def _make_ftor_itq(self, bits=32):
itq_ftor = ItqFunctor(bit_length=bits, random_seed=self.RANDOM_SEED)
def itq_fit(D):
itq_ftor.fit(D)
return itq_ftor, itq_fit
def nearestNeighborIndex(item, user, descriptorSet):
"""
Get the nearest neighbor index from a given item and descriptor set.
:param item: Item to find the nn index from, usually the item that the
user is performing the nearest neighbors search on.
:param user: The owner of the .smqtk folder.
:param descriptorSet: The relevant descriptor set.
"""
folder = ModelImporter.model('folder')
_GirderDataElement = functools.partial(GirderDataElement,
api_root=getApiUrl(),
token=getCurrentToken()['_id'])
smqtkFolder = folder.createFolder(folder.load(item['folderId'], user=user), '.smqtk',
reuseExisting=True)
try:
meanVecFileId = localSmqtkFileIdFromName(smqtkFolder, 'mean_vec.npy')
rotationFileId = localSmqtkFileIdFromName(smqtkFolder, 'rotation.npy')
hash2uuidsFileId = localSmqtkFileIdFromName(smqtkFolder, 'hash2uuids.pickle')
except Exception:
logger.warn('SMQTK files didn\'t exist for performing NN on %s' % item['_id'])
return None
# TODO Should these be Girder data elements? Unnecessary HTTP requests.
functor = ItqFunctor(mean_vec_cache=_GirderDataElement(meanVecFileId),
rotation_cache=_GirderDataElement(rotationFileId))
hash2uuidsKV = MemoryKeyValueStore(_GirderDataElement(hash2uuidsFileId))
return LSHNearestNeighborIndex(functor, descriptorSet,
hash2uuidsKV, read_only=True)
def test_save_model_with_writable_caches(self):
# If one or both cache elements are read-only, no saving.
expected_mean_vec = numpy.array([1, 2, 3])
expected_rotation = numpy.eye(3)
expected_mean_vec_bytes = six.BytesIO()
# noinspection PyTypeChecker
numpy.save(expected_mean_vec_bytes, expected_mean_vec)
expected_mean_vec_bytes = expected_mean_vec_bytes.getvalue()
expected_rotation_bytes = six.BytesIO()
# noinspection PyTypeChecker
numpy.save(expected_rotation_bytes, expected_rotation)
expected_rotation_bytes = expected_rotation_bytes.getvalue()
itq = ItqFunctor()
itq.mean_vec = expected_mean_vec
itq.rotation = expected_rotation
itq.mean_vec_cache_elem = DataMemoryElement(readonly=False)
itq.rotation_cache_elem = DataMemoryElement(readonly=False)
itq.save_model()
self.assertEqual(itq.mean_vec_cache_elem.get_bytes(),
expected_mean_vec_bytes)
self.assertEqual(itq.rotation_cache_elem.get_bytes(),
expected_rotation_bytes)
def test_fit_short_descriptors_for_bit_length(self):
# Should error when input descriptors have fewer dimensions than set bit
# length for output hash codes (limitation of PCA method currently
# used).
fit_descriptors = []
for i in range(3):
d = DescriptorMemoryElement(six.b('test'), i)
d.set_vector([-1 + i, -1 + i])
fit_descriptors.append(d)
itq = ItqFunctor(bit_length=8)
self.assertRaisesRegex(
ValueError,
"Input descriptors have fewer features than requested bit encoding",
itq.fit, fit_descriptors)
self.assertIsNone(itq.mean_vec)
self.assertIsNone(itq.rotation)
# Should behave the same when input is an iterable
self.assertRaisesRegex(
ValueError,
"Input descriptors have fewer features than requested bit encoding",
itq.fit, iter(fit_descriptors))
self.assertIsNone(itq.mean_vec)
self.assertIsNone(itq.rotation)
def test_save_model_with_read_only_cache(self):
