Strom

import torch
from .optimizer import Optimizer


class Storm(Optimizer):
    r"""Implements stochastic
    """

    def __init__(self, params, k=0.1, w=10, c=0.01, weight_decay=0):
        if k < 0.0:
            raise ValueError("Invalid k value: {}".format(k))
        if w < 0.0:
            raise ValueError("Invalid w value: {}".format(w))
        if c < 0.0:
            raise ValueError('Invalid c value:{}'.format(c))

        defaults = dict(k=k, w=w, c=c, weight_decay=weight_decay)

        super(Storm, self).__init__(params, defaults)

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


    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:
            k = group['k']
            w = group['w']
            c = group['c']
            weight_decay = group['weight_decay']

            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)
                param_state = self.state[p]

                if len(param_state)==0:
                    param_state['last_data'] = torch.clone(d_p).detach()
                    buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
                    param_state['accumulated_grad_size'] = torch.norm(torch.clone(d_p).detach())
                    param_state['lr'] = k/(pow((w+pow(param_state['accumulated_grad_size'],2)),1/3))
                    param_state['momentum'] = c * pow(param_state['lr'], 2)
                else:
                    buf = param_state['momentum_buffer']
                    accumulated_grad_size = param_state['accumulated_grad_size']
                    accumulated_grad_size.add_(pow(torch.norm(d_p), 2))
                    param_state['lr'] = k / (pow((w + pow(param_state['accumulated_grad_size'], 2)), 1 / 3))
                    param_state['momentum'] = c * pow(param_state['lr'], 2)
                    # print(param_state['momentum'])
                    buf.mul_(1 - param_state['momentum']).add_(param_state['momentum'] , param_state['last_data']).add_(d_p)
                    # buf.mul_(1-param_state['momentum']).add_(param_state['momentum']-1,d_p)
                    param_state['last_data']=d_p
                    d_p = buf
                p.data.add_(-param_state['lr'], d_p)
        return loss