""" Layer-wise adaptive rate scaling for SGD in PyTorch! """
import torch
from torch.optim.optimizer import Optimizer, required
from torch.optim import *


class LARS(Optimizer):
    r"""Implements layer-wise adaptive rate scaling for SGD.

    Args:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float): base learning rate (\gamma_0)
        momentum (float, optional): momentum factor (default: 0) ("m")
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
            ("\beta")
        eta (float, optional): LARS coefficient
        max_epoch: maximum training epoch to determine polynomial LR decay.

    Based on Algorithm 1 of the following paper by You, Gitman, and Ginsburg.
    Large Batch Training of Convolutional Networks:
        https://arxiv.org/abs/1708.03888

    Example:
        >>> optimizer = LARS(model.parameters(), lr=0.1, eta=1e-3)
        >>> optimizer.zero_grad()
        >>> loss_fn(model(input), target).backward()
        >>> optimizer.step()
    """

    def __init__(self,
                 params,
                 lr=required,
                 momentum=.9,
                 weight_decay=.0005,
                 eta=0.001):
        if lr is not required and lr < 0.0:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if momentum < 0.0:
            raise ValueError("Invalid momentum value: {}".format(momentum))
        if weight_decay < 0.0:
            raise ValueError(
                "Invalid weight_decay value: {}".format(weight_decay))
        if eta < 0.0:
            raise ValueError("Invalid LARS coefficient value: {}".format(eta))

        defaults = dict(
            lr=lr, momentum=momentum, weight_decay=weight_decay, eta=eta)
        super(LARS, self).__init__(params, defaults)

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

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
            epoch: current epoch to calculate polynomial LR decay schedule.
                   if None, uses self.epoch and increments it.
        """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            weight_decay = group['weight_decay']
            momentum = group['momentum']
            eta = group['eta']
            lr = group['lr']

            for p in group['params']:
                if p.grad is None:
                    continue

                param_state = self.state[p]
                d_p = p.grad.data

                weight_norm = torch.norm(p.data)
                grad_norm = torch.norm(d_p)

                # Compute local learning rate for this layer
                local_lr = eta * weight_norm / \
                    (grad_norm + weight_decay * weight_norm)

                # Update the momentum term
                actual_lr = local_lr * lr

                if 'momentum_buffer' not in param_state:
                    buf = param_state['momentum_buffer'] = \
                            torch.zeros_like(p.data)
                else:
                    buf = param_state['momentum_buffer']
                buf.mul_(momentum).add_(actual_lr, d_p + weight_decay * p.data)
                p.data.add_(-buf)

        return loss