# If one or both cache elements are read-only, no saving.
expected_mean_vec = numpy.array([1, 2, 3])
expected_rotation = numpy.eye(3)
itq = ItqFunctor()
itq.mean_vec = expected_mean_vec
itq.rotation = expected_rotation
# read-only mean-vec cache
itq.mean_vec_cache_elem = DataMemoryElement(readonly=True)
itq.rotation_cache_elem = DataMemoryElement(readonly=False)
itq.save_model()
self.assertEqual(itq.mean_vec_cache_elem.get_bytes(), six.b(''))
self.assertEqual(itq.rotation_cache_elem.get_bytes(), six.b(''))
# read-only rotation cache
itq.mean_vec_cache_elem = DataMemoryElement(readonly=False)
itq.rotation_cache_elem = DataMemoryElement(readonly=True)
itq.save_model()
self.assertEqual(itq.mean_vec_cache_elem.get_bytes(), six.b(''))
self.assertEqual(itq.rotation_cache_elem.get_bytes(), six.b(''))
# Both read-only
itq.mean_vec_cache_elem = DataMemoryElement(readonly=True)
itq.rotation_cache_elem = DataMemoryElement(readonly=True)
itq.save_model()
self.assertEqual(itq.mean_vec_cache_elem.get_bytes(), six.b(''))
self.assertEqual(itq.rotation_cache_elem.get_bytes(), six.b(''))
def test_get_hash(self):
fit_descriptors = []
for i in range(5):
d = DescriptorMemoryElement(six.b('test'), i)
d.set_vector([-2. + i, -2. + i])
fit_descriptors.append(d)
# The following "rotation" matrix should cause any 2-feature descriptor
# to the right of the line ``y = -x`` to be True, and to the left as
# False. If on the line, should be True.
itq = ItqFunctor(bit_length=1, random_seed=0)
itq.mean_vec = numpy.array([0., 0.])
itq.rotation = numpy.array([[1. / sqrt(2)], [1. / sqrt(2)]])
numpy.testing.assert_array_equal(itq.get_hash(numpy.array([1, 1])),
[True])
numpy.testing.assert_array_equal(itq.get_hash(numpy.array([-1, -1])),
[False])
numpy.testing.assert_array_equal(itq.get_hash(numpy.array([-1, 1])),
[True])
numpy.testing.assert_array_equal(
itq.get_hash(numpy.array([-1.001, 1])), [False])
numpy.testing.assert_array_equal(
itq.get_hash(numpy.array([-1, 1.001])), [True])
numpy.testing.assert_array_equal(itq.get_hash(numpy.array([1, -1])),
[True])
numpy.testing.assert_array_equal(
itq.get_hash(numpy.array([1, -1.001])), [False])
numpy.testing.assert_array_equal(
itq.get_hash(numpy.array([1.001, -1])), [True])
def test_fit_with_cache(self):
fit_descriptors = []
for i in range(5):
d = DescriptorMemoryElement(six.b('test'), i)
d.set_vector([-2. + i, -2. + i])
fit_descriptors.append(d)
itq = ItqFunctor(DataMemoryElement(),
DataMemoryElement(),
bit_length=1,
random_seed=0)
itq.fit(fit_descriptors)
# TODO: Explanation as to why this is the expected result.
numpy.testing.assert_array_almost_equal(itq.mean_vec, [0, 0])
numpy.testing.assert_array_almost_equal(itq.rotation,
[[1 / sqrt(2)], [1 / sqrt(2)]])
self.assertIsNotNone(itq.mean_vec_cache_elem)
numpy.testing.assert_array_almost_equal(
numpy.load(six.BytesIO(itq.mean_vec_cache_elem.get_bytes())),
[0, 0])
self.assertIsNotNone(itq.rotation_cache_elem)
numpy.testing.assert_array_almost_equal(
numpy.load(six.BytesIO(itq.rotation_cache_elem.get_bytes())),
[[1 / sqrt(2)], [1 / sqrt(2)]])
def _nearestNeighborIndex(sid, descriptor_set):
"""
Retrieve the Nearest neighbor index for a given session.
:param sid: ID of the session
:param descriptor_set: The descriptor set corresponding to the session id,
see _descriptorSetFromSessionId.
:returns: Nearest neighbor index or None if no session exists
:rtype: LSHNearestNeighborIndex|None
"""
session = ModelImporter.model('item').findOne({'_id': ObjectId(sid)})
if not session:
return None
else:
smqtkFolder = {'_id': ObjectId(session['meta']['smqtk_folder_id'])}
functor = ItqFunctor(
smqtkDataElementFromGirderFileId(
localSmqtkFileIdFromName(smqtkFolder, 'mean_vec.npy')),
smqtkDataElementFromGirderFileId(
localSmqtkFileIdFromName(smqtkFolder, 'rotation.npy')))
hash2uuidsKV = MemoryKeyValueStore(
smqtkDataElementFromGirderFileId(
localSmqtkFileIdFromName(smqtkFolder,
'hash2uuids.pickle')))
return LSHNearestNeighborIndex(functor,
descriptor_set,
hash2uuidsKV,
read_only=True)
def test_norm_vector_n2(self):
itq = ItqFunctor(normalize=2)
v = numpy.array([1, 0])
numpy.testing.assert_array_almost_equal(itq._norm_vector(v), [1, 0])
v = numpy.array([1, 1])
numpy.testing.assert_array_almost_equal(itq._norm_vector(v),
[1. / sqrt(2), 1. / sqrt(2)])
def test_fit_has_model(self):
