def get_model(nfets, size, nclass):
Timer.start('Building model')
model = keras.Sequential()
model.add(keras.layers.BatchNormalization(batch_size=300))
model.add(
keras.layers.Conv1D(64,
kernel_size=4 * nfets,
strides=nfets,
input_shape=(size, 1),
activation='elu'))
model.add(keras.layers.Dropout(0.005))
model.add(keras.layers.MaxPooling1D(pool_size=6, strides=4))
model.add(
keras.layers.Conv1D(128, kernel_size=10, strides=2, activation='elu'))
model.add(keras.layers.MaxPooling1D(pool_size=4, strides=2))
model.add(
keras.layers.Conv1D(128, kernel_size=2, strides=2, activation='elu'))
model.add(keras.layers.MaxPooling1D(pool_size=2, strides=2))
model.add(
keras.layers.Conv1D(128, kernel_size=2, strides=2, activation='elu'))
model.add(keras.layers.MaxPooling1D(pool_size=2, strides=2))
model.add(keras.layers.GRU(32, return_sequences=True))
model.add(keras.layers.GRU(32, return_sequences=False))
model.add(keras.layers.Dropout(0.001))
# model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(units=1000, activation='elu'))
model.add(keras.layers.Dropout(0.005))
model.add(keras.layers.Dense(units=nclass, activation='softmax'))
Timer.end('Building model')
return model
def make_data(ext, in_dir, out_dir, n_pack):
mkdirp(out_dir)
allfiles = os.listdir(in_dir)
packs = dist(allfiles, n_pack)
print([len(p) for p in packs])
for pi, files in enumerate(packs):
Timer.start('Making data for pack %i / %i' % (pi + 1, len(packs)))
out_file = '%s/data_%i.%s' % (out_dir, pi, ext)
if os.path.exists(out_file): continue
exe = ThreadPoolExecutor(max_workers=16)
records = {}
meta_dict = {
'hop': hop,
'nfft': nfft,
'win_len': win_len,
'records': records,
'labels': []
}
n = len(files)
pack_data = []
for idx, _file in enumerate(files):
def f(full_file, i):
file_with_lbl, file_ext = os.path.splitext(full_file)
if file_ext not in audio_exts: return
lbl, file = os.path.splitext(file_with_lbl)
lbl = int(lbl)
file = file[1:]
Timer.start('Making data for %s(%i) (%i/%i)' %
(file, pi + 1, i + 1, n))
meta = {'file': file, 'label': lbl}
y, sr = librosa.load('%s/%i.%s%s' %
(in_dir, lbl, file, file_ext))
meta['sr'] = sr
# data = make_file_data(y, sr, hop, nfft, win_len)
data = extract_features(y, sr, nfft, hop)
pack_data.append(data)
records[file] = meta
meta_dict['labels'].append(lbl)
Timer.end('Making data for %s(%i) (%i/%i)' %
(file, pi + 1, i + 1, n))
exe.submit(lambda _file=_file, idx=idx: f(_file, idx))
exe.shutdown(wait=True)
nppack = np.asarray(pack_data)
print(nppack.shape)
meta_dict['rows'] = len(meta_dict['labels'])
meta_dict['nfets'] = nppack.shape[2]
with open('%s/meta_%i.json' % (out_dir, pi), 'w') as fp:
json.dump(meta_dict, fp)
nppack.tofile(out_file)
Timer.end('Making data for pack %i / %i' % (pi + 1, len(packs)))
def f(i, file):
Timer.start('Make test data for file: %i / %i' %
(i + 1, len(pack)))
y, sr = lrs.load('%s/%s' % (dir, file), sr=base_sr)
n = 1 + ((len(y) - base_len) // stride)
for j in range(n):
start = j * stride
end = start + base_len
chunk_data = make_file_data(y[start:end], sr, hop, nfft,
win_len)
data.append(chunk_data)
file_names.append(file)
Timer.end('Make test data for file: %i / %i' %
(i + 1, len(pack)))
def train_file(model, dir, file, x_val, y_val):
nb_epoch = 15
batch_size = 200
print('Training file %s' % file)
Timer.start('Training file %s' % file)
x_train, y_train, meta = load_file(dir, file)
# early_stopping_callback = keras.callbacks.EarlyStopping(monitor='acc', min_delta=0.001, patience=200, verbose=1, mode='auto', baseline=None)
# save_model = keras.callbacks.ModelCheckpoint(model_file, monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=10)
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=nb_epoch,
validation_data=(x_val, y_val),
verbose=2)
Timer.end('Training file %s' % file)
def run():
ext = 'pdt3'
run_id = 3
