No with torch.no_grad() in optim.SGD.step

Part 2 (2019)
1

I was digging in optim.SGD.step and was expecting with torch.no_grad(), but it is absent!
instead, I found

    def step(self, closure=None):
        """Performs a single optimization step.

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            weight_decay = group['weight_decay']
            momentum = group['momentum']
            dampening = group['dampening']
            nesterov = group['nesterov']

            for p in group['params']:
                if p.grad is None:
                    continue
                d_p = p.grad.data
                if weight_decay != 0:
                    d_p.add_(weight_decay, p.data)
                if momentum != 0:
                    param_state = self.state[p]
                    if 'momentum_buffer' not in param_state:
                        buf = param_state['momentum_buffer'] = torch.zeros_like(p.data)
                        buf.mul_(momentum).add_(d_p)
                    else:
                        buf = param_state['momentum_buffer']
                        buf.mul_(momentum).add_(1 - dampening, d_p)
                    if nesterov:
                        d_p = d_p.add(momentum, buf)
                    else:
                        d_p = buf

                p.data.add_(-group['lr'], d_p)

        return loss

Does it mean that using “.data” is like just getting values of grad/parameters and can be used instead of with torch.no_grad. If so why is this way better?

3

.data is available for backwards compataibility in pytorch. It was used earlier (the same time optim.SGD was written). It is not recommended to use .data in pytorch, instead use torch.no_grad.

1 Like
4

@zkylord
.data() is a reference to the value of a variable/tensor in Pytorch, and any operation done on .data is NOT tracked by PyTorch’s autograd. I actually don’t know if there is any advantage of this over with torch.no_grad(). Maybe someone else may be able to answer that.
Regards

5

Other than if .data is just removed from Pytorch in the future, as @kushaj pointed out.

6

Thanks @PalaashAgrawal & @kushaj