def test_LDATransform( self ):
"""LDA transform"""
tempdir = mkdtemp()
import zipfile
zf = zipfile.ZipFile( pychrm_test_dir + sep + 'lymphoma_iicbu2008_subset_EOSIN_ONLY_t5x6_v3.2features.fit.zip', mode='r')
zf.extractall( tempdir )
fitfile_path = tempdir + sep + 'lymphoma_iicbu2008_subset_eosin_t5x6_v3.2features.fit'
try:
kwargs = {}
kwargs['pathname'] = fitfile_path
kwargs['quiet'] = True
# sampling opts: -l -t5x6 implies 5 columns and 6 rows ... I know it's weird.
kwargs['long'] = True
kwargs['tile_num_rows'] = 6
kwargs['tile_num_cols'] = 5
# against self:
fs = FeatureSpace.NewFromFitFile( **kwargs )
self_transformed = fs.LDATransform( reference_features=None, inplace=False )
fit_on_fit_LDA_result = FeatureSpaceClassification.NewWND5(
self_transformed, self_transformed, feature_weights=None )
# against other:
train, test = fs.Split()
train.LDATransform( reference_features=None, inplace=True )
test.LDATransform( reference_features=train, inplace=True )
split_LDA_result = FeatureSpaceClassification.NewWND5(
train, test, feature_weights=None )
finally:
rmtree( tempdir )
def test_FitOnFitClassification( self ):
fitfile_path = wndchrm_test_dir + sep + 'test-l.fit'
#fs = FeatureSet.NewFromFitFile( fitfile_path )
fs = FeatureSpace.NewFromFitFile( fitfile_path )
fs.Normalize( inplace=True, quiet=True )
fw = FisherFeatureWeights.NewFromFeatureSpace( fs ).Threshold(438)
fw.Print(50)
fs.FeatureReduce( fw, inplace=True )
pychrm_split = FeatureSpaceClassification.NewWND5( fs, fs, fw, quiet=False )
pychrm_split = FeatureSpaceClassification.NewWND5_OLD( fs, fs, fw, quiet=False )
from wndcharm.FeatureSpacePredictionExperiment import FeatureSpaceClassificationExperiment
html_path = pychrm_test_dir + sep + 'test-l_training_error_result.html'
html_exp = FeatureSpaceClassificationExperiment.NewFromHTMLReport( html_path, quiet=False )
# single split in this html
html_split = html_exp.individual_results[0]
for i, (html_result, pychrm_result) in enumerate( zip( html_split.individual_results,\
pychrm_split.individual_results ) ):
try:
self.assertEqual( html_result, pychrm_result )
except:
outstr = "Error in comparison # {0}:\n".format( i )
outstr += "HTML result:\n{0}\n Python API res:\n{1}".format( html_result, pychrm_result)
raise
def test_DiscreteTrainTestSplitNoTiling( self ):
"""Uses binucleate test set"""
fitfilepath = wndchrm_test_dir + sep + 'test-l.fit'
fs = FeatureSpace.NewFromFitFile( fitfilepath )
from numpy.random import RandomState
prng = RandomState(42)
full_train, full_test = fs.Split( random_state=prng, quiet=True )
full_train.Normalize( quiet=True )
reduced_fw = FisherFeatureWeights.NewFromFeatureSpace( full_train ).Threshold()
reduced_train = full_train.FeatureReduce( reduced_fw )
reduced_test = full_test.FeatureReduce( reduced_fw )
reduced_test.Normalize( reduced_train, quiet=True )
batch_result = FeatureSpaceClassification.NewWND5( reduced_train,
reduced_test, reduced_fw, quiet=True )
def test_IfNotInterpolatable( self ):
"""You can't graph predicted values if the classes aren't interpolatable."""
