Passing gradient to the backward-func. of NN

Hi everyone,

I want to optimize a NN using Varitional-Optimization instead of pure backprob.

Consider following set:

  1. We have a NN with weights -> f(w)

  2. We have some samples ‘z_i’ from a normal-dist. -> z_i ∼ N(0, 1)​​

Now for applying VO on our NN, we have following equations:

unnamed (3)

unnamed (5)

unnamed (1)


My problem is now how can i moultiply the gradient w.r.t. to the weights of the function g(theta, z) with the grads of the netowrk?

i saw that i can pass a “grad”-variabel to the backward-fucntion but what should be the shape of the variable for applying it to all weights of the network?

this my code right now:

class VOFast(optim.Optimizer):
    def __init__(self, params, sample_size=100, mu=0.0, log_sigma=-4.0, interoptim=optim.Adam, alpha=0.9, **kwargs):
        defaults = dict(sample_size=sample_size)
        params = list(params)
        mean = []
        logsigma = []
        for param in params:
            len = self.param_length(param)
            mean_ = torch.Tensor(len)
            logsigma_ = torch.Tensor(len).fill_(log_sigma)

            if cuda_enabled:
                mean_ = mean_.cuda()
                logsigma_ = logsigma_.cuda()

            mean_ = Variable(mean_, requires_grad=True)
  [...] =[...]
            logsigma_ = Variable(logsigma_, requires_grad=True)


        self.optimizer_ = interoptim([{"params": mean}, {"params": logsigma}], **kwargs)

        param_groups = [dict(params=params, name="params"),
                        dict(params=mean, name="mean"),
                        dict(params=logsigma, name="logsigma")]

        super(VOFast, self).__init__(param_groups, defaults)

        self.baseline = None
        self.baselines = []
        self.alpha = alpha

    def __setstate__(self, state):
        super(VOFast, self).__setstate__(state)

    def zero_grad(self):

    def step(self, closure):
        sample_size = self.param_groups[0]['sample_size']
        params = self.param_groups[0]["params"]
        mean = self.param_groups[1]["params"]
        logsigma = self.param_groups[2]["params"]

        for i in range(sample_size):
            for idx, mu in enumerate(mean):
                samples = Variable(torch.randn(len(mu)), requires_grad=False)
                if cuda_enabled:
                    samples = samples.cuda()
                samples = samples.detach()

                z = torch.exp(logsigma[idx]) * samples + mean[idx]
                if cuda_enabled:
                    z = z.cuda()
                params[idx].data[...] = z.view(params[idx].shape).data

            loss_closure = closure()
            # here we shoudl actualy pass the grads w.r.t. to mean ans sigma to the backward-func!


def voptim(net, loss_func, trainloader, mean_val=1.0,
         log_sigma=-1.0, sample_size=100, lr=0.1, threshold=0.9):
    vanilla natrual evolution strategy
    :param net:
    :param loss_func:
    :param lr:
    kwargs = dict(lr=lr) # adam
    # kwargs = dict(lr=lr, momentum=0.9)  # sgd

    optimizer = VOFast(net.parameters(), sample_size=sample_size, interoptim=optim.Adam,
                        log_sigma=log_sigma, mu=mean_val, alpha=0.9, **kwargs)

    for epoch in range(2):
        for i, data in enumerate(trainloader, 0):

            X_tr, y_tr = data
            if cuda_enabled:
                X_tr = X_tr.cuda()
                y_tr = y_tr.cuda()

            X_train = Variable(X_tr)
            y_train = Variable(y_tr)

            def closure(input=X_train, target=y_train):
                output = net(input)
                loss = loss_func(output, target)
                return loss


            if i % 10 == 0:
                out = net(X_train)
                loss = loss_func(out, y_train)
                _, predicted = torch.max(, 1)
                accuracy = 100 * torch.sum( == predicted) /[0]
                print('[%d.%d], loss = %.2f, accuracy %d %%' % (epoch, i,[0], np.ceil(accuracy)))