def extract(filename, channel):
music_wav, voice_wav = separate(filename, channel)
base_file_name = os.path.splitext(filename)[0]
write_wav(music_wav,
base_file_name + '-h%d-music' % ModelConfig.HIDDEN_UNITS)
write_wav(voice_wav,
base_file_name + '-h%d-voice' % ModelConfig.HIDDEN_UNITS)
def extract(filename, acc_path, voice_path):
if not acc_path and not voice_path:
return
music_wav, voice_wav = separate(filename, -1)
if acc_path:
write_wav(music_wav, os.path.splitext(acc_path)[0])
if voice_path:
write_wav(voice_wav, os.path.splitext(voice_path)[0])
def get_drum(filename):
tf.reset_default_graph()
model = Model()
with tf.Session(config=EvalConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
data = Datas(RunConfig.DATA_ROOT)
model.load_state(sess, EvalConfig.CKPT_PATH)
mixed_wav = data.get_mixture(filename)
print(mixed_wav)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
(pred_src1_mag,
pred_src2_mag) = sess.run(model(),
feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(
pred_src1_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(
pred_src2_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src2_mag = mixed_mag * mask_src2
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase)
filename = filename.replace('.wav', '')
write_wav(pred_src2_wav[0], '{}/{}'.format(RunConfig.RESULT_PATH,
filename))
def test_run():
tf.reset_default_graph()
model = Model()
with tf.Session(config=EvalConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
data = Datas(EvalConfig.DATA_PATH)
model.load_state(sess, EvalConfig.CKPT_PATH)
mixed_wav, src1_wav, src2_wav = data.next_wav(EvalConfig.SECONDS)
print(src1_wav)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
(pred_src1_mag,
pred_src2_mag) = sess.run(model(),
feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(
pred_src1_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(
pred_src2_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src1_mag = mixed_mag * mask_src1
pred_src2_mag = mixed_mag * mask_src2
pred_src1_wav = to_wav(pred_src1_mag, mixed_phase)
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase)
write_wav(mixed_wav[0], '{}/{}'.format(EvalConfig.RESULT_PATH,
'original'))
write_wav(pred_src1_wav[0], '{}/{}'.format(EvalConfig.RESULT_PATH,
'music'))
write_wav(pred_src2_wav[0], '{}/{}'.format(EvalConfig.RESULT_PATH,
'voice'))
def eval(data_path=None, result_path=None):
# Model
model = Model()
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
with tf.Session(config=EvalConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, EvalConfig.CKPT_PATH)
writer = tf.summary.FileWriter(EvalConfig.GRAPH_PATH, sess.graph)
data = Data(data_path) if data_path else Data(EvalConfig.DATA_PATH)
output_path = result_path if result_path else EvalConfig.RESULT_PATH
mixed_wav, src1_wav, src2_wav, wavfiles = data.next_wavs(
EvalConfig.SECONDS, EvalConfig.NUM_EVAL)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
assert (np.all(
np.equal(model.batch_to_spec(mixed_batch, EvalConfig.NUM_EVAL),
padded_mixed_mag)))
(pred_src1_mag,
pred_src2_mag) = sess.run(model(),
feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(
pred_src1_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(
pred_src2_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src1_mag = mixed_mag * mask_src1
pred_src2_mag = mixed_mag * mask_src2
# (magnitude, phase) -> spectrogram -> wav
if EvalConfig.GRIFFIN_LIM:
pred_src1_wav = to_wav_mag_only(
pred_src1_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
pred_src2_wav = to_wav_mag_only(
pred_src2_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
else:
pred_src1_wav = to_wav(pred_src1_mag, mixed_phase)
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase)
# Write the result
tf.summary.audio('GT_mixed',
mixed_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_music',
pred_src1_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_vocal',
pred_src2_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
if EvalConfig.EVAL_METRIC:
# Compute BSS metrics
gnsdr, gsir, gsar = bss_eval_global(mixed_wav, src1_wav, src2_wav,
pred_src1_wav, pred_src2_wav)
# Write the score of BSS metrics
tf.summary.scalar('GNSDR_music', gnsdr[0])
tf.summary.scalar('GSIR_music', gsir[0])
tf.summary.scalar('GSAR_music', gsar[0])
tf.summary.scalar('GNSDR_vocal', gnsdr[1])
tf.summary.scalar('GSIR_vocal', gsir[1])
