import torch
from torch.optim.optimizer import Optimizer

from .types import Betas2, OptFloat, OptLossClosure, Params

__all__ = ('NovoGrad',)


class NovoGrad(Optimizer):
    r"""Implements Novograd optimization algorithm.

    It has been proposed in `Stochastic Gradient Methods with Layer-wise
    Adaptive Moments for Training of Deep Networks`__.

    Arguments:
        params: iterable of parameters to optimize or dicts defining
            parameter groups
        lr: learning rate (default: 1e-3)
        betas: coefficients used for computing
            running averages of gradient and its square (default: (0.95, 0))
        eps: term added to the denominator to improve
            numerical stability (default: 1e-8)
        weight_decay: weight decay (L2 penalty) (default: 0)
        grad_averaging: gradient averaging (default: False)
        amsgrad: whether to use the AMSGrad variant of this
            algorithm from the paper `On the Convergence of Adam and Beyond`
            (default: False)

    Example:
        >>> import torch_optimizer as optim
        >>> optimizer = optim.Yogi(model.parameters(), lr=0.1)
        >>> optimizer.zero_grad()
        >>> loss_fn(model(input), target).backward()
        >>> scheduler = StepLR(optimizer, step_size=1, gamma=0.7)
        >>> optimizer.step()
        >>> scheduler.step()

    __ https://arxiv.org/abs/1905.11286

    Note:
        Reference code: https://github.com/NVIDIA/DeepLearningExamples
    """

    def __init__(
        self,
        params: Params,
        lr: float = 1e-3,
        betas: Betas2 = (0.95, 0),
        eps: float = 1e-8,
        weight_decay: float = 0,
        grad_averaging: bool = False,
        amsgrad: bool = False,
    ):
        if lr <= 0.0:
            raise ValueError('Invalid learning rate: {}'.format(lr))
        if eps < 0.0:
            raise ValueError('Invalid epsilon value: {}'.format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError(
                'Invalid beta parameter at index 0: {}'.format(betas[0])
            )
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError(
                'Invalid beta parameter at index 1: {}'.format(betas[1])
            )
        if weight_decay < 0:
            raise ValueError(
                'Invalid weight_decay value: {}'.format(weight_decay)
            )
        defaults = dict(
            lr=lr,
            betas=betas,
            eps=eps,
            weight_decay=weight_decay,
            grad_averaging=grad_averaging,
            amsgrad=amsgrad,
        )

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

    def __setstate__(self, state: dict) -> None:
        super(NovoGrad, self).__setstate__(state)
        for group in self.param_groups:
            group.setdefault('amsgrad', False)

    def step(self, closure: OptLossClosure = None) -> OptFloat:
        r"""Performs a single optimization step.

        Arguments:
            closure: 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:
            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data
                if grad.is_sparse:
                    msg = (
                        'NovoGrad does not support sparse gradients, '
                        'please consider SparseAdam instead'
                    )
                    raise RuntimeError(msg)
                amsgrad = group['amsgrad']

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    # Exponential moving average of gradient values
                    state['exp_avg'] = torch.zeros_like(p.data)
                    # Exponential moving average of squared gradient values
                    state['exp_avg_sq'] = torch.zeros([]).to(
                        state['exp_avg'].device
                    )
                    if amsgrad:
                        # Maintains max of all exp. moving avg. of sq.
                        # grad. values
                        state['max_exp_avg_sq'] = torch.zeros([]).to(
                            state['exp_avg'].device
                        )

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                if amsgrad:
                    max_exp_avg_sq = state['max_exp_avg_sq']
                beta1, beta2 = group['betas']

                state['step'] += 1

                norm = torch.sum(torch.pow(grad, 2))

                if exp_avg_sq == 0:
                    exp_avg_sq.copy_(norm)
                else:
                    exp_avg_sq.mul_(beta2).add_(norm, alpha=1 - beta2)

                if amsgrad:
                    # Maintains the maximum of all 2nd moment running avg.
                    # till now
                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
                    # Use the max. for normalizing running avg. of gradient
                    denom = max_exp_avg_sq.sqrt().add_(group['eps'])
                else:
                    denom = exp_avg_sq.sqrt().add_(group['eps'])

                grad.div_(denom)
                if group['weight_decay'] != 0:
                    grad.add_(p.data, alpha=group['weight_decay'])
                if group['grad_averaging']:
                    grad.mul_(1 - beta1)
                exp_avg.mul_(beta1).add_(grad)

                p.data.add_(exp_avg, alpha=-group['lr'])

        return loss