test_ctc.py 文件源码

def check_forward(self, t_data, xs_data, l_length, x_length):
        x = tuple(chainer.Variable(x_data) for x_data in xs_data)
        t = chainer.Variable(t_data)

        args = (x, t, self.blank_symbol)
        if self.use_length:
            args += (chainer.Variable(x_length), chainer.Variable(l_length))
        loss = functions.connectionist_temporal_classification(*args)
        loss_value = float(loss.data)

        # compute expected value by recursive computation.
        xp = cuda.get_array_module(self.x)
        xt = xp.swapaxes(self.x, 0, 1)
        for b in range(xt.shape[0]):
            for t in range(xt.shape[1]):
                xt[b][t] = numpy.exp(xt[b][t]) / numpy.sum(numpy.exp(xt[b][t]))
        loss_expect = 0
        batch_size = xt.shape[0]
        path_length = 2 * l_length + 1
        for xtb, lb, xlb, plb in zip(xt, self.l, x_length, path_length):
            loss_expect += -math.log(
                self.alpha(xtb, lb, int(xlb - 1), int(plb - 1)) +
                self.alpha(xtb, lb, int(xlb - 1), int(plb - 2)))
        loss_expect /= batch_size
        self.assertAlmostEqual(loss_expect, loss_value, places=5)