def train():
# Model
model = Model(ModelConfig.HIDDEN_LAYERS, ModelConfig.HIDDEN_UNITS)
# Loss, Optimizer
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
loss_fn = model.loss()
optimizer = tf.train.AdamOptimizer(learning_rate=TrainConfig.LR).minimize(
loss_fn, global_step=global_step)
#optimizer = tf.train.GradientDescentOptimizer(learning_rate=TrainConfig.LR).minimize(loss_fn, global_step=global_step)
model.gnsdr_music = tf.placeholder(dtype=tf.float32,
shape=(),
name='gnsdr_music')
model.gsir_music = tf.placeholder(dtype=tf.float32,
shape=(),
name='gsir_music')
model.gsar_music = tf.placeholder(dtype=tf.float32,
shape=(),
name='gsar_music')
model.gnsdr_vocal = tf.placeholder(dtype=tf.float32,
shape=(),
name='gnsdr_vocal')
model.gsir_vocal = tf.placeholder(dtype=tf.float32,
shape=(),
name='gsir_vocal')
model.gsar_vocal = tf.placeholder(dtype=tf.float32,
shape=(),
name='gsar_vocal')
# Summaries
summary_ops = summaries(model, loss_fn)
with tf.Session(config=TrainConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, TrainConfig.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(TrainConfig.GRAPH_PATH, sess.graph)
# Input source
data = Data(TrainConfig.DATA_PATH)
eval_data = Data(EvalConfig.DATA_PATH)
loss = Diff()
gnsdr, gsir, gsar = np.array([0, 0]), np.array([0,
0]), np.array([0, 0])
intial_global_step = global_step.eval()
for step in range(intial_global_step, TrainConfig.FINAL_STEP):
start_time = time.time()
bss_metric = step % 20 == 0 or step == intial_global_step
bss_eval = ''
if bss_metric:
gnsdr, gsir, gsar = eval(model, eval_data, sess)
bss_eval = 'GNSDR: {} GSIR: {} GSAR: {}'.format(
gnsdr, gsir, gsar)
mixed_wav, src1_wav, src2_wav, _ = data.next_wavs(
TrainConfig.SECONDS, TrainConfig.NUM_WAVFILE)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
src1_spec, src2_spec = to_spectrogram(src1_wav), to_spectrogram(
src2_wav)
src1_mag, src2_mag = get_magnitude(src1_spec), get_magnitude(
src2_spec)
src1_batch, _ = model.spec_to_batch(src1_mag)
src2_batch, _ = model.spec_to_batch(src2_mag)
mixed_batch, _ = model.spec_to_batch(mixed_mag)
l, _, summary = sess.run(
[loss_fn, optimizer, summary_ops],
feed_dict={
model.x_mixed: mixed_batch,
model.y_src1: src1_batch,
model.y_src2: src2_batch,
model.gnsdr_music: gnsdr[0],
model.gsir_music: gsir[0],
model.gsar_music: gsar[0],
model.gnsdr_vocal: gnsdr[1],
model.gsir_vocal: gsir[1],
model.gsar_vocal: gsar[1]
})
loss.update(l)
writer.add_summary(summary, global_step=step)
# Save state
if step % TrainConfig.CKPT_STEP == 0:
tf.train.Saver().save(sess,
TrainConfig.CKPT_PATH + '/checkpoint',
global_step=step)
elapsed_time = time.time() - start_time
print(
'step-{}\ttime={:2.2f}\td_loss={:2.2f}\tloss={:2.3f}\tbss_eval: {}'
.format(step, elapsed_time, loss.diff * 100, loss.value,
bss_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 train():
# Model
model = Model()
# Loss, Optimizer
global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')
loss_fn = model.loss()
optimizer = tf.train.AdamOptimizer(learning_rate=TrainConfig.LR).minimize(
loss_fn, global_step=global_step)
# Summaries
summary_op = summaries(model, loss_fn)
with tf.Session(config=TrainConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, TrainConfig.CKPT_PATH)
writer = tf.summary.FileWriter(TrainConfig.GRAPH_PATH, sess.graph)
# Input source
data = Data(TrainConfig.DATA_PATH)
loss = Diff()
for step in range(global_step.eval(), TrainConfig.FINAL_STEP):
mixed_wav, src1_wav, src2_wav, _ = data.next_wavs(
TrainConfig.SECONDS, TrainConfig.NUM_WAVFILE)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
src1_spec, src2_spec = to_spectrogram(src1_wav), to_spectrogram(
src2_wav)
src1_mag, src2_mag = get_magnitude(src1_spec), get_magnitude(
src2_spec)
src1_batch, _ = model.spec_to_batch(src1_mag)
src2_batch, _ = model.spec_to_batch(src2_mag)
mixed_batch, _ = model.spec_to_batch(mixed_mag)
l, _, summary = sess.run(
[loss_fn, optimizer, summary_op],
feed_dict={
model.x_mixed: mixed_batch,
model.y_src1: src1_batch,
model.y_src2: src2_batch
})
loss.update(l)
print('step-{}\td_loss={:2.2f}\tloss={}'.format(
step, loss.diff * 100, loss.value))
