def __init__(self,
hparams,
conv_filters,
rnn_dim,
bias=False,
conv_act_func=torch.relu,
out_activation_fn=torch.tanh,
drop_rate=0.):
super(ReferenceEncoder, self).__init__()
# ref_enc_filters
channels = [1] + conv_filters + [rnn_dim]
self.convs = nn.ModuleList([
mm.Conv2d(channels[c],
channels[c + 1],
3,
stride=2,
bn=True,
bias=bias,
activation_fn=conv_act_func,
drop_rate=drop_rate) for c in range(len(channels) - 1)
]) # [B, dec_T/r, 128]
self.gru = nn.GRU(rnn_dim * 2, rnn_dim, batch_first=True)
self.fc = nn.Sequential(nn.Linear(rnn_dim, rnn_dim), )
self.activation_fn = out_activation_fn
def __init__(self, in_channels=1, embed_size=128, activation_fn=None):
super(ReferenceEncoder, self).__init__()
self.embed_size = embed_size
channels = [in_channels, 32, 32, 64, 64, 128, embed_size]
self.convs = nn.ModuleList([
mm.Conv2d(channels[c],
channels[c + 1],
3,
stride=2,
bn=True,
bias=False,
activation_fn=torch.relu)
for c in range(len(channels) - 1)
]) # (N, Ty/r, 128)
self.gru = nn.GRU(self.embed_size * 2,
self.embed_size,
batch_first=True)
self.fc = nn.Sequential(nn.Linear(embed_size, embed_size), )
self.activation_fn = activation_fn
def __init__(self, hparams, activation_fn=None):
super(ReferenceEncoder, self).__init__()
self.token_embedding_size = hparams.token_embedding_size
self.in_channels = hparams.n_frames_per_step
# ref_enc_filters
channels = [self.in_channels
] + hparams.ref_enc_filters + [self.token_embedding_size]
self.convs = nn.ModuleList([
mm.Conv2d(channels[c],
channels[c + 1],
3,
stride=2,
bn=True,
bias=False,
activation_fn=torch.relu)
for c in range(len(channels) - 1)
]) # (Batch, Time_domain/r, 128)
self.gru = nn.GRU(self.token_embedding_size * 2,
self.token_embedding_size,
batch_first=True)
self.fc = nn.Sequential(
nn.Linear(self.token_embedding_size, self.token_embedding_size), )
self.activation_fn = activation_fn