def add_classes(db_path, class_csv):
db = datasetSQL.LabelSet(db_path)
with open(class_csv, 'r') as f:
for item in csv.DictReader(f):
class_item = {}
class_item['ASID'] = item['mid']
class_item['class_name'] = item['display_name'].replace("'", "''")
print(item)
db.__insert__(class_item, 'classes')
db.__commit__()
def initiate_database(db_path, segment_csv):
db = datasetSQL.LabelSet(db_path)
db.initialize()
with open(segment_csv, 'r') as f:
for item in csv.DictReader(f):
segment = {}
segment['segment_id'] = item['YTID']
print(segment)
db.__insert__(segment, 'segments')
db.__commit__()
def evaluate(weight_path, db_path, feature_path):
db = datasetSQL.LabelSet(db_path)
h5r = h5py.File(feature_path, 'r')
h5w = h5py.File('/tmp/esc_tmp_1.hdf5', 'w')
db.cursor.execute(
"SELECT segment_id FROM segments WHERE audio_file NOT NULL ORDER BY segment_id ASC;"
)
segment_list = [
record[0].decode('utf-8') for record in db.cursor.fetchall()
]
n_segment = len(segment_list)
h5w.create_dataset('max', data=np.zeros((n_segment, embedding_length)))
emb_net = network_arch.EmbNet().to(device)
emb_net.load_weight(weight_path)
emb_net.eval()
for i, segment_id in enumerate(segment_list):
f = h5r[segment_id][:]
if len(f) < 161:
f = padding(f, 161)
n, D = f.shape
data = np.zeros((1, 1, n, D))
data[0, 0, :, :] = f
torch_data = torch.from_numpy(data).float().to(device)
with torch.no_grad():
#pred = netx(torch_data)[0]
"""
pred = emb_net(torch_data)
#print (pred.size())
if len(pred.size()) > 2:
embedding = torch.max(pred, 2)[0]
embedding = embedding.view(embedding.size(0),-1)
else:
embedding = pred
"""
embedding = emb_net(torch_data)
embedding = embedding.mean(dim=1)
#print (embedding.size())
h5w['max'][i] = embedding.cpu().numpy()[0]
h5w.create_dataset(segment_id, data=embedding.cpu().numpy()[0])
h5w.close()
h5r2 = h5py.File('/tmp/esc_tmp_1.hdf5', 'r')
h5w2 = h5py.File('/tmp/esc_tmp_dist_1.hdf5', 'w')
similarity.Dist_gpu(h5r2, h5w2)
h5w2.close()
h5r3 = h5py.File('/tmp/esc_tmp_dist_1.hdf5', 'r')
return (similarity_analysis.mAP2(h5r3, db))
def add_labels(db_path, segment_csv):
db = datasetSQL.LabelSet(db_path)
with open(segment_csv, 'r') as f:
for line in f.readlines():
segment_id = line.split(',')[0]
if segment_id == "YTID": #title line
continue
labels = ','.join(line.split(',')[3:]).replace('"', '').strip()
for label in labels.split(','):
sql = """
INSERT INTO labels (segment_id, class_id, label_type)
SELECT segments.segment_id, classes.class_id, 0 FROM segments CROSS JOIN classes WHERE segments.segment_id = '{0}' AND classes.ASID = '{1}'
""".format(segment_id, label)
print(sql)
db.cursor.execute(sql)
db.__commit__()
def link_audio_file(db_path, filecsv):
db = datasetSQL.LabelSet(db_path)
with open(filecsv, 'r') as f:
#for item in csv.DictReader(f):
for line in f.readlines():
item = {}
line = line.rstrip()
item["YTID"] = line.split(',')[-1]
item["filename"] = ','.join(line.split(',')[:-1]) + '.wav'
sql = """
UPDATE segments SET audio_file = '{0}'
WHERE segment_id = '{1}'
""".format(item['filename'].replace("'", "''"), item['YTID'])
db.cursor.execute(sql)
db.__commit__()
def compute_features(db_path, feature_path, wav_root):
h5w = h5py.File(feature_path, 'w')
db = datasetSQL.LabelSet(db_path)
#trg_sr = 48000
trg_sr = 16000
sql = """
SELECT segment_id, audio_file FROM segments
WHERE audio_file NOT NULL
"""
segment_list = db.cursor.execute(sql)
for segment_tuple in segment_list:
segment_id, audio_file = segment_tuple[0].decode(
'utf-8'), segment_tuple[1].decode('utf-8')
print(segment_id, audio_file)
"""
y, src_sr = soundfile.read(os.path.join(wav_root, audio_file))
if len(y.shape) > 1:
y = y[:,0]
y = librosa.core.resample(y, src_sr, trg_sr)
"""
y = librosa.load(os.path.join(wav_root, audio_file), trg_sr)[0]
try:
#mel = librosa.feature.melspectrogram(y,trg_sr,n_fft=1024,hop_length=512,n_mels=128)
mel = librosa.feature.melspectrogram(y,
trg_sr,
n_fft=1024,
hop_length=500,
n_mels=64)
log_mel = librosa.power_to_db(mel).T
print(log_mel.shape)
h5w.create_dataset(segment_id, data=log_mel)
except:
print("Failure:", segment_id, audio_file)
sql = """
UPDATE segments SET audio_file=NULL WHERE segment_id = '{0}'
