import torch from torch import nn from torch.nn import functional as F from nn_layers import convolutional_module torch.manual_seed(1234) class Postnet(nn.Module): """Postnet - Five 1-d convolution with 512 channels and kernel size 5 """ def __init__(self, tacotron_hyperparams): super(Postnet, self).__init__() # self.dropout = nn.Dropout(0.5) self.convolutions = nn.ModuleList() self.convolutions.append( nn.Sequential( convolutional_module(tacotron_hyperparams['n_mel_channels'], tacotron_hyperparams['postnet_embedding_dim'], kernel_size=tacotron_hyperparams['postnet_kernel_size'], stride=1, padding=int((tacotron_hyperparams['postnet_kernel_size'] - 1) / 2), dilation=1, w_init_gain='tanh'), nn.BatchNorm1d(tacotron_hyperparams['postnet_embedding_dim'])) ) for i in range(1, tacotron_hyperparams['postnet_n_convolutions'] - 1): self.convolutions.append( nn.Sequential( convolutional_module(tacotron_hyperparams['postnet_embedding_dim'], tacotron_hyperparams['postnet_embedding_dim'], kernel_size=tacotron_hyperparams['postnet_kernel_size'], stride=1, padding=int((tacotron_hyperparams['postnet_kernel_size'] - 1) / 2), dilation=1, w_init_gain='tanh'), nn.BatchNorm1d(tacotron_hyperparams['postnet_embedding_dim'])) ) self.convolutions.append( nn.Sequential( convolutional_module(tacotron_hyperparams['postnet_embedding_dim'], tacotron_hyperparams['n_mel_channels'], kernel_size=tacotron_hyperparams['postnet_kernel_size'], stride=1, padding=int((tacotron_hyperparams['postnet_kernel_size'] - 1) / 2), dilation=1, w_init_gain='linear'), nn.BatchNorm1d(tacotron_hyperparams['n_mel_channels'])) ) def forward(self, x): for i in range(len(self.convolutions) - 1): x = F.dropout(torch.tanh(self.convolutions[i](x)), 0.5, self.training) x = F.dropout(self.convolutions[-1](x), 0.5, self.training) return x