def test_SplitOptions( self ):
from wndcharm.ArtificialFeatureSpace import CreateArtificialFeatureSpace_Discrete
fs_discrete = CreateArtificialFeatureSpace_Discrete( n_samples=1000, n_classes=10,
num_features_per_signal_type=30, noise_gradient=5, initial_noise_sigma=10,
n_samples_per_group=1, interpolatable=True, random_state=42)
# default
train_set, test_set = fs_discrete.Split( random_state=42, quiet=True )
self.assertEqual( train_set.shape, (750, 600) )
self.assertEqual( test_set.shape, (250, 600) )
# Supposed to only return single FeatureSpace instead of 2-tuple of FeatureSpace
# when setting test_size = 0
i = 50
retval = fs_discrete.Split( train_size=i, test_size=0, random_state=42, quiet=True )
self.assertEqual( type(retval), FeatureSpace )
self.assertEqual( retval.num_samples, i * fs_discrete.num_classes )
# dummyproofing
self.assertRaises( ValueError, fs_discrete.Split, train_size='trash' )
self.assertRaises( ValueError, fs_discrete.Split, train_size=1.1 )
self.assertRaises( ValueError, fs_discrete.Split, test_size='trash' )
self.assertRaises( ValueError, fs_discrete.Split, test_size=1.1 )
# What if the feature set number of groups within a class are less than called for
# when specifying by integer?
self.assertRaises( ValueError, test_set.Split, test_size=25 )
# What happens when input fs has unbalanced classes, some of which have enough
# to satisfy train_size/test_size params, and some don't
remove_these = range(250,300) + range(700,750)
fs_class_2_and_7_smaller = \
fs_discrete.SampleReduce( leave_out_sample_group_ids=remove_these )
self.assertRaises( ValueError, fs_class_2_and_7_smaller.Split, train_size=80,
test_size=20 )
# Test balanced_classes:
train_fs, test_fs = fs_class_2_and_7_smaller.Split()
# Training set number rounds down (apparently).
from math import floor
expected_num_samps_per_train_class = int( floor(50*0.75) )
expected_num_samps_per_test_class = 50 - expected_num_samps_per_train_class
err_msg = "Balanced classes {} set split error, class {}, expected {}, got {}"
for i, (n_train, n_test) in enumerate( zip( train_fs.class_sizes, test_fs.class_sizes )):
self.assertEqual( n_train, expected_num_samps_per_train_class, msg=\
err_msg.format( "TRAIN", i, expected_num_samps_per_train_class, n_train ) )
self.assertEqual( n_test, expected_num_samps_per_test_class, msg=\
err_msg.format( "TEST", i, expected_num_samps_per_test_class, n_test ) )
def test_SplitOptions(self):
from wndcharm.ArtificialFeatureSpace import CreateArtificialFeatureSpace_Discrete
fs_discrete = CreateArtificialFeatureSpace_Discrete(
n_samples=1000,
n_classes=10,
num_features_per_signal_type=30,
noise_gradient=5,
initial_noise_sigma=10,
n_samples_per_group=1,
interpolatable=True,
random_state=42)
# default
train_set, test_set = fs_discrete.Split(random_state=42, quiet=True)
self.assertEqual(train_set.shape, (750, 600))
self.assertEqual(test_set.shape, (250, 600))
# Supposed to only return single FeatureSpace instead of 2-tuple of FeatureSpace
# when setting test_size = 0
i = 50
retval = fs_discrete.Split(train_size=i,
test_size=0,
random_state=42,
quiet=True)
self.assertEqual(type(retval), FeatureSpace)
self.assertEqual(retval.num_samples, i * fs_discrete.num_classes)
# dummyproofing
self.assertRaises(ValueError, fs_discrete.Split, train_size='trash')
self.assertRaises(ValueError, fs_discrete.Split, train_size=1.1)
self.assertRaises(ValueError, fs_discrete.Split, test_size='trash')
self.assertRaises(ValueError, fs_discrete.Split, test_size=1.1)
# What if the feature set number of groups within a class are less than called for
# when specifying by integer?
self.assertRaises(ValueError, test_set.Split, test_size=25)
# What happens when input fs has unbalanced classes, some of which have enough
# to satisfy train_size/test_size params, and some don't
remove_these = range(250, 300) + range(700, 750)
fs_class_2_and_7_smaller = \
fs_discrete.SampleReduce( leave_out_sample_group_ids=remove_these )
self.assertRaises(ValueError,
fs_class_2_and_7_smaller.Split,
train_size=80,
test_size=20)
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