def _load_taxon_hierarchy(self):
"""Load the taxon hierarchy.
Must be separate from the constructor because
:meth:`set_photo_count_min` influences the taxon hierarchy.
"""
session, metadata = db.get_session_or_error()
if not self.taxon_hr:
self.taxon_hr = db.get_taxon_hierarchy(session, metadata)
def test_get_taxon_hierarchy(self):
"""Test the get_taxon_hierarchy() method."""
with db.session_scope(META_FILE) as (session, metadata):
expected = {
u'Paphiopedilum': {
u'Brachypetalum': [u'godefroyae', u'wenshanense']
},
u'Selenipedium': {
None: [u'palmifolium']
},
u'Mexipedium': {
None: [u'xerophyticum']
},
u'Cypripedium': {
u'Trigonopedia': [u'fargesii', u'sichuanense'],
u'Obtusipetala': [u'flavum'],
u'Arietinum': [u'plectrochilum']
},
u'Phragmipedium': {
u'Micropetalum': [u'besseae']
}
}
hier = db.get_taxon_hierarchy(session, metadata)
self.assertEqual(str(hier), str(expected))
# Setting `photo_count_min` to 4 should only return the species with
# at least 4 photos.
conf.photo_count_min = 4
expected = {
u'Paphiopedilum': {
u'Brachypetalum': [u'godefroyae']
},
u'Cypripedium': {
u'Obtusipetala': [u'flavum']
},
u'Phragmipedium': {
u'Micropetalum': [u'besseae']
}
}
hier = db.get_taxon_hierarchy(session, metadata)
self.assertEqual(str(hier), str(expected))
def _load_taxon_hierarchy(self):
"""Load the taxon hierarchy.
Must be separate from the constructor because
:meth:`set_photo_count_min` influences the taxon hierarchy.
"""
session, metadata = db.get_session_or_error()
if not self.taxon_hr:
self.taxon_hr = db.get_taxon_hierarchy(session, metadata)
def get_taxon_hierarchy(self):
"""Return the taxon hierarchy.
First tries to get the taxon hierarchy from the metadata database. If
that fails, it will try to get it from the configuration file.
"""
try:
session, metadata = db.get_session_or_error()
hr = db.get_taxon_hierarchy(session, metadata)
except DatabaseSessionError:
hr = self.config.classification.taxa.as_dict()
return hr
def get_taxon_hierarchy(self):
"""Return the taxon hierarchy.
First tries to get the taxon hierarchy from the metadata database. If
that fails, it will try to get it from the configuration file.
"""
try:
session, metadata = db.get_session_or_error()
hr = db.get_taxon_hierarchy(session, metadata)
except DatabaseSessionError:
hr = self.config.classification.taxa.as_dict()
return hr
def test_with_hierarchy(self, test_data_dir, ann_dir, max_error=0.001):
"""Test each ANN in a classification hierarchy and export results.
Returns a 2-tuple ``(correct,total)``.
"""
session, metadata = db.get_session_or_error()
logging.info("Testing the neural networks hierarchy...")
self.classifications = {}
self.classifications_expected = {}
# Get the taxonomic hierarchy from the database.
self.taxon_hr = db.get_taxon_hierarchy(session, metadata)
# Get the classification hierarchy from the configurations.
try:
self.class_hr = self.config.classification.hierarchy
except:
raise ConfigurationError("classification hierarchy not set")
# Get the name of each level in the classification hierarchy.
levels = [l.name for l in self.class_hr]
# Get the prefix for the classification columns.
try:
dependent_prefix = self.config.data.dependent_prefix
except:
dependent_prefix = OUTPUT_PREFIX
# Get the expected and recognized classification for each sample in
# the test data.
for filter_ in classification_hierarchy_filters(levels, self.taxon_hr):
logging.info("Classifying on %s" % readable_filter(filter_))
level_name = filter_.get('class')
level_n = levels.index(level_name)
level = self.class_hr[level_n]
test_file = os.path.join(test_data_dir, level.test_file)
ann_file = os.path.join(ann_dir, level.ann_file)
# Set the maximum error for classification.
try:
max_error = level.max_error
except:
pass
# Replace any placeholders in the paths.
where = filter_.get('where', {})
for key, val in where.items():
val = val if val is not None else '_'
test_file = test_file.replace("__%s__" % key, val)
ann_file = ann_file.replace("__%s__" % key, val)
# Get the class names for this filter.
classes = db.get_classes_from_filter(session, metadata, filter_)
assert len(classes) > 0, \
"No classes found for filter `%s`" % filter_
# Get the codeword for each class.
codewords = get_codewords(classes)
# Load the ANN.
if len(classes) > 1:
ann = libfann.neural_net()
ann.create_from_file(str(ann_file))
# Load the test data.
test_data = TrainData()
test_data.read_from_file(test_file, dependent_prefix)
# Test each sample in the test data.
for label, input_, output in test_data:
assert len(codewords) == len(output), \
"Codeword size mismatch. Codeword has {0} bits, but the " \
"training data has {1} output bits.".\
format(len(codewords), len(output))
