def test_Normalize(self):
"""FIXME: THIS TEST BELONGS IN TEST_FEATURESET.PY"""
from numpy.testing import assert_allclose
result_fs = FeatureSet_Discrete.NewFromFitFile(self.test_fit_path)
result_fs.Normalize()
target_fs = FeatureSet_Discrete.NewFromFitFile(
self.test_normalized_fit_path)
assert_allclose(result_fs.data_matrix,
target_fs.data_matrix,
rtol=self.epsilon)
def test_IfNotInterpolatable(self):
"""You can't graph predicted values if the classes aren't interpolatable."""
testfilename = 'ShouldntBeGraphable.png'
fitfilepath = wndchrm_test_dir + sep + 'test-l.fit'
fs = FeatureSet_Discrete.NewFromFitFile(fitfilepath)
train_set, test_set = fs.Split(randomize=False, quiet=True)
train_set.Normalize()
fw = FisherFeatureWeights.NewFromFeatureSet(train_set).Threshold()
reduced_train_set = train_set.FeatureReduce(fw.names)
reduced_test_set = test_set.FeatureReduce(fw.names)
test_set.Normalize(train_set, quiet=True)
batch_result = DiscreteBatchClassificationResult.New(reduced_train_set,
reduced_test_set,
fw,
quiet=True)
graph = PredictedValuesGraph(batch_result)
tempfile = self.tempdir + sep + testfilename
with self.assertRaises(ValueError):
graph.RankOrderedPredictedValuesGraph()
graph.SaveToFile(tempfile)
def crossValidate(ftb, project, featureThreshold, imagesOnly, numSplits):
message = ''
fullSet = FeatureSet_Discrete()
fullSet.source_path = project.getName()
classId = 0
for ds in project.listChildren():
message += 'Processing dataset id:%d\n' % ds.getId()
message += addToFeatureSet(ftb, ds, fullSet, classId, imagesOnly)
classId += 1
tmp = fullSet.ContiguousDataMatrix()
experiment = DiscreteClassificationExperimentResult(training_set=fullSet)
for i in range(numSplits):
trainSet, testSet = fullSet.Split()
trainSet.Normalize()
testSet.Normalize(trainSet)
weights = FisherFeatureWeights.NewFromFeatureSet(trainSet)
nFeatures = ceil(len(weights.names) * featureThreshold)
message += 'Selecting top %d features\n' % nFeatures
weights = weights.Threshold(nFeatures)
trainSet = reduceFeatures(trainSet, weights)
reducedTestSet = reduceFeatures(testSet, weights)
reducedTrainSet = reduceFeatures(trainSet, weights)
batchResult = DiscreteBatchClassificationResult.New(
reducedTrainSet, reducedTestSet, weights, batch_number=i)
experiment.individual_results.append(batchResult)
out = StringIO()
experiment.Print(output_stream=out)
experiment.PerSampleStatistics(output_stream=out)
pid = project.getId()
WndcharmStorage.addTextFileAnnotationTo(
ftb.conn, out.getvalue(), 'Project', pid,
'Wndcharm_Cross_Validation_Results.txt',
'Wndcharm Cross Validation Results for Project:%d' % pid)
message += 'Attached cross-validation results\n'
#return experiment
return message
def test_predictDiscrete(self):
s1 = self.createSignature(1, -1)
s2 = self.createSignature(2, -2)
s3 = self.createSignature(3, -3)
s4 = self.createSignature(4, -4)
trainFts = FeatureSet_Discrete()
trainFts.AddSignature(s1, 0)
trainFts.AddSignature(s2, 0)
trainFts.AddSignature(s3, 1)
trainFts.AddSignature(s4, 1)
tmp = trainFts.ContiguousDataMatrix()
# Values are chosen so that different weights give different predictions
s5 = self.createSignature(0, -5)
s6 = self.createSignature(5, 0)
testFts = FeatureSet_Discrete()
testFts.AddSignature(s5, 0)
testFts.AddSignature(s6, 0)
