def test_demask():
model = DeMask()
test_input = torch.randn(1, 801)
model_conf = model.serialize()
reconstructed_model = DeMask.from_pretrained(model_conf)
assert_allclose(model(test_input), reconstructed_model(test_input))
def test_sample_rate():
demask = DeMask(hidden_dims=(16, ),
kernel_size=8,
n_filters=8,
stride=4,
sample_rate=9704)
assert demask.sample_rate == 9704
def test_forward(input_type, output_type, fb_name, data):
demask = DeMask(
input_type=input_type,
output_type=output_type,
fb_name=fb_name,
hidden_dims=(16, ),
kernel_size=8,
n_filters=8,
stride=4,
)
demask = demask.eval()
with torch.no_grad():
demask(data)
def test_enhancement_model(small_model_params, test_data):
params = small_model_params["DeMask"]
filter_bank = "free"
device = get_default_device()
inputs = ((torch.rand(1, 201, device=device) - 0.5) * 2, )
test_data = test_data.to(device)
model = DeMask(**params, fb_type=filter_bank).eval().to(device)
traced = torch.jit.trace(model, inputs)
# check forward
with torch.no_grad():
ref = model(test_data)
out = traced(test_data)
assert_allclose(ref, out)
def test_get_model_args():
demask = DeMask()
expected = {
"activation": "relu",
"dropout": 0,
"fb_name": "STFTFB",
"hidden_dims": (1024, ),
"input_type": "mag",
"kernel_size": 512,
"mask_act": "relu",
"n_filters": 512,
"norm_type": "gLN",
"output_type": "mag",
"sample_rate": 16000,
"stride": 256,
}
assert demask.get_model_args() == expected