# When trying to run fit where there is already a mean vector and
# rotation set.
itq = ItqFunctor()
itq.mean_vec = 'sim vec'
itq.rotation = 'sim rot'
self.assertRaisesRegex(RuntimeError,
"Model components have already been loaded.",
itq.fit, [])
def test_get_config_no_cache(self):
itq = ItqFunctor(bit_length=1, itq_iterations=2, normalize=3,
random_seed=4)
c = itq.get_config()
NT.assert_equal(c['bit_length'], 1)
NT.assert_equal(c['itq_iterations'], 2)
NT.assert_equal(c['normalize'], 3)
NT.assert_equal(c['random_seed'], 4)
NT.assert_is_none(c['mean_vec_cache']['type'])
NT.assert_is_none(c['rotation_cache']['type'])
def test_norm_vector_no_normalization(self):
itq = ItqFunctor(normalize=None)
v = numpy.array([0, 1])
numpy.testing.assert_array_equal(itq._norm_vector(v), v)
v = numpy.array([[0, 1, 1, .4, .1]])
numpy.testing.assert_array_equal(itq._norm_vector(v), v)
v = numpy.array([0] * 128)
numpy.testing.assert_array_equal(itq._norm_vector(v), v)
def test_save_model_no_caches(self):
expected_mean_vec = numpy.array([1, 2, 3])
expected_rotation = numpy.eye(3)
# Cache variables should remain None after save.
itq = ItqFunctor()
itq.mean_vec = expected_mean_vec
itq.rotation = expected_rotation
itq.save_model()
self.assertIsNone(itq.mean_vec_cache_elem)
self.assertIsNone(itq.mean_vec_cache_elem)
def test_configuration(self):
i = LSHNearestNeighborIndex(lsh_functor=ItqFunctor(),
descriptor_set=MemoryDescriptorSet(),
hash2uuids_kvstore=MemoryKeyValueStore(),
hash_index=LinearHashIndex(),
distance_method='euclidean',
read_only=True)
for inst in configuration_test_helper(
i): # type: LSHNearestNeighborIndex
assert isinstance(inst.lsh_functor, LshFunctor)
assert isinstance(inst.descriptor_set, MemoryDescriptorSet)
assert isinstance(inst.hash_index, LinearHashIndex)
assert isinstance(inst.hash2uuids_kvstore, MemoryKeyValueStore)
assert inst.distance_method == 'euclidean'
assert inst.read_only is True
def test_has_model(self):
itq = ItqFunctor()
# with no vector/rotation set, should return false.
self.assertFalse(itq.has_model())
# If only one of the two is None, then false should be returned.
itq.mean_vec = 'mean vec'
itq.rotation = None
self.assertFalse(itq.has_model())
itq.mean_vec = None
itq.rotation = 'rotation'
self.assertFalse(itq.has_model())
# If both are not None, return true.
itq.mean_vec = 'mean vec'
itq.rotation = 'rotation'
self.assertTrue(itq.has_model())
def test_get_config_with_cache_elements(self):
itq = ItqFunctor(bit_length=5, itq_iterations=6, normalize=7,
random_seed=8)
itq.mean_vec_cache_elem = DataMemoryElement('cached vec bytes')
itq.rotation_cache_elem = DataMemoryElement('cached rot bytes')
c = itq.get_config()
NT.assert_equal(c['bit_length'], 5)
NT.assert_equal(c['itq_iterations'], 6)
NT.assert_equal(c['normalize'], 7)
NT.assert_equal(c['random_seed'], 8)
NT.assert_equal(c['mean_vec_cache']['type'], "DataMemoryElement")
NT.assert_equal(c['mean_vec_cache']['DataMemoryElement']['bytes'],
'cached vec bytes')
NT.assert_equal(c['rotation_cache']['DataMemoryElement']['bytes'],
'cached rot bytes')
def test_fit(self):
fit_descriptors = []
for i in range(5):
d = DescriptorMemoryElement('test', i)
d.set_vector([-2. + i, -2. + i])
fit_descriptors.append(d)
itq = ItqFunctor(bit_length=1, random_seed=0)
itq.fit(fit_descriptors)
