def __call__(self, s_signals, s_dropout_keep=1.):
with tf.variable_scope(self.name):
s_mid = ops.lyr_linear(
'linear0', s_signals, hparams.FFT_SIZE*2, axis=-1)
s_mid = ops.relu(s_mid, hparams.RELU_LEAKAGE)
s_out = ops.lyr_linear(
'linear1', s_mid,
hparams.FEATURE_SIZE * hparams.EMBED_SIZE, axis=-1)
s_out = tf.reshape(
s_out,
[hparams.BATCH_SIZE,
-1, hparams.FEATURE_SIZE, hparams.EMBED_SIZE])
return s_out
def __call__(self, s_signals, s_dropout_keep=1.):
with tf.variable_scope(self.name):
s_signals = s_signals - tf.reduce_mean(
s_signals, axis=(1,2), keep_dims=True)
hdim = 300
init_range = .75 / sqrt(hdim)
w_initer = tf.random_uniform_initializer(
-init_range, init_range, dtype=hparams.FLOATX)
b_init_value = np.zeros([hdim*4], dtype=hparams.FLOATX)
b_init_value[hdim*1:hdim*2] = 1.5 # input gate
b_init_value[hdim*2:hdim*3] = -1. # forget gate
b_init_value[hdim*3:hdim*4] = 1. # output gate
b_initer = tf.constant_initializer(b_init_value, dtype=hparams.FLOATX)
s_mid0 = _lyr_bilstm(
'lstm0', self.model,
s_signals, hdim,
-2, -1,
w_initer, b_initer, s_dropout_keep)
s_mid1 = _lyr_bilstm(
'lstm1', self.model,
s_mid0, hdim,
-2, -1,
w_initer, b_initer, s_dropout_keep)
s_mid2 = _lyr_bilstm(
'lstm2', self.model,
s_mid1, hdim,
-2, -1,
w_initer, b_initer, s_dropout_keep)
s_out = _lyr_bilstm(
'lstm3', self.model,
s_mid2, hdim,
-2, -1,
w_initer, b_initer, s_dropout_keep)
s_out = s_out - tf.reduce_mean(
s_out, axis=(1,2), keep_dims=True)
# init_range = 2. / sqrt(300)
init_range = 1.85
s_out = ops.lyr_linear(
'output',
s_out,
hparams.FEATURE_SIZE * hparams.EMBED_SIZE,
w_init=tf.random_uniform_initializer(
-init_range, init_range, dtype=hparams.FLOATX),
bias=None)
s_out = tf.reshape(
s_out, [
hparams.BATCH_SIZE, -1,
hparams.FEATURE_SIZE, hparams.EMBED_SIZE])
return s_out