Source code for torch.optim.adamax

# mypy: allow-untyped-decorators
# mypy: allow-untyped-defs
from typing import cast, List, Optional, Tuple, Union

import torch
from torch import Tensor

from .optimizer import (
    _capturable_doc,
    _default_to_fused_or_foreach,
    _differentiable_doc,
    _disable_dynamo_if_unsupported,
    _foreach_doc,
    _get_capturable_supported_devices,
    _get_scalar_dtype,
    _get_value,
    _maximize_doc,
    _params_doc,
    _use_grad_for_differentiable,
    _view_as_real,
    Optimizer,
    ParamsT,
)


__all__ = ["Adamax", "adamax"]


class Adamax(Optimizer):
    def __init__(
        self,
        params: ParamsT,
        lr: Union[float, Tensor] = 2e-3,
        betas: Tuple[float, float] = (0.9, 0.999),
        eps: float = 1e-8,
        weight_decay: float = 0,
        foreach: Optional[bool] = None,
        *,
        maximize: bool = False,
        differentiable: bool = False,
        capturable: bool = False,
    ):
        if isinstance(lr, Tensor) and lr.numel() != 1:
            raise ValueError("Tensor lr must be 1-element")
        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 <= weight_decay:
            raise ValueError(f"Invalid weight_decay value: {weight_decay}")

        defaults = dict(
            lr=lr,
            betas=betas,
            eps=eps,
            weight_decay=weight_decay,
            foreach=foreach,
            maximize=maximize,
            differentiable=differentiable,
            capturable=capturable,
        )
        super().__init__(params, defaults)

    def __setstate__(self, state):
        super().__setstate__(state)
        for group in self.param_groups:
            group.setdefault("foreach", None)
            group.setdefault("maximize", False)
            group.setdefault("differentiable", False)
            group.setdefault("capturable", False)
            for p in group["params"]:
                p_state = self.state.get(p, [])
                if len(p_state) != 0 and not torch.is_tensor(p_state["step"]):
                    step_val = float(p_state["step"])
                    p_state["step"] = (
                        torch.tensor(
                            step_val, dtype=_get_scalar_dtype(), device=p.device
                        )
                        if group["capturable"]
                        else torch.tensor(step_val, dtype=_get_scalar_dtype())
                    )

    def _init_group(
        self, group, params_with_grad, grads, exp_avgs, exp_infs, state_steps
    ):
        has_complex = False
        for p in group["params"]:
            if p.grad is None:
                continue
            has_complex |= torch.is_complex(p)
            params_with_grad.append(p)
            if p.grad.is_sparse:
                raise RuntimeError("Adamax does not support sparse gradients")
            grads.append(p.grad)

            state = self.state[p]

            # State initialization
            if len(state) == 0:
                state["step"] = (
                    torch.zeros((), dtype=_get_scalar_dtype(), device=p.device)
                    if group["capturable"]
                    else torch.tensor(0.0, dtype=_get_scalar_dtype())
                )
                state["exp_avg"] = torch.zeros_like(
                    p, memory_format=torch.preserve_format
                )
                state["exp_inf"] = torch.zeros_like(
                    p, memory_format=torch.preserve_format
                )

            exp_avgs.append(state["exp_avg"])
            exp_infs.append(state["exp_inf"])
            state_steps.append(state["step"])

        return has_complex

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

        Args:
            closure (Callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        self._cuda_graph_capture_health_check()

        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            params_with_grad: List[Tensor] = []
            grads: List[Tensor] = []
            exp_avgs: List[Tensor] = []
            exp_infs: List[Tensor] = []
            state_steps: List[Tensor] = []

            beta1, beta2 = group["betas"]
            eps = group["eps"]
            lr = group["lr"]
            weight_decay = group["weight_decay"]
            foreach = group["foreach"]
            maximize = group["maximize"]
            differentiable = group["differentiable"]
            capturable = group["capturable"]

            has_complex = self._init_group(
                group, params_with_grad, grads, exp_avgs, exp_infs, state_steps
            )

            adamax(
                params_with_grad,
                grads,
                exp_avgs,
                exp_infs,
                state_steps,
                eps=eps,
                beta1=beta1,
                beta2=beta2,
                lr=lr,
                weight_decay=weight_decay,
                foreach=foreach,
                maximize=maximize,
                differentiable=differentiable,
                capturable=capturable,
                has_complex=has_complex,
            )

        return loss


Adamax.__doc__ = (
    r"""Implements Adamax algorithm (a variant of Adam based on infinity norm).

