False, results_table=None, overwrite=False):
'''
Trains a Classifier on a training set and scores the experiment
returns the table of scores as a numpy array.
properties -- Properties instance
ts -- TrainingSet instance
nRules -- number of rules to use
filter_name -- name of a filter to use from the properties file
group -- name of a group to use from the properties file
show_results -- whether or not to show the results in TableViewer
results_table -- table name to save results to or None.
'''
p = properties
db = DBConnect.getInstance()
dm = DataModel.getInstance()
if group == None:
group = 'Image'
if results_table:
if db.table_exists(results_table) and not overwrite:
print 'Table "%s" already exists. Delete this table before running scoreall.'%(results_table)
return None
print ''
print 'properties: ', properties
print 'training set: ', ts
print '# rules: ', nRules
print 'filter: ', filter_name
print 'grouping by: ', group
print 'show results: ', show_results
print 'results table: ', results_table
print 'overwrite: ', overwrite
print ''
nClasses = len(ts.labels)
nKeyCols = len(image_key_columns())
assert 200 > nRules > 0, '# of rules must be between 1 and 200. Value was %s'%(nRules,)
assert filter_name in p._filters.keys()+[None], 'Filter %s not found in properties file. Valid filters are: %s'%(filter_name, ','.join(p._filters.keys()),)
assert group in p._groups.keys()+['Image'], 'Group %s not found in properties file. Valid groups are: %s'%(group, ','.join(p._groups.keys()),)
output = StringIO()
logging.info('Training classifier with %s rules...'%nRules)
t0 = time()
weaklearners = fastgentleboostingmulticlass.train(ts.colnames,
nRules, ts.label_matrix,
ts.values, output)
logging.info('Training done in %f seconds'%(time()-t0))
logging.info('Computing per-image class counts...')
t0 = time()
def update(frac):
logging.info('%d%% '%(frac*100.,))
keysAndCounts = multiclasssql.PerImageCounts(weaklearners, filter_name=(filter_name or None), cb=update)
keysAndCounts.sort()
logging.info('Counts found in %f seconds'%(time()-t0))
if not keysAndCounts:
logging.error('No images are in filter "%s". Please check the filter definition in your properties file.'%(filter_name))
raise Exception('No images are in filter "%s". Please check the filter definition in your properties file.'%(filter_name))
# AGGREGATE PER_IMAGE COUNTS TO GROUPS IF NOT GROUPING BY IMAGE
if group != 'Image':
logging.info('Grouping %s counts by %s...' % (p.object_name[0], group))
t0 = time()
imData = {}
for row in keysAndCounts:
key = tuple(row[:nKeyCols])
imData[key] = np.array([float(v) for v in row[nKeyCols:]])
groupedKeysAndCounts = np.array([list(k)+vals.tolist() for k, vals in dm.SumToGroup(imData, group).items()], dtype=object)
nKeyCols = len(dm.GetGroupColumnNames(group))
logging.info('Grouping done in %f seconds'%(time()-t0))
else:
groupedKeysAndCounts = np.array(keysAndCounts, dtype=object)
# FIT THE BETA BINOMIAL
logging.info('Fitting beta binomial distribution to data...')
counts = groupedKeysAndCounts[:,-nClasses:]
alpha, converged = polyafit.fit_betabinom_minka_alternating(counts)
logging.info(' alpha = %s converged = %s'%(alpha, converged))
logging.info(' alpha/Sum(alpha) = %s'%([a/sum(alpha) for a in alpha]))
# CONSTRUCT ARRAY OF TABLE DATA
logging.info('Computing enrichment scores for each group...')
t0 = time()
tableData = []
for i, row in enumerate(groupedKeysAndCounts):
# Start this row with the group key:
tableRow = list(row[:nKeyCols])
if group != 'Image':
tableRow += [len(dm.GetImagesInGroup(group, tuple(row[:nKeyCols])))]
# Append the counts:
countsRow = [int(v) for v in row[nKeyCols:nKeyCols+nClasses]]
tableRow += [sum(countsRow)]
tableRow += countsRow
if p.area_scoring_column is not None:
# Append the areas
countsRow = [int(v) for v in row[-nClasses:]]
tableRow += [sum(countsRow)]
tableRow += countsRow
# Append the scores:
# compute enrichment probabilities of each class for this image OR group
scores = np.array( dirichletintegrate.score(alpha, np.array(countsRow)) )
# clamp to [0,1] to
scores[scores>1.] = 1.