# Obtain the photo ID from the label.
if not label:
raise ValueError("Test sample is missing a label with " \
"photo ID")
try:
photo_id = self.re_photo_id.search(label).group(1)
photo_id = int(photo_id)
except:
raise RuntimeError("Failed to obtain the photo ID from " \
"the sample label")
# Skip classification if there is only one class for this
# filter.
if not len(classes) > 1:
logging.debug("Not enough classes for filter. Skipping " \
"testing of %s" % ann_file)
self.classifications[photo_id][level_name] = ['']
self.classifications_expected[photo_id][level_name] = ['']
continue
# Set the expected class.
class_expected = get_classification(codewords, output,
max_error)
class_expected = [class_ for mse,class_ in class_expected]
assert len(class_expected) == 1, \
"Class codewords must have one positive bit, found {0}".\
format(len(class_expected))
# Get the recognized class.
codeword = ann.run(input_)
class_ann = get_classification(codewords, codeword,
max_error)
class_ann = [class_ for mse,class_ in class_ann]
# Save the classification at each level.
if level_n == 0:
self.classifications[photo_id] = {}
self.classifications_expected[photo_id] = {}
self.classifications[photo_id][level_name] = class_ann
self.classifications_expected[photo_id][level_name] = class_expected
ann.destroy()
return self.get_correct_count()
def test_with_hierarchy(self, test_data_dir, ann_dir, max_error=0.001):
"""Test each ANN in a classification hierarchy and export results.
Returns a 2-tuple ``(correct,total)``.
"""
session, metadata = db.get_session_or_error()
logging.info("Testing the neural networks hierarchy...")
self.classifications = {}
self.classifications_expected = {}
# Get the taxonomic hierarchy from the database.
self.taxon_hr = db.get_taxon_hierarchy(session, metadata)
# Get the classification hierarchy from the configurations.
try:
self.class_hr = self.config.classification.hierarchy
except:
raise ConfigurationError("classification hierarchy not set")
# Get the name of each level in the classification hierarchy.
levels = [l.name for l in self.class_hr]
# Get the prefix for the classification columns.
try:
dependent_prefix = self.config.data.dependent_prefix
except:
dependent_prefix = OUTPUT_PREFIX
# Get the expected and recognized classification for each sample in
# the test data.
for filter_ in classification_hierarchy_filters(levels, self.taxon_hr):
logging.info("Classifying on %s" % readable_filter(filter_))
level_name = filter_.get('class')
level_n = levels.index(level_name)
level = self.class_hr[level_n]
test_file = os.path.join(test_data_dir, level.test_file)
ann_file = os.path.join(ann_dir, level.ann_file)
# Set the maximum error for classification.
try:
max_error = level.max_error
except:
pass
# Replace any placeholders in the paths.
where = filter_.get('where', {})
for key, val in where.items():
val = val if val is not None else '_'
test_file = test_file.replace("__%s__" % key, val)
ann_file = ann_file.replace("__%s__" % key, val)
# Get the class names for this filter.
classes = db.get_classes_from_filter(session, metadata, filter_)
assert len(classes) > 0, \
"No classes found for filter `%s`" % filter_
# Get the codeword for each class.
codewords = get_codewords(classes)
# Load the ANN.
ann = libfann.neural_net()
if len(classes) > 1:
ann.create_from_file(str(ann_file))
# Load the test data.
test_data = TrainData()
test_data.read_from_file(test_file, dependent_prefix)
# Test each sample in the test data.
for label, input_, output in test_data:
assert len(codewords) == len(output), \
"Codeword size mismatch. Codeword has {0} bits, but the " \
"training data has {1} output bits.".\
format(len(codewords), len(output))
# Obtain the photo ID from the label.
if not label:
raise ValueError("Test sample is missing a label with " \
"photo ID")
try:
photo_id = self.re_photo_id.search(label).group(1)
photo_id = int(photo_id)
except:
raise RuntimeError("Failed to obtain the photo ID from " \
"the sample label")
# Save the classification at each level.
if level_n == 0:
self.classifications[photo_id] = {}
self.classifications_expected[photo_id] = {}
# Skip classification if there is only one class for this
# filter.
if not len(classes) > 1:
logging.debug("Not enough classes for filter. Skipping " \
"testing of %s" % ann_file)
self.classifications[photo_id][level_name] = ['']
self.classifications_expected[photo_id][level_name] = ['']
continue
# Set the expected class.
class_expected = get_classification(codewords, output,
max_error)
class_expected = [class_ for mse,class_ in class_expected]
assert len(class_expected) == 1, \
"Class codewords must have one positive bit, found {0}".\
format(len(class_expected))
# Get the recognized class.
codeword = ann.run(input_)
class_ann = get_classification(codewords, codeword,
max_error)
class_ann = [class_ for mse,class_ in class_ann]
# Save the classification at each level.
self.classifications[photo_id][level_name] = class_ann
self.classifications_expected[photo_id][level_name] = class_expected
ann.destroy()
return self.get_correct_count()