testfilename = 'ShouldntBeGraphable.png'
small_fs = CreateArtificialFeatureSpace_Discrete(
n_samples=20, n_classes=2, random_state=42, interpolatable=False )
train_set, test_set = small_fs.Split( random_state=False, quiet=True )
train_set.Normalize()
fw = FisherFeatureWeights.NewFromFeatureSpace( train_set ).Threshold()
reduced_train_set = train_set.FeatureReduce( fw )
reduced_test_set = test_set.FeatureReduce( fw )
test_set.Normalize( train_set, quiet=True )
batch_result = FeatureSpaceClassification.NewWND5(
reduced_train_set, reduced_test_set, fw, quiet=True )
with self.assertRaises( ValueError ):
graph = PredictedValuesGraph( batch_result )
def test_TiledTrainTestSplit(self):
"""Uses a fake FeatureSpace"""
from wndcharm.ArtificialFeatureSpace import CreateArtificialFeatureSpace_Discrete
fs_kwargs = {}
fs_kwargs['name'] = "DiscreteArtificialFS 10-class"
fs_kwargs['n_samples'] = 1000
fs_kwargs['n_classes'] = 10 # 100 samples per class
fs_kwargs['num_features_per_signal_type'] = 25
fs_kwargs['initial_noise_sigma'] = 40
fs_kwargs['noise_gradient'] = 20
fs_kwargs['n_samples_per_group'] = 4 # 25 images, 2x2 tiling scheme
fs_kwargs['interpolatable'] = True
fs_kwargs['random_state'] = 43
fs_kwargs['singularity'] = False
fs_kwargs['clip'] = False
fs = CreateArtificialFeatureSpace_Discrete(**fs_kwargs)
train, test = fs.Split(random_state=False, quiet=True)
train.Normalize(inplace=True, quiet=True)
fw = FisherFeatureWeights.NewFromFeatureSpace(train).Threshold()
train.FeatureReduce(fw, inplace=True)
test.FeatureReduce(fw, inplace=True,
quiet=True).Normalize(train,
inplace=True,
quiet=True)
result = FeatureSpaceClassification.NewWND5(train, test, fw)
result.Print()
for class_name in result.test_set.class_names:
try:
self.assertEqual(
result.similarity_matrix[class_name][class_name], float(1))
except:
print "offending class: {0}, val: {1}".format(
class_name,
result.similarity_matrix[class_name][class_name])
raise
def __init__( self, training_set, feature_weights, test_image_path,
chart_title=None, max_num_features=300 ):
self.timing_axes = None
import time
timings = []
from wndcharm.FeatureSpacePredictionExperiment import FeatureSpaceClassificationExperiment
from wndcharm.SingleSamplePrediction import SingleSampleClassification
from wndcharm.FeatureSpacePrediction import FeatureSpaceClassification
experiment = FeatureSpaceClassificationExperiment( training_set, training_set, feature_weights )
for number_of_features_to_use in range( 1, max_num_features + 1 ):
reduced_ts = None
reduced_fw = None
three_timings = []
# Take the best of 3
for timing in range( 3 ):
# Time the creation and classification of a single signature
t1 = time.time()
reduced_fw = feature_weights.Threshold( number_of_features_to_use )
sig = FeatureVector( source_filepath=test_image_path, feature_names=reduced_fw.feature_names ).GenerateFeatures()
reduced_ts = training_set.FeatureReduce( reduced_fw )
sig.Normalize( reduced_ts )
result = SingleSampleClassification.NewWND5( reduced_ts, reduced_fw, sig )
result.Print()
# FIXME: save intermediates just in case of interruption or parallization
# result.PickleMe()
t2 = time.time()
three_timings.append( t2 - t1 )
timings.append( min( three_timings ) )
# now, do a fit-on-fit test to measure classification accuracy
split_result = FeatureSpaceClassification.NewWND5( reduced_ts, reduced_ts, reduced_fw )
split_result.Print()
experiment.individual_results.append( split_result )
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
x_vals = list( range( 1, max_num_features + 1 ) )
self.figure = plt.figure()
self.main_axes = self.figure.add_subplot(111)
if chart_title == None:
self.chart_title = "Feature timing v. classification accuracy"
else:
self.chart_title = chart_title
self.main_axes.set_title( self.chart_title )
self.main_axes.set_xlabel( 'Number of features' )
self.main_axes.set_ylabel( 'Classification accuracy (%)', color='b' )
classification_accuracies = \
[ split_result.classification_accuracy * 100 for split_result in experiment.individual_results ]
self.main_axes.plot( x_vals, classification_accuracies, color='b', linewidth=2 )
for tl in self.main_axes.get_yticklabels():
tl.set_color('b')
self.timing_axes = self.main_axes.twinx()
self.timing_axes.set_ylabel( 'Time to calculate features (s)', color='r' )
self.timing_axes.plot( x_vals, timings, color='r' )
for tl in self.timing_axes.get_yticklabels():
tl.set_color('r')
class TestGraphs(unittest.TestCase):
"""Test WND-CHARM's graph-making functionality."""