    .. math::
       \begin{aligned}
            &\rule{110mm}{0.4pt}                                                                 \\
            &\textbf{input}      : \gamma \text{ (lr)}, \beta_1, \beta_2
                \text{ (betas)},\theta_0 \text{ (params)},f(\theta) \text{ (objective)},
                \: \lambda \text{ (weight decay)},                                                \\
            &\hspace{13mm}    \epsilon \text{ (epsilon)}                                          \\
            &\textbf{initialize} :  m_0 \leftarrow 0 \text{ ( first moment)},
                u_0 \leftarrow 0 \text{ ( infinity norm)}                                 \\[-1.ex]
            &\rule{110mm}{0.4pt}                                                                 \\
            &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do}                         \\
            &\hspace{5mm}g_t           \leftarrow   \nabla_{\theta} f_t (\theta_{t-1})           \\
            &\hspace{5mm}if \: \lambda \neq 0                                                    \\
            &\hspace{10mm} g_t \leftarrow g_t + \lambda  \theta_{t-1}                            \\
            &\hspace{5mm}m_t      \leftarrow   \beta_1 m_{t-1} + (1 - \beta_1) g_t               \\
            &\hspace{5mm}u_t      \leftarrow   \mathrm{max}(\beta_2 u_{t-1}, |g_{t}|+\epsilon)   \\
            &\hspace{5mm}\theta_t \leftarrow \theta_{t-1} - \frac{\gamma m_t}{(1-\beta^t_1) u_t} \\
            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
            &\bf{return} \:  \theta_t                                                     \\[-1.ex]
            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
       \end{aligned}

    For further details regarding the algorithm we refer to `Adam: A Method for Stochastic Optimization`_.
    """
    + rf"""
    Args:
        {_params_doc}
        lr (float, Tensor, optional): learning rate (default: 2e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its square
        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)
        {_foreach_doc}
        {_maximize_doc}
        {_differentiable_doc}
        {_capturable_doc}

    .. _Adam\: A Method for Stochastic Optimization:
        https://arxiv.org/abs/1412.6980

    """
)


def _single_tensor_adamax(
    params: List[Tensor],
    grads: List[Tensor],
    exp_avgs: List[Tensor],
    exp_infs: List[Tensor],
    state_steps: List[Tensor],
    *,
    eps: float,
    beta1: float,
    beta2: float,
    lr: float,
    weight_decay: float,
    maximize: bool,
    differentiable: bool,
    capturable: bool,
    has_complex: bool,
):
    for i, param in enumerate(params):
        grad = grads[i]
        grad = grad if not maximize else -grad
        exp_avg = exp_avgs[i]
        exp_inf = exp_infs[i]
        step_t = state_steps[i]

        # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
        if not torch.compiler.is_compiling() and capturable:
            capturable_supported_devices = _get_capturable_supported_devices()
            assert (
                param.device.type == step_t.device.type
                and param.device.type in capturable_supported_devices
            ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."

        # update step
        step_t += 1

        if weight_decay != 0:
            grad = grad.add(param, alpha=weight_decay)

        if torch.is_complex(param):
            param = torch.view_as_real(param)
            grad = torch.view_as_real(grad)
            exp_avg = torch.view_as_real(exp_avg)
            exp_inf = torch.view_as_real(exp_inf)

        # Update biased first moment estimate.
        exp_avg.lerp_(grad, 1 - beta1)
        # Update the exponentially weighted infinity norm.
        if not differentiable:
            torch.maximum(
                exp_inf.mul_(beta2),
                grad.abs().add_(eps),
                out=exp_inf,
            )
        else:
            norm_buf = torch.cat(
                [exp_inf.mul_(beta2).unsqueeze(0), grad.abs().add_(eps).unsqueeze_(0)],
                0,
            )
            exp_inf.copy_(torch.amax(norm_buf, 0, keepdim=False))

        if capturable:
            # why jump through extra hoops and negate bias_correction? check out #121238
            # once fixed, we should use bias_correction with addcdiv value=-1 for readability
            neg_bias_correction = beta1**step_t - 1
            neg_bias_correction.div_(lr)
            denom = exp_inf * neg_bias_correction
            param.addcdiv_(exp_avg, denom)
        else:
            bias_correction = 1 - beta1 ** _get_value(step_t)
            clr = lr / bias_correction

            param.addcdiv_(exp_avg, exp_inf, value=-clr)


def _multi_tensor_adamax(
    params: List[Tensor],
    grads: List[Tensor],
    exp_avgs: List[Tensor],
    exp_infs: List[Tensor],
    state_steps: List[Tensor],
    *,
    eps: float,
    beta1: float,
    beta2: float,
    lr: float,
    weight_decay: float,
    maximize: bool,
    differentiable: bool,
    capturable: bool,
    has_complex: bool,
):
    assert not differentiable, "_foreach ops don't support autograd"

    if len(params) == 0:
        return

    # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
    if not torch.compiler.is_compiling() and capturable:
        capturable_supported_devices = _get_capturable_supported_devices(
            supports_xla=False
        )
        assert all(
            p.device.type == step.device.type
            and p.device.type in capturable_supported_devices
            for p, step in zip(params, state_steps)
        ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."

