def load_piece_list(piece_names, aug_config=NO_AUGMENT):
"""
Collect piece data
"""
all_images = []
all_specs = []
all_o2c_maps = []
#comment changed range from len(piece_names) to 2
for ip in tqdm(2, ncols=70):
piece_name = piece_names[ip]
try:
image, specs, o2c_maps = prepare_piece_data(DATA_ROOT_MSMD, piece_name,
aug_config=aug_config, require_audio=False)
except:
print("Problems with loading piece %s" % piece_name)
print(sys.exc_info()[0])
continue
# keep stuff
all_images.append(image)
all_specs.append(specs)
all_o2c_maps.append(o2c_maps)
return all_images, all_specs, all_o2c_maps
def initialize_sheet_db(self, pieces, keep_snippets=False):
""" init sheet music data base """
print("Initializing sheet music db ...")
self.id_to_piece = dict()
self.sheet_snippet_ids = np.zeros(0, dtype=np.int)
self.sheet_snippet_codes = np.zeros((0, self.embed_network.code_dim),
dtype=np.float32)
self.sheet_snippets = np.zeros(
(0, self.snippet_shape[0] // 2, self.snippet_shape[1] // 2),
dtype=np.uint8)
# initialize retrieval pool
for piece_idx, piece in enumerate(pieces):
print(" (%03d / %03d) %s" % (piece_idx + 1, len(pieces), piece))
# load piece
self.id_to_piece[piece_idx] = piece
piece_image, piece_specs, piece_o2c_maps = prepare_piece_data(
ROOT_DIR, piece, require_audio=False)
# initialize data pool with piece
piece_pool = AudioScoreRetrievalPool([piece_image], [piece_specs],
[piece_o2c_maps],
data_augmentation=NO_AUGMENT,
shuffle=False)
# embed sheet snippets of piece
snippets = np.zeros((piece_pool.shape[0], 1, self.sheet_shape[0],
self.sheet_shape[1]),
dtype=np.uint8)
for i in range(piece_pool.shape[0]):
# get image snippet
snippet, _ = piece_pool[i:i + 1]
snippet = snippet[0, 0]
snippets[i, 0] = snippet
# keep sheet snippets
if keep_snippets:
snippet = resize_image(snippet, rsz=0.5).astype(
np.uint8)[np.newaxis]
self.sheet_snippets = np.concatenate(
(self.sheet_snippets, snippet))
# compute sheet snippet codes
codes = self.embed_network.compute_view_1(snippets)
# keep codes
self.sheet_snippet_codes = np.concatenate(
(self.sheet_snippet_codes, codes))
# save id of piece
piece_ids = np.ones(piece_pool.shape[0], dtype=np.int) * piece_idx
self.sheet_snippet_ids = np.concatenate(
(self.sheet_snippet_ids, piece_ids))
print("%s sheet snippet codes of %d pieces collected" %
(self.sheet_snippet_codes.shape[0], len(pieces)))
def initialize_audio_db(self, pieces, augment, keep_snippets=False):
""" init audio data base """
print("Initializing audio db ...")