fs_kwargs = {}
fs_kwargs['name'] = "DiscreteArtificialFS 10-class"
fs_kwargs['n_samples'] = 1000
fs_kwargs['n_classes'] = 10
fs_kwargs['num_features_per_signal_type'] = 25
fs_kwargs['initial_noise_sigma'] = 40
fs_kwargs['noise_gradient'] = 20
fs_kwargs['n_samples_per_group'] = 1
fs_kwargs['interpolatable'] = True
fs_kwargs['random_state'] = 43
fs_kwargs['singularity'] = False
fs_kwargs['clip'] = False
fs = CreateArtificialFeatureSpace_Discrete(**fs_kwargs)
train_set, test_set = fs.Split(random_state=False, quiet=True)
train_set.Normalize(quiet=True)
fw = FisherFeatureWeights.NewFromFeatureSpace(train_set).Threshold()
reduced_train_set = train_set.FeatureReduce(fw)
reduced_test_set = test_set.FeatureReduce(fw)
reduced_test_set.Normalize(reduced_train_set, quiet=True)
batch_result = FeatureSpaceClassification.NewWND5(reduced_train_set,
reduced_test_set,
fw,
quiet=True)
def setUp(self):
self.tempdir = mkdtemp()
def tearDown(self):
rmtree(self.tempdir)
def CompareGraphs(self, graph, testfilename):
"""Helper function to check output graphs"""
# Uncoment to see what graph looks like!
#graph.SaveToFile( testfilename + 'GRAPH.png' )
# We used to output the graphs to a png file and do a binary diff on a reference png
# but there are superficial differences between matplotlib versions that result in
# the points still being in the right place, but the font is slightly larger,
# or the text is subtlely offset. So now, we interrogate the matplotlib.figure
# object and retrieve its coordinates and check them against blessed numpy arrays
# saved to a npy file.
axessubplot = graph.figure.gca()
if len(axessubplot.lines) > 0:
# line plot
try:
all_coords = np.dstack(
tuple([
group._path._vertices for group in axessubplot.lines
]))
except AttributeError:
# older version of matplotlib didn't include leading underscore in attribute
# "_vertices"
all_coords = np.dstack(
tuple(
[group._path.vertices for group in axessubplot.lines]))
elif len(axessubplot.collections) > 0:
# scatter plot
all_coords = np.dstack(
tuple([group._offsets for group in axessubplot.collections]))
else:
self.fail("Graph doesn't have any lines nor points")
# uncomment to replace old coords
#np.save( testfilename, all_coords )
#from os.path import splitext
#testfilename_base, ext = splitext( testfilename )
#np.save( testfilename_base + 'NEW.npy', all_coords )
reference_array = np.load(testfilename)
if not np.array_equal(all_coords, reference_array):
if not np.allclose(all_coords, reference_array):
errmsg = 'Reference graph "{0}" coordinates '.format(testfilename) + \
'do not concur with coordinates generated by this test.'