writer.add_summary(summary, global_step=step)
# Save state
if step % TrainConfig.CKPT_STEP == 0:
tf.train.Saver().save(sess,
TrainConfig.CKPT_PATH + '/checkpoint',
global_step=step)
writer.close()
from data import Data
from librosa import amplitude_to_db, stft
from librosa.display import specshow
from preprocess import to_spectrogram, get_magnitude
from pylab import savefig
import matplotlib.pyplot as plt
import numpy as np
data = Data(TrainConfig.DATA_PATH)
mixed_wav, src1_wav, src2_wav, _ = data.next_wavs(TrainConfig.SECONDS,
TrainConfig.NUM_WAVFILE)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
src1_spec, src2_spec = to_spectrogram(src1_wav), to_spectrogram(src2_wav)
src1_mag, src2_mag = get_magnitude(src1_spec), get_magnitude(src2_spec)
sr = ModelConfig.SR
y = src1_wav[0]
def plot_wav_as_spec(wav, sr=ModelConfig.SR, s=0.5, path='foo.png'):
"""Plots a spectrogram
Will save as foo.png in script directory
Arguments:
wav {array} -- audio data
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()
def train():
dsd_train, dsd_test = GetData.getDSDFilelist("DSD100.xml")
dataset = dict()
dataset[
"train_sup"] = dsd_train # 50 training tracks from DSD100 as supervised dataset
dataset["valid"] = dsd_test[:25]
dataset["test"] = dsd_test[25:]
with open('dataset.pkl', 'wb') as file:
pickle.dump(dataset, file)
print("Created dataset structure")
# Model
model = Model()
# Loss, Optimizer
#global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
loss_fn = model.loss()
lr = ((CONFIG_MAP['flat-R-VAE'].hparams.learning_rate -
CONFIG_MAP['flat-R-VAE'].hparams.min_learning_rate) *
tf.pow(CONFIG_MAP['flat-R-VAE'].hparams.decay_rate,
tf.to_float(global_step)) +
CONFIG_MAP['flat-R-VAE'].hparams.min_learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(
loss_fn, global_step=model.global_step)
# Summaries
summary_op = summaries(model, loss_fn)
with tf.Session(config=TrainConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
#model.load_state(sess, TrainConfig.CKPT_PATH)
writer = tf.summary.FileWriter(TrainConfig.GRAPH_PATH, sess.graph)
# Input source
btch_size = CONFIG_MAP['flat-R-VAE'].hparams.batch_size
loss = Diff()
i = 0
for step in range(
model.global_step.eval(),
TrainConfig.FINAL_STEP): # changed xrange to range for py3
if (i > 50):
i = 0
batch_ = dsd_train[i:i + btch_size]
i = i + btch_size
mixed_wav, drums_wav = get_random_wav(batch_, TrainConfig.SECONDS,
ModelConfig.SR)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
drums_spec = to_spectrogram(drums_wav)
drums_mag = get_magnitude(drums_spec)
mixed_batch, _ = model.spec_to_batch(mixed_mag)
drums_batch, _ = model.spec_to_batch(drums_mag)
l, _, summary = sess.run(
[loss_fn, optimizer, summary_op],
feed_dict={
model.x_mixed: mixed_batch,
model.x_drums: drums_batch,
model.y_drums: drums_batch
})
loss.update(l)
print('step-{}\td_loss={:2.2f}\tloss={}'.format(
step, loss.diff * 100, loss.value))
writer.add_summary(summary, global_step=step)
# Save state
if step % TrainConfig.CKPT_STEP == 0:
tf.train.Saver().save(sess,
TrainConfig.CKPT_PATH + '/checkpoint',
global_step=step)
writer.close()
def separate(filename, channel):
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')
total_samples, origin_samples, samples = decode_input(filename)
channels = origin_samples.shape[0]
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)
mixed_wav, src1_wav, src2_wav = samples, samples, samples
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,
1)[:, :, :seq_len]
pred_src2_mag = model.batch_to_spec(pred_src2_mag,
1)[:, :, :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)
def stack(data):
size = data.shape[0] // channels
elements = []
for i in range(channels):
elements.append(data[size * i:size * (i + 1)])
return np.dstack(elements)[0]
music_data = pred_src1_wav
voice_data = pred_src2_wav
if channel >= 0:
def filter_samples(data):
for i in range(origin_samples.shape[0]):
if i != channel:
data[i, :] = origin_samples[i, 0:data.shape[1]]
return data
music_data = filter_samples(music_data)
voice_data = filter_samples(voice_data)
music_wav = np.dstack(music_data)[0]
voice_wav = np.dstack(voice_data)[0]
return music_wav, voice_wav
return None