ppg_tacotron/model/Net2.py

85 lines
3.5 KiB
Python

import torch
from torch.nn import Module, Linear
from .modules import PreNet, CBHG
import hparams
class Net2(Module):
def __init__(self, in_dims, hidden_units, dropout_rate, num_conv1d_banks, num_highway_blocks):
super(Net2, self).__init__()
# in_dims = phones_len, out_dims_1 = 2*out_dims_2 = net2_hidden_units
self.pre_net = PreNet(in_dims=in_dims,
out_dims_1=hidden_units,
dropout_rate=dropout_rate)
# num_conv1d_banks = net2_num_conv1d_banks, num_highway_blocks = net2_num_highway_blocks
# in_dims = net2_hidden_units // 2, out_dims = net2_hidden_units // 2
# activation=torch.nn.ReLU()
self.cbhg_mel = CBHG(num_conv1d_banks=num_conv1d_banks,
num_highway_blocks=num_highway_blocks,
in_dims=hidden_units // 2,
out_dims=hidden_units // 2,
activation=torch.nn.ReLU())
# num_conv1d_banks = net2_num_conv1d_banks, num_highway_blocks = net2_num_highway_blocks
# in_dims = net2_hidden_units // 2, out_dims = net2_hidden_units // 2
# activation=torch.nn.ReLU()
self.cbhg_spec = CBHG(num_conv1d_banks=num_conv1d_banks,
num_highway_blocks=num_highway_blocks,
in_dims=hidden_units // 2,
out_dims=hidden_units // 2,
activation=torch.nn.ReLU())
self.pred_mel = Linear(in_features=hidden_units, out_features=hparams.timit_n_mels)
self.prepare_spec = Linear(in_features=hparams.timit_n_mels, out_features=hidden_units // 2)
self.pred_spec = Linear(in_features=hidden_units, out_features=hparams.timit_n_fft // 2 + 1)
def forward(self, inputs):
# inputs : (N, L_in, in_dims)
# in_dims = phns_len
# PreNet
pre_net_outputs = self.pre_net(inputs)
# pre_net_outputs : (N, L_in, net2_hidden_units // 2)
# Change data format
cbhg_mel_inputs = pre_net_outputs.transpose(1, 2)
# cbhg_mel_inputs : (N, net2_hidden_units // 2, L_in)
# CBHG : mel-scale
cbhg_mel_outputs = self.cbhg_mel(cbhg_mel_inputs)
# cbhg_mel_outputs : (N, L_in, net2_hidden_units)
# Pred mel
pred_mel = self.pred_mel(cbhg_mel_outputs)
# pred_mel : (N, L_in, n_mels)
# Change data format
cbhg_spec_inputs = self.prepare_spec(pred_mel)
# cbhg_spec_inputs : (N, L_in, net2_hidden_units // 2)
cbhg_spec_inputs = cbhg_spec_inputs.transpose(1, 2)
# cbhg_spec_inputs : (N, net2_hidden_units // 2, L_in)
# Pred spec
cbhg_spec_outputs = self.cbhg_spec(cbhg_spec_inputs)
# cbhg_spec_outputs : (N, L_in, net2_hidden_units)
# Pred spec
pred_spec = self.pred_spec(cbhg_spec_outputs)
# pred_spec : (N, L_in, n_fft//2 + 1)
debug = False
if debug:
print("inputs.shape : " + str(inputs.shape))
print("pre_net_outputs.shape : " + str(pre_net_outputs.shape))
print("cbhg_mel_inputs.shape : " + str(cbhg_mel_inputs.shape))
print("cbhg_mel_outputs.shape : " + str(cbhg_mel_outputs.shape))
print("pred_mel.shape : " + str(pred_mel.shape))
print("cbhg_spec_inputs.shape : " + str(cbhg_spec_inputs.shape))
print("cbhg_spec_outputs.shape : " + str(cbhg_spec_outputs.shape))
print("pred_spec.shape : " + str(pred_spec.shape))
return pred_spec, pred_mel