self.fail(errmsg)
@unittest.skipIf(HasMatplotlib, "Skipped if matplotlib IS installed")
def test_ErrMsgIfMatplotibNotInstalled(self):
"""Fail gracefully with informative message if matplotlib"""
graph = PredictedValuesGraph(self.batch_result)
with self.assertRaises(ImportError):
graph.RankOrderedPredictedValuesGraph()
with self.assertRaises(ImportError):
graph.KernelSmoothedDensityGraph()
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOT installed")
@unittest.expectedFailure
def test_RankOrderedFromBatchClassificationResult(self):
"""Rank Ordered Predicted values graph from a single split"""
testfilename = 'test_graph_rank_ordered_interpolated_discrete.npy'
graph = PredictedValuesGraph(self.batch_result)
graph.RankOrderedPredictedValuesGraph()
self.CompareGraphs(graph, testfilename)
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOT installed")
@unittest.expectedFailure
def test_KernelSmoothedFromBatchClassificationResult(self):
"""Kernel Smoothed Probability density graph from a single split"""
testfilename = 'test_graph_kernel_smoothed.npy'
graph = PredictedValuesGraph(self.batch_result)
graph.KernelSmoothedDensityGraph()
self.CompareGraphs(graph, testfilename)
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOT installed")
def test_FromDiscreteClassificationExperimentResults(self):
"""Rank Ordered Predicted values graph from an experiment result (multiple splits)"""
testfilename = 'test_graph_rank_ordered_experiment.npy'
# Make a smaller featureset to do multiple splits
fs_kwargs = {}
fs_kwargs['name'] = "DiscreteArtificialFS RANK ORDERED SHUFFLE SPLIT"
fs_kwargs['n_samples'] = 100 # smaller
fs_kwargs['n_classes'] = 5 # smaller, 20 samples per class
fs_kwargs['num_features_per_signal_type'] = 10 # smaller
fs_kwargs['initial_noise_sigma'] = 50
fs_kwargs['noise_gradient'] = 20
fs_kwargs['n_samples_per_group'] = 1
fs_kwargs['interpolatable'] = True
fs_kwargs['random_state'] = 42
fs_kwargs['singularity'] = False
fs_kwargs['clip'] = False
small_fs = CreateArtificialFeatureSpace_Discrete(**fs_kwargs)
ss_kwargs = {}
ss_kwargs['quiet'] = True
ss_kwargs['n_iter'] = n_iter = 10
ss_kwargs['train_size'] = train_size = 18 # per-class
ss_kwargs['test_size'] = test_size = 2 # per-class
ss_kwargs['random_state'] = 42
exp = FeatureSpaceClassificationExperiment.NewShuffleSplit(
small_fs, **ss_kwargs)
graph = PredictedValuesGraph(exp, use_averaged_results=False)
graph.RankOrderedPredictedValuesGraph()
self.CompareGraphs(graph, testfilename)
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOT installed")
def test_HyperparameterOptimizationGraph(self):
"""Accuracy vs. # features or samples with and without LDA feature space transform"""
testfilename = 'test_graph_rank_ordered_experiment.npy'
# Make a smaller featureset to do multiple splits
fs_kwargs = {}
fs_kwargs['name'] = "DiscreteArtificialFS RANK ORDERED SHUFFLE SPLIT"
fs_kwargs['n_samples'] = 100 # smaller
fs_kwargs['n_classes'] = 5 # smaller, 20 samples per class
fs_kwargs['num_features_per_signal_type'] = 10 # smaller
fs_kwargs['initial_noise_sigma'] = 50
fs_kwargs['noise_gradient'] = 20
fs_kwargs['n_samples_per_group'] = 1
fs_kwargs['interpolatable'] = True
fs_kwargs['random_state'] = 42
fs_kwargs['singularity'] = False
fs_kwargs['clip'] = False
small_fs = CreateArtificialFeatureSpace_Discrete(**fs_kwargs)
ss_kwargs = {}
ss_kwargs['quiet'] = False
ss_kwargs['n_iter'] = n_iter = 10
ss_kwargs['train_size'] = train_size = 18 # per-class
ss_kwargs['test_size'] = test_size = 2 # per-class
ss_kwargs['random_state'] = 42
ss_kwargs['show_raw'] = True
ss_kwargs['show_lda'] = True