tf.summary.scalar('GSAR_vocal', gsar[1])
if EvalConfig.WRITE_RESULT:
# Write the result
for i in range(len(wavfiles)):
name = 'video'
print output_path
write_wav(mixed_wav[i],
'{}/{}-{}'.format(output_path, name, 'original'))
write_wav(pred_src1_wav[i],
'{}/{}-{}'.format(output_path, name, 'music'))
write_wav(pred_src2_wav[i],
'{}/{}-{}'.format(output_path, name, 'voice'))
writer.add_summary(sess.run(tf.summary.merge_all()),
global_step=global_step.eval())
writer.close()
def eval():
# Model
model = Model()
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
with tf.Session(config=EvalConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, EvalConfig.CKPT_PATH)
writer = tf.summary.FileWriter(EvalConfig.GRAPH_PATH, sess.graph)
data = Data(EvalConfig.DATA_PATH, TrainConfig.NOISE_DATA_PATH,
TrainConfig.VOICE_DATA_PATH)
mixed_wav, src1_wav, src2_wav, wavfiles = data.next_wavs_eval(
EvalConfig.SECONDS, EvalConfig.NUM_EVAL)
start = time.time()
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
assert (np.all(
np.equal(model.batch_to_spec(mixed_batch, EvalConfig.NUM_EVAL),
padded_mixed_mag)))
(pred_src1_mag,
pred_src2_mag) = sess.run(model(),
feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(
pred_src1_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(
pred_src2_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src1_mag = mixed_mag * mask_src1
pred_src2_mag = mixed_mag * mask_src2
# (magnitude, phase) -> spectrogram -> wav
if EvalConfig.GRIFFIN_LIM:
pred_src1_wav = to_wav_mag_only(
pred_src1_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
pred_src2_wav = to_wav_mag_only(
pred_src2_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
else:
pred_src1_wav = to_wav(pred_src1_mag, mixed_phase)
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase)
end = time.time()
print("Time elapsed: {0}".format(end - start))
# Write the result
tf.summary.audio('GT_mixed',
mixed_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_music',
pred_src1_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_vocal',
pred_src2_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
# Write the result
for i in range(len(wavfiles)):
name = wavfiles[i].replace('/', '-').replace('.wav', '')
write_wav(
mixed_wav[i], '{}/{}-{}'.format(EvalConfig.RESULT_PATH, name,
'original'))
write_wav(
pred_src1_wav[i], '{}/{}-{}'.format(EvalConfig.RESULT_PATH,
name, 'background'))
write_wav(pred_src2_wav[i],
'{}/{}-{}'.format(EvalConfig.RESULT_PATH, name, 'voice'))
writer.add_summary(sess.run(tf.summary.merge_all()),
global_step=global_step.eval())
writer.close()
def eval(n):
overall_gnsdr, overall_gsir, overall_gsar = [], [], []
for i in range(n):
with tf.Graph().as_default():
# Model
model = Model(ModelConfig.HIDDEN_LAYERS, ModelConfig.HIDDEN_UNITS)
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
with tf.Session(config=EvalConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, EvalConfig.CKPT_PATH)
print('num trainable parameters: %s' % (np.sum([
np.prod(v.get_shape().as_list())
for v in tf.trainable_variables()
])))
writer = tf.summary.FileWriter(EvalConfig.GRAPH_PATH,
sess.graph)
data = Data(EvalConfig.DATA_PATH)
mixed_wav, src1_wav, src2_wav, wavfiles = data.next_wavs(
EvalConfig.SECONDS, EvalConfig.NUM_EVAL)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
assert (np.all(
np.equal(
model.batch_to_spec(mixed_batch, EvalConfig.NUM_EVAL),
padded_mixed_mag)))
(pred_src1_mag, pred_src2_mag) = sess.run(
model(), feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(
pred_src1_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(
pred_src2_mag, EvalConfig.NUM_EVAL)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src1_mag = mixed_mag * mask_src1
pred_src2_mag = mixed_mag * mask_src2
# (magnitude, phase) -> spectrogram -> wav
if EvalConfig.GRIFFIN_LIM:
pred_src1_wav = to_wav_mag_only(
pred_src1_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
pred_src2_wav = to_wav_mag_only(
pred_src2_mag,
init_phase=mixed_phase,
num_iters=EvalConfig.GRIFFIN_LIM_ITER)
else:
pred_src1_wav = to_wav(pred_src1_mag, mixed_phase)
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase)
# Write the result
tf.summary.audio('GT_mixed',
mixed_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_music',
pred_src1_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
