def __call__(self, noise, clean1, clean2, face1, face2):
noise = xp.asarray(noise).astype(xp.float32)
clean1 = xp.asarray(clean1).astype(xp.float32)
clean2 = xp.asarray(clean2).astype(xp.float32)
face1 = xp.asarray(face1).astype(xp.float32)[:, :, :, xp.newaxis]
face2 = xp.asarray(face2).astype(xp.float32)[:, :, :, xp.newaxis]
clean = xp.concatenate((clean1, clean2), axis=3)
compressed_noise, _ = op.compress_audio(noise)
compressed_clean, _ = op.compress_audio(clean)
mask1, mask2 = self.estimate_mask(spec=compressed_noise,
face1=face1,
face2=face2)
separated1 = op.mul(mask1, compressed_noise)
separated2 = op.mul(mask2, compressed_noise)
separated = F.concat((separated1, separated2),
axis=3) # (6, 2, 301, 514)
loss = evaluate_loss(self, separated, compressed_clean)
return loss
def get_batch_mnist(ds, index, repeat, image_size=112):
nt = ds.num_target
# print(index)
batch_size = index.shape[0]
return_x = np.empty((batch_size, 1, image_size, image_size))
return_t = np.zeros((batch_size, nt))
for bi in range(batch_size):
return_x[bi] = ds[index[bi]][0]
return_t[bi] = ds[index[bi]][1]
return_x = return_x.reshape(batch_size, 1, image_size, image_size).astype(np.float32)
return_t = return_t.astype(np.float32)
return_x = xp.asarray(xp.tile(return_x, (repeat, 1, 1, 1)))
return_t = xp.asarray(xp.tile(return_t, (repeat, 1)))
return return_x, return_t
def __call__(self, noise, clean):
noise = xp.asarray(noise).astype(xp.float32)
clean = xp.asarray(clean).astype(xp.float32)
compressed_noise, _ = op.compress_audio(noise)
compressed_clean, _ = op.compress_audio(clean)
mask, _ = self.estimate_mask(spec=compressed_noise)
separated = op.mul(mask, compressed_noise)
loss = evaluate_loss(self, separated, compressed_clean)
return loss