scores[scores<0.] = 0.
tableRow += scores.tolist()
# Append the logit scores:
# Special case: only calculate logit of "positives" for 2-classes
if nClasses==2:
tableRow += [np.log10(scores[0])-(np.log10(1-scores[0]))] # compute logit of each probability
else:
tableRow += [np.log10(score)-(np.log10(1-score)) for score in scores] # compute logit of each probability
tableData.append(tableRow)
tableData = np.array(tableData, dtype=object)
logging.info('Enrichments computed in %f seconds'%(time()-t0))
# CREATE COLUMN LABELS LIST
# if grouping isn't per-image, then get the group key column names.
if group != 'Image':
colnames = dm.GetGroupColumnNames(group)
else:
colnames = list(image_key_columns())
# record the column indices for the keys
key_col_indices = [i for i in range(len(colnames))]
if group != 'Image':
colnames += ['Number_of_Images']
colnames += ['Total_%s_Count'%(p.object_name[0].capitalize())]
for i in xrange(nClasses):
colnames += ['%s_%s_Count'%(ts.labels[i].capitalize(), p.object_name[0].capitalize())]
if p.area_scoring_column is not None:
colnames += ['Total_%s_Area'%(p.object_name[0].capitalize())]
for i in xrange(nClasses):
colnames += ['%s_%s_Area'%(ts.labels[i].capitalize(), p.object_name[0].capitalize())]
for i in xrange(nClasses):
colnames += ['pEnriched_%s'%(ts.labels[i])]
if nClasses==2:
colnames += ['Enriched_Score_%s'%(ts.labels[0])]
else:
for i in xrange(nClasses):
colnames += ['Enriched_Score_%s'%(ts.labels[i])]
title = results_table or "Enrichments_per_%s"%(group,)
if filter_name:
title += "_filtered_by_%s"%(filter_name,)
title += ' (%s)'%(os.path.split(p._filename)[1])
if results_table:
print 'Creating table %s'%(results_table)
success = db.CreateTableFromData(tableData, colnames, results_table, temporary=False)
if not success:
print 'Failed to create results table :('
if show_results:
import tableviewer
tableview = tableviewer.TableViewer(None, title=title)
if results_table and overwrite:
tableview.load_db_table(results_table)
else:
tableview.table_from_array(tableData, colnames, group, key_col_indices)
tableview.set_fitted_col_widths()
tableview.Show()
return tableData
self.tile_collection.cv.acquire()
heappush(self.tile_collection.loadq, ((0, 0, 0), '<ABORT>'))
self.tile_collection.cv.notify()
self.tile_collection.cv.release()
################# FOR TESTING ##########################
if __name__ == "__main__":
app = wx.PySimpleApp()
from datamodel import DataModel
p = Properties.getInstance()
p.LoadFile('../properties/nirht_test.properties')
db = DBConnect.getInstance()
db.connect()
dm = DataModel.getInstance()
test = TileCollection.getInstance()
f = wx.Frame(None)
for i in xrange(10):
obKey = dm.GetRandomObject()
test.GetTileData((0, 1, 1), f)
for t in threading.enumerate():
if t != threading.currentThread():
t.abort()
f.Destroy()
app.MainLoop()
def score_objects(properties,
ts,
gt,
nRules,
filter_name=None,
group='Image',
show_results=False,
results_table=None,
overwrite=False):
'''
Trains a Classifier on a training set and scores the experiment
returns the table of scores as a numpy array.
properties -- Properties instance
ts -- TrainingSet instance
gt -- Ground Truth instance
nRules -- number of rules to use
filter_name -- name of a filter to use from the properties file
group -- name of a group to use from the properties file
show_results -- whether or not to show the results in TableViewer
results_table -- table name to save results to or None.
