def test_attention_location():
encoder_out = Variable(torch.zeros(152, 2, 256)) # seq, batch, dim
query_vector = Variable(torch.zeros(1, 2, 256)) # seq, batch, dim
mask = Variable(torch.zeros(2, 152, 1)) # seq, batch, dim
attention = LocationAttention(dim=256)
context, mask = attention(query_vector, encoder_out, mask)
assert context.size() == (1, 2, 256)
assert mask.size() == (2, 152, 1) # batch, input_seq_len
def test_attention_location_softmax():
encoder_out = Variable(torch.randn(152, 2, 256)) # seq, batch, dim
query_vector = Variable(torch.randn(1, 2, 1024)) # seq, batch, dim
mask = Variable(torch.randn(2, 152, 1)) # batch, seq1, seq2
attention = LocationAttention(encoded_dim=256,
query_dim=1024,
attention_dim=128)
context, mask = attention(query_vector, encoder_out, mask)
assert float(mask[:, 0, :].sum().data) == 1.0 # batch, input_seq_len
def test_attention_location_sizes():
encoder_out = Variable(torch.randn(152, 2, 256)) # seq, batch, dim
query_vector = Variable(torch.randn(1, 2, 1024)) # seq, batch, dim
mask = Variable(torch.randn(2, 152, 1)) # batch, seq1, seq2
attention = LocationAttention(encoded_dim=256,
query_dim=1024,
attention_dim=128)
context, mask = attention(query_vector, encoder_out, mask)
assert context.size() == (1, 2, 256)
assert mask.size() == (1, 2, 152) # seq2, batch, seq1
def __init__(self, hidden_size=1024, num_layers=2):
super(Decoder, self).__init__()
self.prenet = PreNet(in_features=80, out_features=256)
self.attention = LocationAttention(dim=256)
self.rnn = nn.GRU(input_size=512,
hidden_size=hidden_size,
num_layers=num_layers,
dropout=0.1)
self.spec_out = nn.Linear(in_features=1024 + 256, out_features=80)
self.stop_out = nn.Linear(in_features=1024 + 256, out_features=1)
self.postnet = PostNet()
def test_attention():
"""
Attention should output a fixed length context vector (seq len = 1)
and and a weight for each item in the input sequence
"""
encoder_out = Variable(torch.zeros(152, 2, 256)) # seq, batch, dim
query_vector = Variable(torch.zeros(1, 2, 256)) # seq, batch, dim
attention = LocationAttention(dim=256)
context, mask = attention(query_vector, encoder_out)
assert context.size() == (1, 2, 256) # seq, batch, dim
assert mask.size() == (2, 152, 1) # batch, input_seq_len
def test_attention_sizes():
"""
Attention should output a fixed length context vector (seq len = 1)
and and a weight for each item in the input sequence
"""
encoder_out = Variable(torch.randn(152, 2, 256)) # seq, batch, dim
query_vector = Variable(torch.randn(1, 2, 1024)) # seq, batch, dim
attention = LocationAttention(encoded_dim=256,
query_dim=1024,
attention_dim=128)
context, mask = attention(query_vector, encoder_out)
assert context.size() == (1, 2, 256) # seq, batch, dim
assert mask.size() == (1, 2, 152) # seq2, batch, seq1
def __init__(self,
hidden_size=1024,
num_layers=2,
num_mels=80,
num_prenet_features=256):
super(Decoder, self).__init__()
self.num_layers = num_layers
self.hidden_size = hidden_size
self.num_mels = num_mels
self.prenet = PreNet(in_features=num_mels,
out_features=num_prenet_features)
self.attention = LocationAttention(encoded_dim=256,
query_dim=hidden_size,
attention_dim=128)
self.rnn = nn.LSTM(input_size=num_prenet_features + 256,
hidden_size=hidden_size,
num_layers=num_layers,
dropout=0.1)
self.spec_out = nn.Linear(in_features=hidden_size + 256,
out_features=num_mels)
self.stop_out = nn.Linear(in_features=hidden_size + 256,
out_features=1)
self.postnet = PostNet()