Can't use chainer.grad for LSTM

I tried to differentiate a LSTM block with chainer.grad function along with enable_double_backprop=True option. Then got the error:

Code:

import numpy as np
import chainer
import chainer.links as L

x = chainer.Variable(np.random.rand(10, 20).astype('f'))
lstm = L.LSTM(20, 20)
y = lstm(x)
dydx, = chainer.grad([y], [x], enable_double_backprop=True)

Error:

~/.pyenv/versions/anaconda3-5.0.0/lib/python3.6/site-packages/chainer/functions/activation/lstm.py in backward(self, indexes, grads)
    111     def backward(self, indexes, grads):
    112         grad_inputs = (
--> 113             self.get_retained_inputs() + self.get_retained_outputs() + grads)
    114         return LSTMGrad()(*grad_inputs)

Is this a bug of LSTM module or my wrong usage of LSTM? I’m using Chainer v4.0.0b3.

Any comments will help. Thanks.

[toslunar]

dydx, = chainer.grad(
    [y, lstm.c, lstm.h], [x],
    [np.ones_like(y), np.zeros_like(lstm.c), np.zeros_like(lstm.h)],
    enable_double_backprop=True)  # Now success with #4320

loss = 10. * F.mean_squared_error(dydx, np.ones_like(dydx.array))

lstm.cleargrads()
loss.backward()

seems to work if the lines

        if ggc_prev is None:
            ggc_prev = xp.zeros_like(c)

are added to LSTMGrad.backward.

[toslunar]

Sorry,

    [np.ones_like(y), np.zeros_like(lstm.c), np.zeros_like(lstm.h)],

should be

    [np.ones_like(y.data), np.zeros_like(lstm.c.data), np.zeros_like(lstm.h.data)],

It seems that leaving lstm.h.grad to be None is fine:

dydx, = chainer.grad(
    [y, lstm.c], [x],
    [np.ones_like(y.data), np.zeros_like(lstm.c.data)],
    enable_double_backprop=True)  # Now success with #4320