def load_from_file(folder_specs, data, type='data'):
data_dict = load_matrix(folder_specs, data)
if type == 'data':
return data_dict['x_data'], data_dict['y_data']
elif type == 'stats':
return data_dict['mean_x'], data_dict['std_x'], data_dict['fnames'], data_dict['sample_rate']
else:
print("Not supported")
def train_func(train_dir,
model_type='LRNN',
matrix_file='',
n_hid=1024,
n_recur=1,
epochs=100,
batch_size=10,
n_to_load=1,
highest_freq=5000,
clip_len=2,
mat_dirs=None,
chunks_per_sec=4,
down_sampling=False,
root_to_folder='/instrument_samples/',
save_weights=True,
add_spectra=False):
if matrix_file is '':
matrix_file = train_dir
dpath = config.datapath + root_to_folder + train_dir
fpath = dpath + '/' + matrix_file + '.npy'
d_mat_name = '/' + matrix_file + '_' + str(n_to_load) + 'files'
if down_sampling:
d_mat_name += str(chunks_per_sec) + 'res'
d_mat_name += str(highest_freq) + 'maxf'
else:
d_mat_name += '_raw'
if add_spectra:
d_mat_name += '_spec'
# see if matrix file exists
if not os.path.isfile(fpath):
# if not, create it from data
dpath = config.datapath + root_to_folder + train_dir
if os.path.isdir(dpath):
audios = AudioPipeline((root_to_folder + train_dir),
n_to_load=n_to_load,
highest_freq=highest_freq,
clip_len=clip_len,
mat_dirs=mat_dirs,
chunks_per_sec=chunks_per_sec,
down_sampling=down_sampling,
add_spectra=add_spectra)
audios.create_train_matrix(f_name_out=d_mat_name)
else:
print('both entered paths are invalid. No data loaded')
print('train directory: ', dpath)
print('matrix file: ', fpath)
return
data = load_matrix(root_to_folder + train_dir + '/', d_mat_name)
else:
data = load_matrix(root_to_folder + train_dir + '/', matrix_file)
x_data = data['x_data']
y_data = data['y_data']
num_frequency_dimensions = x_data.shape[2]
# create model
if model_type == 'LRNN':
model = create_lstm_network(num_frequency_dimensions, n_hid, n_recur)
elif model_type == 'CRNN':
y_data = x_data[:, :, :]
x_data = np.reshape(
x_data, (x_data.shape[0], x_data.shape[1], x_data.shape[2], 1))
model = create_conv_lstm_network(num_frequency_dimensions, n_hid,
n_recur)
model.summary()
w_mat_name = ""
print('Start Training')
model.fit(x_data,
y_data,
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_split=0.0)
print('Training complete')
w_mat_name = d_mat_name + '_' + str(n_hid) + 'hid_' + str(epochs) + 'ep'
if save_weights:
json_string = model.to_json()
model_output = config.datapath + '/weight_matrices/' + w_mat_name + '_model.json'
fout = open(model_output, 'w')
json.dump(json_string, fout)
fout.close()
print('saved model to: ' + model_output)
weights_output = config.datapath + '/weight_matrices/' + w_mat_name + '_weights.h5'
model.save_weights(weights_output, overwrite=True)
print('saved weights to: ' + weights_output)
return model, w_mat_name
def train_func(train_dir, matrix_file='', n_hid_neurons=1024, n_rec_layers=1, epochs=100, batch_size=5,
n_to_load=1, highest_freq=5000, clip_len=2, mat_dirs=None, chunks_per_sec=4,
down_sampling=False, root_to_folder='/instrument_samples/', save_weights=True, add_spectra=False,
activation='linear', architecture=1, mean_std_per_file=False):
if matrix_file is '':
matrix_file = train_dir.replace('/', '_') # just in case it happens to be a path
dpath = config.datapath + root_to_folder + train_dir
fpath = dpath + '/' + matrix_file + '.npy'
d_mat_name = '/' + matrix_file + '_' + str(n_to_load) + 'files'
d_mat_name += '_' + str(clip_len) + 'sec'
if down_sampling:
d_mat_name += '_' + str(chunks_per_sec) + 'res'
d_mat_name += '_' + str(highest_freq) + 'maxf'
else:
d_mat_name += '_raw'
if add_spectra:
d_mat_name += '_spec'
# see if matrix file exists
if not os.path.isfile(fpath):
# if not, create it from data
dpath = config.datapath + root_to_folder + train_dir
if os.path.isdir(dpath):
audios = AudioPipeline((root_to_folder + train_dir), n_to_load=n_to_load, highest_freq=highest_freq,
clip_len=clip_len, mat_dirs=mat_dirs, chunks_per_sec=chunks_per_sec,
down_sampling=down_sampling, add_spectra=add_spectra)
audios.create_train_matrix(f_name_out=d_mat_name, mean_std_per_file=mean_std_per_file)
else:
print('both entered paths are invalid. No data loaded')
print('train directory: ', dpath)
print('matrix file: ', fpath)
return
data = load_matrix(root_to_folder + train_dir + '/', d_mat_name)
else:
data = load_matrix(root_to_folder + train_dir + '/', matrix_file)
x_data = data['x_data']
y_data = data['y_data']
num_frequency_dimensions = x_data.shape[2]
# create model
if architecture == '1':
model = create_lstm_network(num_frequency_dimensions, n_hid_neurons, n_rec_layers, l_activation=activation)
elif architecture == '2':
x_data = np.reshape(x_data, (x_data.shape[0], x_data.shape[1], x_data.shape[2], 1))
model = create_conv_lstm_network(num_frequency_dimensions, n_hid_neurons, n_rec_layers)
model.summary()
print('Start Training')
model.fit(x_data, y_data, batch_size=batch_size, nb_epoch=epochs, verbose=1, validation_split=0.0)
print('Training complete')
w_mat_name = d_mat_name + '_m' + str(architecture) + '_' + str(n_hid_neurons) + 'hid_'
w_mat_name += str(n_rec_layers) + 'lyrs_' + str(epochs) + 'ep_' + str(activation) + 'act'
if save_weights:
json_string = model.to_json()
model_output = config.datapath + '/weight_matrices/' + w_mat_name + '_model.json'
fout = open(model_output, 'w')
json.dump(json_string, fout)
fout.close()
print('saved model to: ' + model_output)
weights_output = config.datapath + '/weight_matrices/' + w_mat_name + '_weights.h5'
model.save_weights(weights_output, overwrite=True)
print('saved weights to: ' + weights_output)
return model, w_mat_name, d_mat_name
from __future__ import print_function
from lstm_utils import create_lstm_network
from audio_preprocessing.pipeline import load_matrix, AudioPipeline
from audio_preprocessing.cconfig import config
data = 'cello_arco_train'
folder_spec = 'cello_arco_train/'
data = 'train_flute'
folder_spec = 'D - data_flute_vib/'
max_instru_freq = 8000 # cello
max_instru_freq = 5000
# get trainings data
myAudios = AudioPipeline(folder_spec, 90, max_instru_freq, 2, chunks_per_sec=4)
myAudios.create_train_matrix(data)
x_data, y_data = load_matrix(folder_spec, data)
num_frequency_dimensions = x_data.shape[2]
num_hidden_dimensions = 1024
batch_size = 10
epochs = 100
model = create_lstm_network(num_frequency_dimensions, num_hidden_dimensions)
print(model.summary())
print('Start Training')
model.fit(x_data,
y_data,
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_split=0.0)