'''
p = properties
#db = DBConnect.getInstance() ## Removed writing to db. Results_table should be 'None' anyway
dm = DataModel.getInstance()
#if group == None:
#group = 'Image'
if results_table:
if db.table_exists(results_table) and not overwrite:
print 'Table "%s" already exists. Delete this table before running scoreall.' % (
results_table)
return None
print ''
print 'properties: ', properties
print 'initial training set: ', ts
print 'ground truth training set: ', gt
print '# rules: ', nRules
print 'filter: ', filter_name
print 'grouping by: ', group
print 'show results: ', show_results
print 'results table: ', results_table
print 'overwrite: ', overwrite
print ''
nClasses = len(ts.labels)
nKeyCols = len(image_key_columns())
assert 200 > nRules > 0, '# of rules must be between 1 and 200. Value was %s' % (
nRules, )
assert filter_name in p._filters.keys() + [
None
], 'Filter %s not found in properties file. Valid filters are: %s' % (
filter_name,
','.join(p._filters.keys()),
)
assert group in p._groups.keys() + [
'Image', 'None'
], 'Group %s not found in properties file. Valid groups are: %s' % (
group,
','.join(p._groups.keys()),
)
output = StringIO()
logging.info('Training classifier with %s rules...' % nRules)
t0 = time()
weaklearners = fastgentleboostingmulticlass.train(ts.colnames, nRules,
ts.label_matrix,
ts.values, output)
logging.info('Training done in %f seconds' % (time() - t0))
t0 = time()
#def update(frac):
#logging.info('%d%% '%(frac*100.,))
## Score Ground Truth using established classifier
gt_predicted_scores = per_cell_scores(weaklearners, gt.values, gt.colnames)
#plt.hist(gt_predicted_scores)
#plt.show()
gt_predicted_signs = np.sign(gt_predicted_scores)
## Compare Ground Truth score signs with the actual ground truth values
numclasses = ts.labels.size
gt_actual_signs = gt.label_matrix[:, 0]
cm_unrotated = metrics.confusion_matrix(gt_actual_signs,
gt_predicted_signs)
## sklearn.metrics.confusion_matrix -- 2D confusion matrix is inverted from convention.
## https://github.com/scikit-learn/scikit-learn/issues/1664
cm = np.rot90(np.rot90(cm_unrotated))
fpr, sens, thresholds = metrics.roc_curve(gt_actual_signs,
gt_predicted_signs)
spec = 1 - fpr
s = np.sum(cm, axis=1)
percent = [100 * cm[i, i] / float(s[i]) for i in range(len(s))]
avg = np.mean(percent)
avgTotal = 100 * np.trace(cm) / float(np.sum(cm))
print 'accuracy = %f' % avgTotal
print 'Confusion Matrix = ... '
print cm
my_sens = cm[0, 0] / float(cm[0, 0] + cm[0, 1]) #TP/(TP+FN)
my_spec = cm[1, 1] / float(cm[1, 1] + cm[1, 0]) #TN/(TN+FP)
print 'My_Sensitivity = %f' % my_sens
print 'My_Specificity = %f' % my_spec
print 'Sensitivity = ...'
print sens
print 'Specificity = ...'
print spec
print 'Done calculating'
############
## Confusion Matrix code from here: http://stackoverflow.com/questions/5821125/how-to-plot-confusion-matrix-with-string-axis-rather-than-integer-in-python
conf_arr = cm
norm_conf = []
## This normalizes each *row* to the color map, but I chose to normalize the whole confusion matrix to the same scale
##for i in conf_arr:
##a = 0
##tmp_arr = []
##a = sum(i, 0)
##for j in i:
##tmp_arr.append(float(j)/float(a))
##norm_conf.append(tmp_arr)
norm_conf = conf_arr / float(np.max(conf_arr))
if DISPLAY_CONFUSION_MATRIX:
fig = plt.figure()
plt.clf()
ax = fig.add_subplot(111)
ax.set_aspect(1)
res = ax.imshow(np.array(norm_conf),
cmap=plt.cm.jet,
interpolation='nearest')
width = len(conf_arr)
height = len(conf_arr[0])
for x in xrange(width):
for y in xrange(height):
ax.annotate(str(conf_arr[x][y]),
xy=(y, x),
horizontalalignment='center',
verticalalignment='center')
cb = fig.colorbar(res)
#cb.set_cmap = [0,1]
if width == 2 and height == 2:
plt.xticks([0, 1], ['FP', 'TN'])
plt.yticks([0, 1], ['TP', 'FP'])
else:
alphabet = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
plt.xticks(range(width), alphabet[:width])
plt.yticks(range(height), alphabet[:height])
plt.show()
print 'Done'
