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 recurrent_forward(self, xm, lm, sm):
ls = xp.concatenate([lm.data, sm.data], axis=1)
hgl = F.relu(self.glimpse_loc(Variable(ls)))
h = self.glimpse_forward(xm)
hr1 = F.relu(self.rnn_1(hgl * h))
hr2 = F.relu(self.rnn_2(hr1))
l = F.sigmoid(self.attention_loc(hr2))
s = F.sigmoid(self.attention_scale(hr2))
y = F.softmax(self.full_2(hgl))
b = F.sigmoid(self.baseline(Variable(hr2.data)))
return l, s, y, b