tmp = testFts.ContiguousDataMatrix()
weights = FisherFeatureWeights()
weights.names = ['ft [0]', 'ft [1]']
weights.values = [2.0, 1.0]
pred = DiscreteBatchClassificationResult.New(trainFts, testFts,
weights)
self.assertEqual(len(pred.individual_results), 2)
r1, r2 = pred.individual_results
np.testing.assert_almost_equal(r1.marginal_probabilities,
[0.975, 0.025],
decimal=3)
np.testing.assert_almost_equal(r2.marginal_probabilities,
[0.025, 0.975],
decimal=3)
weights = FisherFeatureWeights()
weights.names = ['ft [0]', 'ft [1]']
weights.values = [1.0, 2.0]
pred = DiscreteBatchClassificationResult.New(trainFts, testFts,
weights)
self.assertEqual(len(pred.individual_results), 2)
r1, r2 = pred.individual_results
np.testing.assert_almost_equal(r1.marginal_probabilities,
[0.025, 0.975],
decimal=3)
np.testing.assert_almost_equal(r2.marginal_probabilities,
[0.975, 0.025],
decimal=3)
def test_PerSampleStatisticsWITHOUTPredictedValue(self):
"""DISCRETE ShuffleSplit/PerSampleStatistics w/ mini binucleate test set (no predicted value)"""
fs = FeatureSet_Discrete.NewFromFitFile('../wndchrm_tests/test-l.fit')
exp = DiscreteClassificationExperimentResult.NewShuffleSplit(
fs, quiet=True)
exp.PerSampleStatistics()
self.assertTrue(True)
def test_NewFromFeatureSet(self):
"""Fisher score calculation"""
feature_set = FeatureSet_Discrete.NewFromFitFile(self.test_fit_path)
feature_set.Normalize()
result_weights = FisherFeatureWeights.NewFromFeatureSet(feature_set)
# test weights generated from test-l.fit:
# wndchrm classify -l -f1.0 -vtest_fit-l.weights test-l.fit test-l.fit
target_weights = FisherFeatureWeights.NewFromFile(
self.test_feat_weight_path)
for target_val, res_val in zip(target_weights.values,
result_weights.values):
self.assertAlmostEqual(target_val, res_val, delta=self.epsilon)
def test_fisherFeatureWeights(self):
sig1, sig2, sig3, sig4 = self.createSignatures()
fts = FeatureSet_Discrete()
fts.AddSignature(sig1, 0)
fts.AddSignature(sig2, 0)
fts.AddSignature(sig3, 1)
fts.AddSignature(sig4, 1)
tmp = fts.ContiguousDataMatrix()
# TODO: weight[1]==0, presumably because the intra-class variance=0,
# even though feature[1] is a perfect discriminator?
fts.Normalize()
wts = FisherFeatureWeights.NewFromFeatureSet(fts)
np.testing.assert_almost_equal(wts.values, [4.0, 0.0])
self.assertEqual(wts.names, ['ft [0]', 'ft [1]'])
def test_createFeatureSet(self):
sig1, sig2, sig3, sig4 = self.createSignatures()
fts = FeatureSet_Discrete()
# Add classes out of order
fts.AddSignature(sig3, 1)
self.assertEqual(fts.num_classes, 2)
self.assertEqual(fts.num_features, 2)
self.assertEqual(fts.num_images, 1)
self.assertEqual(len(fts.data_list), 2)
self.assertIsNone(fts.data_list[0])
#self.assertSequenceEqual
np.testing.assert_almost_equal(fts.data_list[1], sig3.values)
self.assertEqual(fts.classsizes_list, [0, 1])
self.assertEqual(fts.classnames_list, ['UNKNOWN1', 'UNKNOWN2'])
fts.AddSignature(sig1, 0)
self.assertEqual(fts.num_classes, 2)
self.assertEqual(fts.num_features, 2)
self.assertEqual(fts.num_images, 2)
self.assertEqual(len(fts.data_list), 2)
np.testing.assert_almost_equal(fts.data_list[0], sig1.values)
np.testing.assert_almost_equal(fts.data_list[1], sig3.values)
self.assertEqual(fts.classsizes_list, [1, 1])
self.assertEqual(fts.classnames_list, ['UNKNOWN1', 'UNKNOWN2'])