# TODO: Explanation as to why this is the expected result.
numpy.testing.assert_array_almost_equal(itq.mean_vec, [0, 0])
numpy.testing.assert_array_almost_equal(itq.rotation,
[[1 / sqrt(2)], [1 / sqrt(2)]])
NT.assert_is_none(itq.mean_vec_cache_elem)
NT.assert_is_none(itq.rotation_cache_elem)
def test_get_config_with_cache_elements(self):
itq = ItqFunctor(bit_length=5, itq_iterations=6, normalize=7,
random_seed=8)
itq.mean_vec_cache_elem = DataMemoryElement(b'cached vec bytes')
itq.rotation_cache_elem = DataMemoryElement(b'cached rot bytes')
c = itq.get_config()
self.assertEqual(c['bit_length'], 5)
self.assertEqual(c['itq_iterations'], 6)
self.assertEqual(c['normalize'], 7)
self.assertEqual(c['random_seed'], 8)
self.assertEqual(c['mean_vec_cache']['type'], "DataMemoryElement")
# Check using string encodings of set bytes (JSON compliant).
self.assertEqual(c['mean_vec_cache']['DataMemoryElement']['bytes'],
'cached vec bytes')
self.assertEqual(c['rotation_cache']['DataMemoryElement']['bytes'],
'cached rot bytes')
def compute_hash_codes(task, folderId, **kwargs):
"""
Celery task for computing hash codes on a given folder (descriptor index).
:param task: Celery provided task object.
:param folderId: The folder to train ITQ for, note this is only used to
infer the descriptor index.
"""
task.job_manager.updateProgress(message='Computing Hash Codes',
forceFlush=True)
index = descriptorIndexFromFolderId(task.girder_client, folderId)
smqtkFolder = getCreateFolder(task.girder_client, folderId, '.smqtk')
meanVecFileId = smqtkFileIdFromName(task.girder_client, smqtkFolder,
'mean_vec.npy')
rotationFileId = smqtkFileIdFromName(task.girder_client, smqtkFolder,
'rotation.npy')
hash2uuidsFile = initializeItemWithFile(
task.girder_client,
createOverwriteItem(task.girder_client, smqtkFolder['_id'],
'hash2uuids.pickle'))
functor = ItqFunctor(
mean_vec_cache=GirderDataElement(
meanVecFileId,
api_root=task.request.apiUrl,
token=task.request.jobInfoSpec['headers']['Girder-Token']),
rotation_cache=GirderDataElement(
rotationFileId,
api_root=task.request.apiUrl,
token=task.request.jobInfoSpec['headers']['Girder-Token']))
hash2uuids = compute_functions.compute_hash_codes(index.iterkeys(),
index,
functor,
use_mp=False)
data = pickle.dumps(dict((y, x) for (x, y) in hash2uuids))
task.girder_client.uploadFileContents(hash2uuidsFile['_id'],
six.BytesIO(data), len(data))
def itq(task, folderId, **kwargs):
"""
Celery task for training ITQ on a given folder.
This trains ITQ on all descriptors within the index. Since this
is typically called after computing descriptors, it will often
only contain what's in the folder.
:param task: Celery provided task object.
:param folderId: The folder to train ITQ for, note this is only used to
infer the descriptor index.
"""
task.job_manager.updateProgress(message='Training ITQ', forceFlush=True)
index = descriptorIndexFromFolderId(task.girder_client, folderId)
if not index.count():
# TODO SMQTK should account for this?
raise Exception('Descriptor index is empty, cannot train ITQ.')
smqtkFolder = getCreateFolder(task.girder_client, folderId, '.smqtk')
meanVecFile = initializeItemWithFile(
task.girder_client,
createOverwriteItem(task.girder_client, smqtkFolder['_id'],
'mean_vec.npy'))
rotationFile = initializeItemWithFile(
task.girder_client,
createOverwriteItem(task.girder_client, smqtkFolder['_id'],
'rotation.npy'))
functor = ItqFunctor(
mean_vec_cache=GirderDataElement(
meanVecFile['_id'],
api_root=task.request.apiUrl,
token=task.request.jobInfoSpec['headers']['Girder-Token']),
rotation_cache=GirderDataElement(
rotationFile['_id'],
api_root=task.request.apiUrl,
token=task.request.jobInfoSpec['headers']['Girder-Token']))
functor.fit(index.iterdescriptors(), use_multiprocessing=False)
def test_build_index_read_only(self):
index = LSHNearestNeighborIndex(ItqFunctor(),
MemoryDescriptorIndex(),
MemoryKeyValueStore(),
read_only=True)
ntools.assert_raises(ReadOnlyError, index.build_index, [])