ss_kwargs['param'] = 'features'
ss_kwargs['text_angle'] = -30
graph = HyperparameterOptimizationGraph(small_fs)
graph.GridSearch(**ss_kwargs)
#graph.savefig( '/Users/colettace/test_features.png' )
ss_kwargs['param'] = 'samples'
ss_kwargs['quiet'] = False
ss_kwargs['text_angle'] = -30
graph = HyperparameterOptimizationGraph(small_fs)
graph.GridSearch(**ss_kwargs)
#graph.savefig( '/Users/colettace/test_samples.png' )
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOT installed")
def test_FromHTML(self):
"""Rank Ordered Predicted values graph from an experiment result (multiple splits)"""
testfilename = 'test_graph_fromHTML.npy'
# Inflate the zipped html file into a temp file
import zipfile
#zipped_file_path = pychrm_test_dir + sep + 'c_elegans_terminal_bulb.html'
#import zlib
#zf = zipfile.ZipFile( zipped_file_path + '.zip', mode='w' )
#zf.write( zipped_file_path, compress_type=zipfile.ZIP_DEFLATED )
#zf.close()
zipped_file_path = pychrm_test_dir + sep + 'c_elegans_terminal_bulb.html.zip'
zf = zipfile.ZipFile(zipped_file_path, mode='r')
zf.extractall(self.tempdir)
htmlfilepath = self.tempdir + sep + zf.namelist()[0]
graph = PredictedValuesGraph.NewFromHTMLReport(
htmlfilepath, use_averaged_results=False)
graph.RankOrderedPredictedValuesGraph()
self.CompareGraphs(graph, testfilename)
@unittest.skipUnless(HasMatplotlib,
"Skipped if matplotlib IS NOTinstalled")
def test_IfNotInterpolatable(self):
"""You can't graph predicted values if the classes aren't interpolatable."""
testfilename = 'ShouldntBeGraphable.png'
small_fs = CreateArtificialFeatureSpace_Discrete(n_samples=20,
n_classes=2,
random_state=42,
interpolatable=False)
train_set, test_set = small_fs.Split(random_state=False, quiet=True)
train_set.Normalize()
fw = FisherFeatureWeights.NewFromFeatureSpace(train_set).Threshold()
reduced_train_set = train_set.FeatureReduce(fw)
reduced_test_set = test_set.FeatureReduce(fw)
test_set.Normalize(train_set, quiet=True)
batch_result = FeatureSpaceClassification.NewWND5(reduced_train_set,
reduced_test_set,
fw,
quiet=True)
with self.assertRaises(ValueError):
graph = PredictedValuesGraph(batch_result)
def test_FitOnFit(self):
"""Uses a curated subset of the IICBU 2008 Lymphoma dataset, preprocessed as follows:
auto-deconvolved, eosin channel only, tiled 5x6, 3 classes, 10 imgs per class,
300 samples per class.
"""
# Inflate the zipped test fit into a temp file
import zipfile
zipped_file_path = pychrm_test_dir + sep + 'lymphoma_iicbu2008_subset_EOSIN_ONLY_t5x6_v3.2features.fit.zip'
zf = zipfile.ZipFile(zipped_file_path, mode='r')
tempdir = mkdtemp()
zf.extractall(tempdir)
try:
fitfilepath = tempdir + sep + zf.namelist()[0]
# Do fit on fit WITHOUT tiling and compare with fit on fit results
# generated with wndchrm 1.60
fs = FeatureSpace.NewFromFitFile(fitfilepath).Normalize(
inplace=True, quiet=True)
#fs = FeatureSpace.NewFromFitFile( wndchrm_test_dir + sep + 'test-l.fit' )
#fs.ToFitFile( 'temp.fit' )
fw = FisherFeatureWeights.NewFromFeatureSpace(fs).Threshold()
fs.FeatureReduce(fw, inplace=True)
# #fw.Print()
# #fs.Print(verbose=True)
pychrm_res = FeatureSpaceClassification.NewWND5(fs, fs, fw)
pychrm_res.Print()
#
# import cProfile as pr
# #import profile as pr
# import tempfile
# import pstats
# prof = tempfile.NamedTemporaryFile()
# cmd = 'no_tile_pychrm_result = DiscreteBatchClassificationResult.New( reduced_fs, reduced_fs, fw )'
# pr.runctx( cmd, globals(), locals(), prof.name)
# p = pstats.Stats(prof.name)
# p.sort_stats('time').print_stats(20)
# prof.close()
self.maxDiff = None