""".format(segment_id)
db.cursor.execute(sql)
db.__commit__()
continue
return
def train(db_path, feature_path):
db = datasetSQL.LabelSet(db_path)
h5r = h5py.File(feature_path, 'r')
db.cursor.execute("SELECT COUNT(*) FROM classes;")
n_class = db.cursor.fetchone()[0]
cascade_net = network_arch.CascadeNet(n_class).to(device)
if os.path.isfile(tmp_model_weight_path1):
cascade_net.load_weight(tmp_model_weight_path1)
optimizer = optim.Adam(cascade_net.parameters(), lr=1e-4)
loss_function = nn.BCELoss()
cascade_net.train()
best_epoch = 0
class_list = np.random.permutation(range(2, n_class + 1))
for i_epoch in range(100):
losses = 0
db.cursor.execute(
"SELECT segment_id FROM segments WHERE audio_file NOT NULL")
#segment_list = [record[0].decode('utf-8') for record in db.cursor.fetchall()]
segment_list = [
record[0].decode('utf-8').strip()
for record in db.cursor.fetchall()
]
n_segment = len(segment_list)
print(i_epoch, n_segment)
order_list = np.random.permutation(range(n_segment))
for start_index in range(0, n_segment, training_batch_size):
if start_index + training_batch_size > n_segment:
continue
batch_data = np.zeros((training_batch_size, 1, target_input_length,
target_input_dim))
batch_target = np.zeros((training_batch_size, n_class))
optimizer.zero_grad()
for i in range(training_batch_size):
segment_id = segment_list[order_list[start_index + i]]
try:
f = h5r[segment_id][:]
if len(f) != target_input_length:
f = padding(f)
batch_data[i, 0, :, :] = f
sql = """
SELECT class_id FROM labels WHERE segment_id = '{0}'
""".format(segment_id)
db.cursor.execute(sql)
records = db.cursor.fetchall()
for record in records:
batch_target[i, record[0] - 1] = 1
except:
print(segment_id + ' None!')
torch_train = torch.from_numpy(batch_data).float().to(device)
torch_target = torch.from_numpy(batch_target).float().to(device)
for i in range(1):
torch_output = cascade_net(torch_train)
loss = loss_function(torch_output, torch_target)
loss.backward()
optimizer.step()
print(start_index, loss)
losses += loss
print("epoch {0} loss: {1}".format(i_epoch, losses))
torch.save(cascade_net.state_dict(), tmp_model_weight_path1)
torch.save(cascade_net.emb_net.state_dict(), tmp_model_weight_path0)
print("Model has been saved...")
if i_epoch % 1 == 0:
criteria_i = evaluate(tmp_model_weight_path0,
'database/db_esc10.sqlite',
'database/mel_esc10.hdf5')
if criteria_i > best_epoch:
best_epoch = criteria_i
copyfile(tmp_model_weight_path0, best_model_weight_path0)
copyfile(tmp_model_weight_path1, best_model_weight_path1)
def refine_labels(db_path):
#Strategy: loop five times to ensure all the existing labels have their parental classes labeled.
import json
with open(ontology_json, 'r') as f:
ontology_list = json.load(f)
db = datasetSQL.LabelSet(db_path)
sql = """
ALTER TABLE classes
ADD COLUMN leaf_node BOOL
"""
try:
db.cursor.execute(sql)
except:
pass
for loop_i in range(5):
for class_dict in ontology_list:
parent_asid = class_dict['id']
sql = """
SELECT class_id from classes WHERE ASID = '{0}'
""".format(parent_asid)
db.cursor.execute(sql)
record = db.cursor.fetchone()
if record: #Some class in ontology does not exist in dataset
parent_id = record[0]
else:
continue
if len(class_dict['child_ids']) == 0:
sql = """
UPDATE classes SET leaf_node=1
WHERE class_id = {0}
""".format(parent_id)
else:
sql = """
UPDATE classes SET leaf_node=0
WHERE class_id = {0}
""".format(parent_id)
db.cursor.execute(sql)
for child_asid in class_dict['child_ids']:
sql = """
SELECT class_id from classes WHERE ASID = '{0}'
""".format(child_asid)
db.cursor.execute(sql)
record = db.cursor.fetchone()
if record: #Some class in ontology does not exist in dataset
child_id = record[0]
else:
continue
print(parent_asid, child_asid)
sql = """
INSERT INTO labels (segment_id, class_id, label_type)
SELECT segment_id, {1},0 from segments
WHERE segment_id IN
(SELECT segment_id FROM labels WHERE class_id = {0})
AND segment_id NOT IN
(SELECT segment_id FROM labels WHERE class_id = {1})
""".format(child_id, parent_id)
db.cursor.execute(sql)
#print (db.cursor.fetchall())
db.__commit__()
db.__close__()