tf.summary.audio('Pred_vocal',
pred_src2_wav,
ModelConfig.SR,
max_outputs=EvalConfig.NUM_EVAL)
if EvalConfig.EVAL_METRIC:
# Compute BSS metrics
gnsdr, gsir, gsar = bss_eval_global(
mixed_wav, src1_wav, src2_wav, pred_src1_wav,
pred_src2_wav)
# Write the score of BSS metrics
tf.summary.scalar('GNSDR_music', gnsdr[0])
tf.summary.scalar('GSIR_music', gsir[0])
tf.summary.scalar('GSAR_music', gsar[0])
tf.summary.scalar('GNSDR_vocal', gnsdr[1])
tf.summary.scalar('GSIR_vocal', gsir[1])
tf.summary.scalar('GSAR_vocal', gsar[1])
print('GNSDR: ', gnsdr)
print('GSIR: ', gsir)
print('GSAR: ', gsar)
overall_gnsdr.append(gnsdr)
overall_gsir.append(gsir)
overall_gsar.append(gsar)
if EvalConfig.WRITE_RESULT:
# Write the result
for i in range(len(wavfiles)):
name = wavfiles[i].replace('/',
'-').replace('.wav', '')
write_wav(
mixed_wav[i],
'{}/{}-{}'.format(EvalConfig.RESULT_PATH, name,
'original'))
write_wav(
pred_src1_wav[i],
'{}/{}-{}'.format(EvalConfig.RESULT_PATH, name,
'music'))
write_wav(
pred_src2_wav[i],
'{}/{}-{}'.format(EvalConfig.RESULT_PATH, name,
'voice'))
writer.add_summary(sess.run(tf.summary.merge_all()),
global_step=global_step.eval())
writer.close()
if n > 1:
overall_gnsdr = np.array(overall_gnsdr)
overall_gsir = np.array(overall_gsir)
overall_gsar = np.array(overall_gsar)
overall_gnsdr = np.mean(overall_gnsdr, axis=0)
overall_gsir = np.mean(overall_gsir, axis=0)
overall_gsar = np.mean(overall_gsar, axis=0)
print('OVERALL GNSDR: ', overall_gnsdr)
print('OVERALL GSIR: ', overall_gsir)
print('OVERALL GSAR: ', overall_gsar)
def eval(model, data, sr, len_frame, num_wav, glim, glim_iter, ckpt_path,
graph_path, result_path):
len_hop = closest_power_of_two(len_frame / 4)
global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
with tf.Session() as sess:
if not os.path.exists(result_path):
os.makedirs(result_path)
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, ckpt_path)
writer = tf.summary.FileWriter("{}/{}".format(graph_path, "eval"), sess.graph)
mixed_wav, src1_wav, src2_wav, med_names = data.next_wavs(num_wav)
mixed_spec = to_spectrogram(mixed_wav, len_frame, len_hop)
mixed_mag = get_magnitude(mixed_spec)
mixed_batch, padded_mixed_mag = model.spec_to_batch(mixed_mag)
mixed_phase = get_phase(mixed_spec)
assert (np.all(np.equal(model.batch_to_spec(mixed_batch, num_wav),
padded_mixed_mag)))
(pred_src1_mag, pred_src2_mag) = sess.run(model(),
feed_dict={model.x_mixed: mixed_batch})
seq_len = mixed_phase.shape[-1]
pred_src1_mag = model.batch_to_spec(pred_src1_mag, num_wav)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(pred_src2_mag, num_wav)[:, :, :seq_len]
# Time-frequency masking
mask_src1 = soft_time_freq_mask(pred_src1_mag, pred_src2_mag)
# mask_src1 = hard_time_freq_mask(pred_src1_mag, pred_src2_mag)
mask_src2 = 1. - mask_src1
pred_src1_mag = mixed_mag * mask_src1
pred_src2_mag = mixed_mag * mask_src2
# (magnitude, phase) -> spectrogram -> wav
if glim:
pred_src1_wav = to_wav_mag_only(pred_src1_mag, mixed_phase, len_frame,
len_hop, num_iters=glim_iter)
pred_src2_wav = to_wav_mag_only(pred_src2_mag, mixed_phase, len_frame,
len_hop, num_iters=glim_iter)
else:
pred_src1_wav = to_wav(pred_src1_mag, mixed_phase, len_hop)
pred_src2_wav = to_wav(pred_src2_mag, mixed_phase, len_hop)
# Write the result
tf.summary.audio('GT_mixed', mixed_wav, sr, max_outputs=num_wav)
tf.summary.audio('Pred_music', pred_src1_wav, sr, max_outputs=num_wav)
tf.summary.audio('Pred_vocal', pred_src2_wav, sr, max_outputs=num_wav)
# Compute BSS metrics
gnsdr, gsir, gsar = bss_eval_global(mixed_wav, src1_wav, src2_wav, pred_src1_wav,
pred_src2_wav, num_wav)
# Write the score of BSS metrics
tf.summary.scalar('GNSDR_music', gnsdr[0])
tf.summary.scalar('GSIR_music', gsir[0])
tf.summary.scalar('GSAR_music', gsar[0])
tf.summary.scalar('GNSDR_vocal', gnsdr[1])
tf.summary.scalar('GSIR_vocal', gsir[1])
tf.summary.scalar('GSAR_vocal', gsar[1])
# Write the result
for i in range(len(med_names)):
write_wav(mixed_wav[i], '{}/{}-{}'.format(result_path, med_names[i],
'all_stems_mixed'), sr)
write_wav(pred_src1_wav[i], '{}/{}-{}'.format(result_path, med_names[i],
'target_instrument'), sr)
write_wav(pred_src2_wav[i], '{}/{}-{}'.format(result_path, med_names[i],
'other_stems_mixed'), sr)
writer.add_summary(sess.run(tf.summary.merge_all()), global_step=global_step.eval())
writer.close()
