some user reports suggestion for better accuracy and less params

[WenchenLi]

model.py
line 119：
embedded_f0_channel = tf.reshape(embedded_chars_expanded,
[-1, self.f0, self.s, self.embedding_size])
changed to：
embedded_f0_channel = tf.reshape(embedded_chars_expanded,
[-1, self.s, self.embedding_size, self.f0])

line 157：
 multiplier = max_pool.get_shape()[2].value
changed to:
 multiplier = max_pool.get_shape()[1].value

for better accuracy and less params