def __init__(self, X, Y, speakers, dropout=0.5, training=True):
self.X = X
self.Y = Y
self.speakers = speakers
stft_X, D_X = tf_featurization.get_stft(self.X, T=512 * 3)
self.stft = []
for i in range(speakers):
self.stft.append(tf_featurization.get_stft(self.Y[i], T=512 * 3))
params = {'conv_n_filters': [48 * (2**i) for i in range(6)]}
logits = unet.Model(
D_X,
cout=speakers,
output_mask_logit=True,
dropout=dropout,
training=training,
params=params,
).logits
logits = tf.nn.sigmoid(logits)
self.outputs = []
for i in range(speakers):
self.outputs.append(logits[:, :, :, i:i + 1] * D_X)
self.loss = []
for i in range(speakers):
self.loss.append(
tf.reduce_mean(tf.abs(self.outputs[i] - self.stft[i][1]),
axis=[1, 2, 3]))
self.cost = tf.reduce_sum(self.loss, axis=-1)
self.cost = tf.reduce_mean(self.cost)
def __init__(self, X, Y):
self.X = X
self.Y = Y
stft_X, D_X = tf_featurization.get_stft(self.X)
self.stft = []
for i in range(len(self.Y)):
self.stft.append(tf_featurization.get_stft(self.Y[i]))
self.outputs = []
for i in range(len(self.Y)):
with tf.variable_scope(f'model_{i}'):
self.outputs.append(unet.Model(D_X, num_layers=6).logits)
self.loss = []
for i in range(len(self.Y)):
self.loss.append(
tf.reduce_mean(tf.abs(self.outputs[i] - self.stft[i][1])))
self.cost = tf.reduce_sum(self.loss)
def body(i, stft, D):
stft_x, D_x = tf_featurization.get_stft(X[i])
return i + 1, stft.write(i, stft_x), D.write(i, D_x)