timm/optim/adahessian.py

""" AdaHessian Optimizer

Lifted from https://github.com/davda54/ada-hessian/blob/master/ada_hessian.py
Originally licensed MIT, Copyright 2020, David Samuel
"""
import torch


class Adahessian(torch.optim.Optimizer):
    """
    Implements the AdaHessian algorithm from "ADAHESSIAN: An Adaptive Second OrderOptimizer for Machine Learning"

    Arguments:
        params (iterable): iterable of parameters to optimize or dicts defining parameter groups
        lr (float, optional): learning rate (default: 0.1)
        betas ((float, float), optional): coefficients used for computing running averages of gradient and the
            squared hessian trace (default: (0.9, 0.999))
        eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-8)
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0.0)
        hessian_power (float, optional): exponent of the hessian trace (default: 1.0)
        update_each (int, optional): compute the hessian trace approximation only after *this* number of steps
            (to save time) (default: 1)
        n_samples (int, optional): how many times to sample `z` for the approximation of the hessian trace (default: 1)
    """

    def __init__(self, params, lr=0.1, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.0,
                 hessian_power=1.0, update_each=1, n_samples=1, avg_conv_kernel=False):
        if not 0.0 <= lr:
            raise ValueError(f"Invalid learning rate: {lr}")
        if not 0.0 <= eps:
            raise ValueError(f"Invalid epsilon value: {eps}")
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError(f"Invalid beta parameter at index 0: {betas[0]}")
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError(f"Invalid beta parameter at index 1: {betas[1]}")
        if not 0.0 <= hessian_power <= 1.0:
            raise ValueError(f"Invalid Hessian power value: {hessian_power}")

        self.n_samples = n_samples
        self.update_each = update_each
        self.avg_conv_kernel = avg_conv_kernel

        # use a separate generator that deterministically generates the same `z`s across all GPUs in case of distributed training
        self.seed = 2147483647
        self.generator = torch.Generator().manual_seed(self.seed)

        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, hessian_power=hessian_power)
        super(Adahessian, self).__init__(params, defaults)

        for p in self.get_params():
            p.hess = 0.0
            self.state[p]["hessian step"] = 0

    @property
    def is_second_order(self):
        return True

    def get_params(self):
        """
        Gets all parameters in all param_groups with gradients
        """

        return (p for group in self.param_groups for p in group['params'] if p.requires_grad)

    def zero_hessian(self):
        """
        Zeros out the accumalated hessian traces.
        """

        for p in self.get_params():
            if not isinstance(p.hess, float) and self.state[p]["hessian step"] % self.update_each == 0:
                p.hess.zero_()

    @torch.no_grad()
    def set_hessian(self):
        """
        Computes the Hutchinson approximation of the hessian trace and accumulates it for each trainable parameter.
        """

        params = []
        for p in filter(lambda p: p.grad is not None, self.get_params()):
            if self.state[p]["hessian step"] % self.update_each == 0:  # compute the trace only each `update_each` step
                params.append(p)
            self.state[p]["hessian step"] += 1

        if len(params) == 0:
            return

        if self.generator.device != params[0].device:  # hackish way of casting the generator to the right device
            self.generator = torch.Generator(params[0].device).manual_seed(self.seed)

        grads = [p.grad for p in params]

        for i in range(self.n_samples):
            # Rademacher distribution {-1.0, 1.0}
            zs = [torch.randint(0, 2, p.size(), generator=self.generator, device=p.device) * 2.0 - 1.0 for p in params]
            h_zs = torch.autograd.grad(
                grads, params, grad_outputs=zs, only_inputs=True, retain_graph=i < self.n_samples - 1)
            for h_z, z, p in zip(h_zs, zs, params):
                p.hess += h_z * z / self.n_samples  # approximate the expected values of z*(H@z)

    @torch.no_grad()
    def step(self, closure=None):
        """
        Performs a single optimization step.
        Arguments:
            closure (callable, optional) -- a closure that reevaluates the model and returns the loss (default: None)
        """

        loss = None
        if closure is not None:
            loss = closure()

        self.zero_hessian()
        self.set_hessian()

        for group in self.param_groups:
            for p in group['params']:
                if p.grad is None or p.hess is None:
                    continue

                if self.avg_conv_kernel and p.dim() == 4:
                    p.hess = torch.abs(p.hess).mean(dim=[2, 3], keepdim=True).expand_as(p.hess).clone()

                # Perform correct stepweight decay as in AdamW
                p.mul_(1 - group['lr'] * group['weight_decay'])

                state = self.state[p]

                # State initialization
                if len(state) == 1:
                    state['step'] = 0
                    # Exponential moving average of gradient values
                    state['exp_avg'] = torch.zeros_like(p)
                    # Exponential moving average of Hessian diagonal square values
                    state['exp_hessian_diag_sq'] = torch.zeros_like(p)

                exp_avg, exp_hessian_diag_sq = state['exp_avg'], state['exp_hessian_diag_sq']
                beta1, beta2 = group['betas']
                state['step'] += 1

                # Decay the first and second moment running average coefficient
                exp_avg.mul_(beta1).add_(p.grad, alpha=1 - beta1)
                exp_hessian_diag_sq.mul_(beta2).addcmul_(p.hess, p.hess, value=1 - beta2)

                bias_correction1 = 1 - beta1 ** state['step']
                bias_correction2 = 1 - beta2 ** state['step']

                k = group['hessian_power']
                denom = (exp_hessian_diag_sq / bias_correction2).pow_(k / 2).add_(group['eps'])

                # make update
                step_size = group['lr'] / bias_correction1
                p.addcdiv_(exp_avg, denom, value=-step_size)

        return loss