html_path = pychrm_test_dir + sep + 'lymphoma_iicbu2008_subset_eosin_t5x6_v3.2feats_REFERENCE_RESULTS_900_samples_TRAINING_ERROR.html'
wres = FeatureSpaceClassificationExperiment.NewFromHTMLReport(
html_path)
wres.Print()
wc_batch_result = wres.individual_results[
0] # only 1 split in fit-on-fit
# This takes WAY too long:
#self.assertSequenceEqual( wc_batch_result.individual_results, pychrm_res.individual_results )
wc_result = np.empty((3 * len(wc_batch_result.individual_results)))
for i, single_result in enumerate(
wc_batch_result.individual_results):
wc_result[i * 3:(i + 1) *
3] = single_result.marginal_probabilities
pc_result = np.empty((3 * len(pychrm_res.individual_results)))
for i, single_result in enumerate(pychrm_res.individual_results):
# HTML report only has 3 decimal places
pc_result[ i*3 : (i+1)*3 ] = \
[ float( "{0:0.3f}".format( val ) ) for val in single_result.marginal_probabilities ]
from numpy.testing import assert_allclose
assert_allclose(actual=pc_result, desired=wc_result, atol=0.003)
#wc_batch_result.Print()
#pres.Print()
# ==========================================================
# Now do the same with tiling, reusing fs from before:
num_samples_per_group = 30
n_groups = fs.num_samples / num_samples_per_group
new_sg_ids = [
i for i in xrange(n_groups)
for j in xrange(num_samples_per_group)
]
fs.Update( tile_num_rows=5, tile_num_cols=6, num_samples_per_group=30,\
_contiguous_sample_group_ids=new_sg_ids )._RebuildViews()
with_tile_pychrm_result = FeatureSpaceClassification.NewWND5(
fs, fs, fw)
html_path = pychrm_test_dir + sep + 'lymphoma_iicbu2008_subset_eosin_t5x6_v3.2feats_REFERENCE_RESULTS_30_samples_tiled_TRAINING_ERROR.html'
with_tile_wndchrm_result = \
FeatureSpaceClassificationExperiment.NewFromHTMLReport( html_path ).individual_results[0]
#self.assertSequenceEqual( with_tile_pychrm_result.averaged_results, with_tile_wndchrm_result.individual_results )
wc_result = np.empty(
(3 * len(with_tile_wndchrm_result.individual_results)))
for i, single_result in enumerate(
with_tile_wndchrm_result.individual_results):
wc_result[i * 3:(i + 1) *
3] = single_result.marginal_probabilities
pc_result = np.empty(
(3 * len(with_tile_pychrm_result.averaged_results)))
for i, single_result in enumerate(
with_tile_pychrm_result.averaged_results):
# HTML report only has 3 decimal places
pc_result[ i*3 : (i+1)*3 ] = \
[ float( "{0:0.3f}".format( val ) ) for val in single_result.marginal_probabilities ]
assert_allclose(actual=pc_result, desired=wc_result, atol=0.003)
finally:
rmtree(tempdir)
if feature_usage_fraction:
if feature_usage_fraction < 0 or feature_usage_fraction > 1.0:
raise Exception('Feature usage fraction must be on interval [0,1]')
num_features = int( feature_usage_fraction * train_set.num_features )
if num_features:
print "Using top {0} Fisher-ranked features.".format( num_features )
else:
print "Using top 15% Fisher-ranked features."
experiment = FeatureSpaceClassificationExperiment( training_set=train_set )
train_set.Normalize( inplace=True )
weights = FisherFeatureWeights.NewFromFeatureSpace( train_set ).Threshold( num_features )
train_set.FeatureReduce( weights, inplace=True )
if train_set != test_set:
test_set.FeatureReduce( weights, inplace=True ).Normalize( train_set )
for i in range( num_splits ):
split = FeatureSpaceClassification.NewWND5( train_set, test_set, weights, batch_number=i )
experiment.individual_results.append( split )
if outpath:
experiment.Print( output_filepath=outpath, mode='w' )
#experiment.PerSampleStatistics( output_filepath=outpath, mode= 'a' )
else:
experiment.Print()
#experiment.PerSampleStatistics()