run_key = '%s:%i' % (ext, run_id)
dir = '/Users/salm/Desktop/zalo_music/extra_%s' % ext
model_file = '/Users/salm/Desktop/zalo_music/model_%s.%s.mdl' % (run_id,
ext)
files = list(
filter(lambda f: os.path.splitext(f)[1] == '.%s' % ext,
os.listdir(dir)))
val_file = files[len(files) - 1]
files = files[:len(files) - 1]
classes = [
'cai_luong', 'cach_mang', 'dan_ca', 'dance', 'khong_loi', 'thieu_nhi',
'trinh', 'tru_tinh', 'rap', 'rock'
]
x_val, y_val, meta = load_file(dir, val_file)
nfets = meta['nfets']
if os.path.isfile(model_file):
model = keras.models.load_model(model_file)
else:
model = get_model(nfets, x_val.shape[1], len(classes))
model.compile(loss='sparse_categorical_crossentropy',
optimizer=keras.optimizers.Adam(),
metrics=['accuracy'])
nb_runs = 10
r = int(rd.get('zlm:%s:runs' % run_key) or 0)
while r < nb_runs:
Timer.start('Training run %i / %i' % (r + 1, nb_runs))
i = int(rd.hget('zlm:%s:file' % run_key, r) or 0)
while i < len(files):
file = files[i]
train_file(model, dir, file, x_val, y_val)
model.save(model_file)
i = rd.hincrby('zlm:%s:file' % run_key, r)
r = rd.incr('zlm:%s:runs' % run_key)
Timer.end('Training run %i / %i' % (r + 1, nb_runs))
def make_test_data(dir, out_dir, ext):
mkdirp(out_dir)
files = list(
filter(lambda f: os.path.splitext(f)[1] in music_exts,
os.listdir(dir)))
packs = dist(files, len(files) // 300)
n_pack = len(packs)
for pi, pack in enumerate(packs):
out_file = '%s/data_%i.%s' % (out_dir, pi, ext)
if os.path.isfile(out_file): continue
Timer.start('Make test data pack %i / %i' % (pi + 1, n_pack))
exe = ThreadPoolExecutor(max_workers=4)
file_names = []
data = []
for _i, _file in enumerate(pack):
def f(i, file):
Timer.start('Make test data for file: %i / %i' %
(i + 1, len(pack)))
y, sr = lrs.load('%s/%s' % (dir, file), sr=base_sr)
n = 1 + ((len(y) - base_len) // stride)
for j in range(n):
start = j * stride
end = start + base_len
chunk_data = make_file_data(y[start:end], sr, hop, nfft,
win_len)
data.append(chunk_data)
file_names.append(file)
Timer.end('Make test data for file: %i / %i' %
(i + 1, len(pack)))
exe.submit(lambda _file=_file, _i=_i: f(_i, _file))
exe.shutdown(wait=True)
np.asarray(data).tofile(out_file)
with open('%s/meta_%i.json' % (out_dir, pi), 'w') as fp:
json.dump(file_names, fp)
Timer.end('Make test data pack %i / %i' % (pi + 1, n_pack))
def f(full_file, i):
file_with_lbl, file_ext = os.path.splitext(full_file)
if file_ext not in audio_exts: return
lbl, file = os.path.splitext(file_with_lbl)
lbl = int(lbl)
file = file[1:]
Timer.start('Making data for %s(%i) (%i/%i)' %
(file, pi + 1, i + 1, n))
meta = {'file': file, 'label': lbl}
y, sr = librosa.load('%s/%i.%s%s' %
(in_dir, lbl, file, file_ext))
meta['sr'] = sr
# data = make_file_data(y, sr, hop, nfft, win_len)
data = extract_features(y, sr, nfft, hop)
pack_data.append(data)
records[file] = meta
meta_dict['labels'].append(lbl)
Timer.end('Making data for %s(%i) (%i/%i)' %
(file, pi + 1, i + 1, n))
def predict(model, dir, ext):
files = list(
filter(lambda f: os.path.splitext(f)[1] == '.%s' % ext,
os.listdir(dir)))
results = []
Timer.start('Prediction')
for file in files:
Timer.start('Predicting file %s' % file)
idx = int(file.split('.')[0].split('_')[1])
meta_file = 'meta_%s.json' % idx
with open('%s/%s' % (dir, meta_file), 'r') as fp:
file_names = json.load(fp)
x_predict = np.fromfile('%s/%s' % (dir, file))
x_predict = x_predict.reshape(len(file_names), -1, 1)
predictions = model.predict(x_predict)
for file_name, prediction in zip(file_names, predictions):
results.append({'file': file_name, 'predict': prediction.tolist()})
Timer.end('Predicting file %s' % file)
Timer.end('Prediction')
return results
def f(i, file):
Timer.start('Resampling file %i / %i' % (i + 1, n))
resampling('%s/%s' % (dir, file), lbl, out_dir)
Timer.end('Resampling file %i / %i' % (i + 1, n))