madgrad_wd.py

# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

# modifications  - 4/4/2021  @lessw2020  (decay issue spotted by @nestordemeure )
# weight decay has been implemented AdamW style instead of the original madgrad Adam style.
# in initial image classification testing, this outperformed 0 weight decay or original style weight decay.

# closure is checked if callable or not since some code passes loss directly, rather than in closure param

import math
from typing import Collection, TYPE_CHECKING, Any, Callable, Optional

import torch
import torch.optim
import collections

if TYPE_CHECKING:
    from torch.optim.optimizer import _params_t
else:
    _params_t = Any


class madgrad_wd(torch.optim.Optimizer):
    """
    MADGRAD_: A Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic
    Optimization.

    .. _MADGRAD: https://arxiv.org/abs/2101.11075

    MADGRAD is a general purpose optimizer that can be used in place of SGD or
    Adam may converge faster and generalize better. Currently GPU-only.
    Typically, the same learning rate schedule that is used for SGD or Adam may
    be used. The overall learning rate is not comparable to either method and
    should be determined by a hyper-parameter sweep.

    MADGRAD requires less weight decay than other methods, often as little as
    zero. Momentum values used for SGD or Adam's beta1 should work here also.

    On sparse problems both weight_decay and momentum should be set to 0.

    Arguments:
        params (iterable):
            Iterable of parameters to optimize or dicts defining parameter groups.
        lr (float):
            Learning rate (default: 1e-2).
        momentum (float):
            Momentum value in  the range [0,1) (default: 0.9).
        weight_decay (float):
            Weight decay, i.e. a L2 penalty (default: 0).
        eps (float):
            Term added to the denominator outside of the root operation to improve numerical stability. (default: 1e-6).
    """

    def __init__(
        self,
        params: _params_t,
        lr: float = 1e-2,
        momentum: float = 0.9,
        weight_decay: float = 0,
        eps: float = 1e-6,
    ):
        if momentum < 0 or momentum >= 1:
            raise ValueError(f"Momentum {momentum} must be in the range [0,1]")
        if lr <= 0:
            raise ValueError(f"Learning rate {lr} must be positive")
        if weight_decay < 0:
            raise ValueError(f"Weight decay {weight_decay} must be non-negative")
        if eps < 0:
            raise ValueError(f"Eps must be non-negative")

        defaults = dict(lr=lr, eps=eps, momentum=momentum, weight_decay=weight_decay)
        super().__init__(params, defaults)

    @property
    def supports_memory_efficient_fp16(self) -> bool:
        return False

    @property
    def supports_flat_params(self) -> bool:
        return True

    def step(self, closure: Optional[Callable[[], float]] = None) -> Optional[float]:
        """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 and isinstance(closure, collections.Callable):
            loss = closure()

        # step counter must be stored in state to ensure correct behavior under
        # optimizer sharding
        if "k" not in self.state:
            self.state["k"] = torch.tensor([0], dtype=torch.long)
        k = self.state["k"].item()

        for group in self.param_groups:
            eps = group["eps"]
            lr = group["lr"] + eps
            decay = group["weight_decay"]
            momentum = group["momentum"]

            ck = 1 - momentum
            lamb = lr * math.pow(k + 1, 0.5)

            for p in group["params"]:
                if p.grad is None:
                    continue
                grad = p.grad.data
                state = self.state[p]

                if "grad_sum_sq" not in state:
                    state["grad_sum_sq"] = torch.zeros_like(p.data).detach()
                    state["s"] = torch.zeros_like(p.data).detach()
                    if momentum != 0:
                        state["x0"] = torch.clone(p.data).detach()

                if momentum != 0.0 and grad.is_sparse:
                    raise RuntimeError(
                        "momentum != 0 is not compatible with sparse gradients"
                    )

                grad_sum_sq = state["grad_sum_sq"]
                s = state["s"]

                # Apply weight decay - L2 / AdamW style
                if decay:
                    p.data.mul_(1 - lr * decay)

                """ original impl:
                if decay != 0:
                    if grad.is_sparse:
                        raise RuntimeError("weight_decay option is not compatible with sparse gradients")

                    grad.add_(p.data, alpha=decay)
                """

                if grad.is_sparse:
                    grad = grad.coalesce()
                    grad_val = grad._values()

                    p_masked = p.sparse_mask(grad)
                    grad_sum_sq_masked = grad_sum_sq.sparse_mask(grad)
                    s_masked = s.sparse_mask(grad)

                    # Compute x_0 from other known quantities
                    rms_masked_vals = grad_sum_sq_masked._values().pow(1 / 3).add_(eps)
                    x0_masked_vals = p_masked._values().addcdiv(
                        s_masked._values(), rms_masked_vals, value=1
                    )

                    # Dense + sparse op
                    grad_sq = grad * grad
                    grad_sum_sq.add_(grad_sq, alpha=lamb)
                    grad_sum_sq_masked.add_(grad_sq, alpha=lamb)

                    rms_masked_vals = grad_sum_sq_masked._values().pow_(1 / 3).add_(eps)

                    s.add_(grad, alpha=lamb)
                    s_masked._values().add_(grad_val, alpha=lamb)

                    # update masked copy of p
                    p_kp1_masked_vals = x0_masked_vals.addcdiv(
                        s_masked._values(), rms_masked_vals, value=-1
                    )
                    # Copy updated masked p to dense p using an add operation
                    p_masked._values().add_(p_kp1_masked_vals, alpha=-1)
                    p.data.add_(p_masked, alpha=-1)
                else:
                    if momentum == 0:
                        # Compute x_0 from other known quantities
                        rms = grad_sum_sq.pow(1 / 3).add_(eps)
                        x0 = p.data.addcdiv(s, rms, value=1)
                    else:
                        x0 = state["x0"]

                    # Accumulate second moments
                    grad_sum_sq.addcmul_(grad, grad, value=lamb)
                    rms = grad_sum_sq.pow(1 / 3).add_(eps)

                    # Update s
                    s.data.add_(grad, alpha=lamb)

                    # Step
                    if momentum == 0:
                        p.data.copy_(x0.addcdiv(s, rms, value=-1))
                    else:
                        z = x0.addcdiv(s, rms, value=-1)

                        # p is a moving average of z
                        p.data.mul_(1 - ck).add_(z, alpha=ck)

        self.state["k"] += 1
        return loss