# fts.ContiguousDataMatrix() fails unless there are at least two images
# per class, is this really necessary?
#tmp = fts.ContiguousDataMatrix()
fts.AddSignature(sig2, 0)
fts.AddSignature(sig4, 1)
self.assertEqual(fts.classsizes_list, [2, 2])
self.assertEqual(fts.num_images, 4)
tmp = fts.ContiguousDataMatrix()
self.assertEqual(fts.data_matrix.shape, (4, 2))
np.testing.assert_almost_equal(fts.data_matrix[0], sig1.values)
np.testing.assert_almost_equal(fts.data_matrix[1], sig2.values)
np.testing.assert_almost_equal(fts.data_matrix[2], sig3.values)
np.testing.assert_almost_equal(fts.data_matrix[3], sig4.values)
class TestWND5Classification(unittest.TestCase):
"""WND5 Classification"""
epsilon = 0.00001
# Define paths to original files
test_sig_path = join(test_dir, 't1_s01_c05_ij-l_precalculated.sig')
test_fit_path = join(test_dir, 'test-l.fit')
test_feat_wght_path = join(test_dir, 'test_fit-l.weights')
test_tif_path = join(test_dir, 't1_s01_c05_ij.tif')
# Here are the correct values that Python API needs to return:
# wndchrm classify -l -f1.0 test-l.fit t1_s01_c05_ij.tif
# t1_s01_c05_ij.tif 1.6e-27 0.083 0.917 * 4cell 3.835
# wndchrm classify -l -f0.14765 test-l.fit t1_s01_c05_ij.tif
# t1_s01_c05_ij.tif 3.23e-27 0.076 0.924 * 4cell 3.848
# wndchrm classify -l -f0.0685 test-l.fit t1_s01_c05_ij.tif
# t1_s01_c05_ij.tif 7.05e-27 0.069 0.931 * 4cell 3.862
correct_marg_probs = {}
correct_marg_probs[2919] = [0.083, 0.917]
correct_marg_probs[431] = [0.076, 0.924]
#correct_marg_probs[200] = [0.044, 0.956]
# slight difference in marg probs due to my use of round() below
correct_marg_probs[200] = [0.069, 0.931]
# Load the original files once and only once for all this class's tests
feature_set = FeatureSet_Discrete.NewFromFitFile(test_fit_path)
feature_set.Normalize()
test_sample = Signatures.NewFromSigFile(test_sig_path, test_tif_path)
test_sample.Normalize(feature_set)
all_weights = FisherFeatureWeights.NewFromFile(test_feat_wght_path)
# --------------------------------------------------------------------------
def Check(self, num_feats=None):
weights = self.all_weights.Threshold(num_feats)
feat_set = self.feature_set.FeatureReduce(weights.names)
sample = self.test_sample.FeatureReduce(weights.names)
result = DiscreteImageClassificationResult.NewWND5(
feat_set, weights, sample)
result_marg_probs = [ round( val, 3 ) \
for val in result.marginal_probabilities ]
self.assertSequenceEqual(self.correct_marg_probs[num_feats],
result_marg_probs)
# --------------------------------------------------------------------------
def test_WND5_all_features(self):
"""WND5 classification with entire large feature set (2919 features)"""
self.Check(2919)
# --------------------------------------------------------------------------
def test_WND5_15percent_threshold(self):
"""WND5 classification with large feature set 15% threshold (431 features)"""
self.Check(431)
# --------------------------------------------------------------------------
def test_WND5_200_feat_threshold(self):
"""WND5 classification with large feature set & 200 feature threshold"""
self.Check(200)
def test_incompatibleFeatureVersion(self):
s = self.createSignature(1, 10)
fts = FeatureSet_Discrete()
fts.feature_vector_version = "0.0"
self.assertRaises(ValueError, fts.AddSignature, s, 1)
def test_incompatibleFeatureVersion(self):
s = self.createSignature(1, 10)
fts = FeatureSet_Discrete()
fts.feature_vector_version = '0.0'
self.assertRaises(ValueError, fts.AddSignature, s, 1)
