def test_tf_mask_net_consistency(
n_fft,
win_length,
hop_length,
rnn_type,
layer,
unit,
dropout,
num_spk,
none_linear,
utt_mvn,
):
model = TFMaskingNet(
n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
rnn_type=rnn_type,
layer=layer,
unit=unit,
dropout=dropout,
num_spk=num_spk,
none_linear=none_linear,
utt_mvn=utt_mvn,
)
model.eval()
random_input_numpy = np.random.randn(2, 16) # np.float64
random_input_torch = (
torch.from_numpy(random_input_numpy - 1.0).float() + 1.0
) # torch.float32
random_input_numpy = torch.from_numpy(
random_input_numpy.astype("float32")
) # np.float64-->np.float32-->torch.float32
est_speech_numpy, flens, masks = model(
random_input_numpy, ilens=torch.LongTensor([16, 12])
)
est_speech_torch, flens, masks = model(
random_input_torch, ilens=torch.LongTensor([16, 12])
)
assert (est_speech_torch[0] - est_speech_numpy[0]).abs().mean() < 1e-5
assert (
np.abs(
(est_speech_torch[-1] - est_speech_numpy[-1]).detach().real.numpy()
).mean()
< 1e-5
)
def test_tf_mask_net_forward_backward(
n_fft,
win_length,
hop_length,
rnn_type,
layer,
unit,
dropout,
num_spk,
nonlinear,
utt_mvn,
mask_type,
loss_type,
):
model = TFMaskingNet(
n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
rnn_type=rnn_type,
layer=layer,
unit=unit,
dropout=dropout,
num_spk=num_spk,
nonlinear=nonlinear,
utt_mvn=utt_mvn,
mask_type=mask_type,
loss_type=loss_type,
)
model.train()
if loss_type.startswith("mask"):
# mask backward
est_speech, flens, masks = model(torch.randn(2, 16,
requires_grad=True),
ilens=torch.LongTensor([16, 12]))
loss = sum([masks[key].mean() for key in masks])
loss.backward()
else:
# spectrums backward
est_speech, flens, masks = model(torch.randn(2, 16,
requires_grad=True),
ilens=torch.LongTensor([16, 12]))
loss = sum([abs(est).mean() for est in est_speech])
loss.backward()
def test_tf_mask_net_output():
inputs = torch.randn(2, 16)
ilens = torch.LongTensor([16, 12])
for num_spk in range(1, 3):
model = TFMaskingNet(
n_fft=8,
win_length=None,
hop_length=2,
rnn_type="blstm",
layer=3,
unit=8,
dropout=0.0,
num_spk=num_spk,
)
model.eval()
specs, _, masks = model(inputs, ilens)
assert isinstance(specs, list)
assert isinstance(masks, dict)
for n in range(num_spk):
assert "spk{}".format(n + 1) in masks
assert specs[n].shape == masks["spk{}".format(n + 1)].shape
def test_tf_mask_net_invalid_norm_type():
with pytest.raises(ValueError):
TFMaskingNet(
n_fft=8,
win_length=None,
hop_length=2,
rnn_type="blstm",
layer=3,
unit=8,
dropout=0.0,
num_spk=2,
nonlinear="fff",
)
def test_tf_mask_net_invalid_loss_type():
with pytest.raises(ValueError):
TFMaskingNet(loss_type="fff")