wrong size of conv1d in CBHG, Post-processing net

[msobhan69]

See https://github.com/Kyubyong/tacotron/blob/master/networks.py#L94 , according to Table 1 of the paper, the size of 2'th layer of Conv1D projections should be 80.

[tozangezan]

Here. (networks.py; L112~)

dec = conv1d(dec, hp.embed_size//2, 3, scope="conv1d_1") # (N, T', E) dec = normalize(dec, type="bn", is_training=is_training, activation_fn=tf.nn.relu) dec = conv1d(dec, hp.embed_size//2, 3, scope="conv1d_2") # (N, T', E) I think first method should be conv1d(dec, hp.embed_size, 3, scope="conv1d_1") and the second should be conv1d(dec, 80, 3, scope="conv1d_1")

[Kyubyong]

You guys are right. I've changed. Thanks.

[candlewill]

@Kyubyong The dimension is updated. However, this change leads to a new issue in highway net:

outputs = H * T + inputs * C

The dimension of inputs and C does not match. The inputs tensor has shape [N, T, W], W=80, while the C tensor has shape [N, T, num_units], num_units=128.

My suggestion is to add a dense layer before highway layer. This dense layer transforms the inputs into a tensor of shape [N, T, num_units].

[Kyubyong]

You're right candlewill. But I don't see any particular reason why we should make things complicated, so I'll just change the output units of the second conv1d layer to 128 for simpliicity.