PyTorch `reinforce()` function sucks so I keep the alternative solution here

reinforce.py

import torch as T
import numpy as np
x = T.autograd.Variable(T.randn(5, 8), requires_grad=True)
p = T.nn.functional.softmax(x)
y = p.multinomial()
y.reinforce(T.ones(y.size()))
y.backward()
d = x.grad.data.clone().numpy()
x.grad.data.zero_()
logp = T.nn.functional.log_softmax(x)
logp_selected = logp.gather(1, T.autograd.Variable(y.data))
logp_selected.backward(-T.ones(y.size()))   # notice the minus
d2 = x.grad.data.clone().numpy()

assert np.all(np.abs(d - d2) < 1e-3)