self.id_to_perform = dict()
self.perform_excerpt_ids = np.zeros(0, dtype=np.int)
self.perform_excerpt_codes = np.zeros((0, self.embed_network.code_dim),
dtype=np.float32)
self.perform_excerpts = np.zeros(
(0, self.excerpt_shape[0] // 2, self.excerpt_shape[1] // 2),
dtype=np.uint8)
# initialize retrieval pool
for piece_idx, piece in enumerate(pieces):
print(" (%03d / %03d) %s" % (piece_idx + 1, len(pieces), piece))
# load piece
self.id_to_perform[piece_idx] = piece
piece_image, piece_specs, piece_o2c_maps = prepare_piece_data(
ROOT_DIR, piece, aug_config=augment, require_audio=False)
# initialize data pool with piece
piece_pool = AudioScoreRetrievalPool([piece_image], [piece_specs],
[piece_o2c_maps],
data_augmentation=augment,
shuffle=False)
# embed audio excerpt of piece
excerpts = np.zeros((piece_pool.shape[0], 1, self.spec_shape[0],
self.spec_shape[1]),
dtype=np.float32)
for j in range(piece_pool.shape[0]):
# get spectrogram excerpt
_, spec = piece_pool[j:j + 1]
excerpts[j, 0] = spec
# TODO: keep excerpt snippets
# don't know yet how much sense this makes
if keep_snippets:
pass
# compute audio excerpt code
codes = self.embed_network.compute_view_2(excerpts)
# keep code
self.perform_excerpt_codes = np.concatenate(
(self.perform_excerpt_codes, codes))
# save id of piece
piece_ids = np.ones(piece_pool.shape[0], dtype=np.int) * piece_idx
self.perform_excerpt_ids = np.concatenate(
(self.perform_excerpt_ids, piece_ids))
print("%s audio excerpts of %d pieces collected" %
(self.perform_excerpt_codes.shape[0], len(pieces)))
sheet_win_shape = model.INPUT_SHAPE_1[1:]
spec_win_shape = model.INPUT_SHAPE_2[1:]
# select pieces
split = load_split(args.train_split)
te_pieces = split["test"]
# pieces = [te_pieces[te_pieces.index("BachJS__BWV790__bwv790")]]
pieces = te_pieces
# collect pixel errors for all pieces
piece_pxl_errors = dict()
for piece in pieces:
print("\nTarget Piece: %s" % piece)
piece_image, piece_specs, piece_o2c_maps = prepare_piece_data(
ROOT_DIR, piece, aug_config=test_augment, require_audio=False)
# initialize data pool with piece
piece_pool = AudioScoreRetrievalPool([piece_image], [piece_specs],
[piece_o2c_maps],
data_augmentation=test_augment,
shuffle=False)
# compute spectrogram from file
if args.real_audio:
audio_file = "/home/matthias/cp/data/sheet_localization/real_music/0_real_audio/%s.flac" % piece
if not os.path.exists(audio_file):
continue
spec = processor.process(audio_file).T
# use pre-computed spectrogram
else:
a2s_srv.initialize_audio_db(pieces=te_pieces, augment=test_augment, keep_snippets=False)
a2s_srv.save_audio_db_file("audio_db_file.pkl")
# load sheet music data base
else:
a2s_srv.load_audio_db_file("audio_db_file.pkl")
# run full evaluation
if args.full_eval:
print(col.print_colored("\nRunning full evaluation:", col.UNDERLINE))
ranks = []
for tp in te_pieces:
# load piece
piece_image, _, _ = prepare_piece_data(ROOT_DIR, tp, aug_config=test_augment, require_audio=False)
# detect piece from spectrogram
ret_result, ret_votes = a2s_srv.detect_performance(piece_image, top_k=len(te_pieces), n_candidates=args.n_candidates, verbose=False)
if tp in ret_result:
rank = ret_result.index(tp) + 1
ratio = ret_votes[ret_result.index(tp)]
else:
rank = len(ret_result)
ratio = 0.0
ranks.append(rank)
color = col.OKBLUE if ranks[-1] == 1 else col.WARNING
print(col.print_colored("rank: %02d (%.2f) " % (ranks[-1], ratio), color) + tp)
# report results
ranks = np.asarray(ranks)
# path to MSMD data set
DATA_ROOT_MSMD = '/Users/abdelrahman/Desktop/Bachelor References/msmd_aug'
# for now we select just some pieces
# (this could be also the entire training set)
piece_names = [
'BachCPE__cpe-bach-rondo__cpe-bach-rondo', 'BachJS__BWV259__bwv-259',
'AdamA__giselle__giselle'
]
all_piece_images = []
all_piece_specs = []
all_piece_o2c_maps = []
for piece_name in piece_names:
piece_image, piece_specs, piece_o2c_maps = prepare_piece_data(
DATA_ROOT_MSMD, piece_name, require_audio=False)
# keep stuff
all_piece_images.append(piece_image)
all_piece_specs.append(piece_specs)
all_piece_o2c_maps.append(piece_o2c_maps)
# path to network parameters
param_file = "/Users/abdelrahman/Desktop/resultdumper/mutopia_ccal_cont_est_UV/params_all_split_mutopia_full_aug.pkl"
# this function is called before a snippet is fed to the network
# (for the present model it resizes the image by a factor of 2)
prepare_sheet_img = model.prepare
# initialize retrieval wrapper
embed_network = RetrievalWrapper(model,