def test_Encoder_forward_backward(
input_layer,
positionwise_layer_type,
interctc_layer_idx,
interctc_use_conditioning,
):
encoder = TransformerEncoder(
20,
output_size=40,
input_layer=input_layer,
positionwise_layer_type=positionwise_layer_type,
interctc_layer_idx=interctc_layer_idx,
interctc_use_conditioning=interctc_use_conditioning,
)
if input_layer == "embed":
x = torch.randint(0, 10, [2, 10])
else:
x = torch.randn(2, 10, 20, requires_grad=True)
x_lens = torch.LongTensor([10, 8])
if len(interctc_layer_idx) > 0:
ctc = None
if interctc_use_conditioning:
vocab_size = 5
output_size = encoder.output_size()
ctc = CTC(odim=vocab_size, encoder_output_size=output_size)
encoder.conditioning_layer = torch.nn.Linear(
vocab_size, output_size)
y, _, _ = encoder(x, x_lens, ctc=ctc)
y = y[0]
else:
y, _, _ = encoder(x, x_lens)
y.sum().backward()
def test_Encoder_output_size():
encoder = TransformerEncoder(20, output_size=256)
assert encoder.output_size() == 256
hop_length=16,
n_mels=32,
)
diar_encoder = TransformerEncoder(
input_layer="linear",
num_blocks=1,
linear_units=32,
output_size=16,
attention_heads=2,
input_size=tcn_separator.output_dim + diar_frontend.output_size(),
)
diar_decoder = LinearDecoder(
num_spk=2,
encoder_output_size=diar_encoder.output_size(),
)
@pytest.mark.parametrize("label_aggregator", [label_aggregator])
@pytest.mark.parametrize("enh_encoder, enh_decoder",
[(enh_encoder, enh_decoder)])
@pytest.mark.parametrize("enh_separator", [tcn_separator])
@pytest.mark.parametrize("mask_module", [mask_module])
@pytest.mark.parametrize("training", [True, False])
@pytest.mark.parametrize("loss_wrappers", [[fix_order_solver]])
@pytest.mark.parametrize("diar_frontend", [diar_frontend])
@pytest.mark.parametrize("diar_encoder, diar_decoder",
[(diar_encoder, diar_decoder)])
def test_enh_diar_model(
enh_encoder,
hop_length=16,
n_mels=10,
)
encoder = TransformerEncoder(
input_size=10,
input_layer="linear",
num_blocks=1,
linear_units=32,
output_size=16,
attention_heads=2,
)
decoder = LinearDecoder(
num_spk=2,
encoder_output_size=encoder.output_size(),
)
rnn_attractor = RnnAttractor(unit=16,
encoder_output_size=encoder.output_size())
label_aggregator = LabelAggregate(
win_length=32,
hop_length=16,
)
@pytest.mark.parametrize(
"frontend, encoder, decoder, label_aggregator",
[(frontend, encoder, decoder, label_aggregator)],
)