def __init__(self, glyph_classifier_fn=None):
"""Creates the engine and TF graph for running OMR.
Args:
glyph_classifier_fn: Callable that loads the glyph classifier into the
graph. Accepts a `Structure` as the single argument, and returns an
instance of `BaseGlyphClassifier`. The function typically loads a TF
saved model or other external data, and wraps the classification in a
concrete glyph classifier subclass. If the classifier uses a
`StafflineExtractor` for classification, it must set the
`staffline_extractor` attribute of the `Structure`. Otherwise, glyph x
coordinates will not be scaled back to image coordinates.
"""
glyph_classifier_fn = (glyph_classifier_fn
or saved_classifier_fn.build_classifier_fn())
self.graph = tf.Graph()
self.session = tf.Session(graph=self.graph)
with self.graph.as_default():
with self.session.as_default():
with tf.name_scope('OMREngine'):
self.png_path = tf.placeholder(tf.string,
name='png_path',
shape=())
self.image = image.decode_music_score_png(
tf.read_file(self.png_path, name='page_image'))
self.structure = structure_module.create_structure(
self.image)
# Loading saved models happens outside of the name scope, because scopes
# can rename tensors from the model and cause dangling references.
# TODO(ringw): TF should be able to load models gracefully within a
# name scope.
self.glyph_classifier = glyph_classifier_fn(self.structure)
def testGetPage_x_scale(self):
# Random staffline images matching the dimensions of PREDICTIONS.
dummy_stafflines = np.random.random((2, 3, 5, 6))
classifier = glyphs_testing.DummyGlyphClassifier(glyphs_testing.PREDICTIONS)
image = np.random.randint(0, 255, (30, 20), dtype=np.uint8)
staves = staves_testing.FakeStaves(
image_t=image,
staves_t=np.asarray([[[0, 10], [19, 10]], [[0, 20], [19, 20]]],
np.int32),
staffline_distance_t=np.asarray([5, 20], np.int32),
staffline_thickness_t=np.asarray(1, np.int32))
structure = structure_module.create_structure(image,
lambda unused_image: staves)
class DummyStafflineExtractor(object):
"""A placeholder for StafflineExtractor.
It only contains the constants necessary to scale the x coordinates.
"""
staffline_distance_multiple = 2
target_height = 10
omr = engine.OMREngine(lambda _: classifier)
page = omr.process_image(
# Feed in a dummy image. It doesn't matter because FakeStaves has
# hard-coded staff values.
np.random.randint(0, 255, (100, 100)),
process_structure=False)
page = staff_processor.StaffProcessor(structure,
DummyStafflineExtractor()).apply(page)
self.assertEqual(len(page.system[0].staff), 2)
# The first staff has a staffline distance of 5.
# The extracted staffline slices have an original height of
# staffline_distance * staffline_distance_multiple (10), which equals
# target_height here, so there is no scaling.
self.assertEqual(
musicscore_pb2.Staff(glyph=page.system[0].staff[0].glyph),
glyphs_testing.GLYPHS_PAGE.system[0].staff[0])
# Glyphs in the second staff have a scaled x coordinate.
self.assertEqual(
len(page.system[0].staff[1].glyph),
len(glyphs_testing.GLYPHS_PAGE.system[0].staff[1].glyph))
for glyph in glyphs_testing.GLYPHS_PAGE.system[0].staff[1].glyph:
expected_glyph = copy.deepcopy(glyph)
# The second staff has a staffline distance of 20. The extracted staffline
# slice would be 4 times the size of the scaled staffline, so x
# coordinates are scaled by 4. Also, the glyphs may be in a different
# order.
expected_glyph.x *= 4
self.assertIn(expected_glyph, page.system[0].staff[1].glyph)
def testCompute(self):
filename = os.path.join(tf.resource_loader.get_data_files_path(),
'../testdata/IMSLP00747-000.png')
image = image_module.decode_music_score_png(tf.read_file(filename))
struct = structure.create_structure(image)
with self.test_session():
struct = struct.compute()
self.assertEqual(np.int32, struct.staff_detector.staves.dtype)
# Expected number of staves for the corpus image.
self.assertEqual((12, 2, 2), struct.staff_detector.staves.shape)
self.assertEqual(np.int32, struct.verticals.lines.dtype)
self.assertEqual(3, struct.verticals.lines.ndim)
self.assertEqual((2, 2), struct.verticals.lines.shape[1:])
def testSaveAndLoadDummyClassifier(self):
with tempfile.TemporaryDirectory() as base_dir:
export_dir = os.path.join(base_dir, 'export')
with self.test_session() as sess:
patches = tf.placeholder(tf.float32, shape=(None, 18, 15))
num_patches = tf.shape(patches)[0]
# Glyph.NONE is number 1.
class_ids = tf.ones([num_patches], tf.int32)
signature = tf.saved_model.signature_def_utils.build_signature_def(
# pyformat: disable
{'input': tf.saved_model.utils.build_tensor_info(patches)},
{
'class_ids':
tf.saved_model.utils.build_tensor_info(class_ids)
},
'serve')
builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
builder.add_meta_graph_and_variables(
sess, ['serve'],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
signature
})
builder.save()
tf.reset_default_graph()
# Load the saved model.
with self.test_session() as sess:
filename = os.path.join(
tf.resource_loader.get_data_files_path(),
'../testdata/IMSLP00747-000.png')
page = image.decode_music_score_png(tf.read_file(filename))
clazz = saved_classifier.SavedConvolutional1DClassifier(
structure.create_structure(page), export_dir)
# Run min length should be the default.
self.assertEqual(clazz.run_min_length,
convolutional.DEFAULT_RUN_MIN_LENGTH)
predictions = clazz.staffline_predictions.eval()
self.assertEqual(predictions.ndim,
3) # Staff, staff position, x
self.assertGreater(predictions.size, 0)
# Predictions are all musicscore_pb2.Glyph.NONE.
self.assertAllEqual(
predictions,
np.full(predictions.shape, musicscore_pb2.Glyph.NONE,
np.int32))