    grouped_tensors = Optimizer._group_tensors_by_device_and_dtype(
        [params, grads, exp_avgs, exp_infs, state_steps]  # type: ignore[list-item]
    )
    for (
        grouped_params_,
        grouped_grads_,
        grouped_exp_avgs_,
        grouped_exp_infs_,
        grouped_state_steps_,
    ), _ in grouped_tensors.values():
        grouped_params = cast(List[Tensor], grouped_params_)
        grouped_grads = cast(List[Tensor], grouped_grads_)
        grouped_exp_avgs = cast(List[Tensor], grouped_exp_avgs_)
        grouped_exp_infs = cast(List[Tensor], grouped_exp_infs_)
        grouped_state_steps = cast(List[Tensor], grouped_state_steps_)

        if has_complex:
            _view_as_real(
                grouped_params, grouped_grads, grouped_exp_avgs, grouped_exp_infs
            )

        if maximize:
            grouped_grads = torch._foreach_neg(grouped_grads)  # type: ignore[assignment]

        # Update steps
        # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over
        # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just
        # wrapped it once now. The alpha is required to assure we go to the right overload.
        if not torch.compiler.is_compiling() and grouped_state_steps[0].is_cpu:
            torch._foreach_add_(
                grouped_state_steps, torch.tensor(1.0, device="cpu"), alpha=1.0
            )
        else:
            torch._foreach_add_(grouped_state_steps, 1)

        if weight_decay != 0:
            if maximize:
                # Re-use the intermediate memory (grouped_grads) already allocated for maximize
                torch._foreach_add_(grouped_grads, grouped_params, alpha=weight_decay)
            else:
                grouped_grads = torch._foreach_add(  # type: ignore[assignment]
                    grouped_grads, grouped_params, alpha=weight_decay
                )

        # Update biased first moment estimate.
        torch._foreach_lerp_(grouped_exp_avgs, grouped_grads, 1 - beta1)

        # Update the exponentially weighted infinity norm.
        torch._foreach_mul_(grouped_exp_infs, beta2)

        # in this case, we need to introduce a copy of the grads
        # since one has not been introduced previously
        if not maximize and weight_decay == 0:
            grouped_grads = torch._foreach_abs(grouped_grads)  # type: ignore[assignment]
        else:
            torch._foreach_abs_(grouped_grads)

        torch._foreach_add_(grouped_grads, eps)
        torch._foreach_maximum_(grouped_exp_infs, grouped_grads)

        bias_corrections: Union[Tuple[Tensor, ...], List[Tensor]]
        if capturable:
            bias_corrections = torch._foreach_pow(beta1, grouped_state_steps)
            # foreach_sub doesn't allow a scalar as the first arg
            torch._foreach_sub_(bias_corrections, 1)
            torch._foreach_div_(bias_corrections, lr)

            denom = torch._foreach_mul(grouped_exp_infs, bias_corrections)
            torch._foreach_addcdiv_(grouped_params, grouped_exp_avgs, denom)
        else:
            bias_corrections = [
                1 - beta1 ** _get_value(step) for step in grouped_state_steps
            ]
            step_size = [(_get_value(lr) / bc) * -1 for bc in bias_corrections]
            torch._foreach_addcdiv_(
                grouped_params, grouped_exp_avgs, grouped_exp_infs, step_size
            )


@_disable_dynamo_if_unsupported(single_tensor_fn=_single_tensor_adamax)
def adamax(
    params: List[Tensor],
    grads: List[Tensor],
    exp_avgs: List[Tensor],
    exp_infs: List[Tensor],
    state_steps: List[Tensor],
    # kwonly args with defaults are not supported by functions compiled with torchscript issue #70627
    # setting this as kwarg for now as functional API is compiled by torch/distributed/optim
    foreach: Optional[bool] = None,
    maximize: bool = False,
    differentiable: bool = False,
    capturable: bool = False,
    has_complex: bool = False,
    *,
    eps: float,
    beta1: float,
    beta2: float,
    lr: float,
    weight_decay: float,
):
    r"""Functional API that performs adamax algorithm computation.

    See :class:`~torch.optim.Adamax` for details.
    """

    if not torch.compiler.is_compiling() and not all(
        isinstance(t, torch.Tensor) for t in state_steps
    ):
        raise RuntimeError(
            "API has changed, `state_steps` argument must contain a list of singleton tensors"
        )

    if foreach is None:
        _, foreach = _default_to_fused_or_foreach(
            params, differentiable, use_fused=False
        )

    if foreach and torch.jit.is_scripting():
        raise RuntimeError("torch.jit.script not supported with foreach optimizers")

    if foreach and not torch.jit.is_scripting():
        func = _multi_tensor_adamax
    else:
        func = _single_tensor_adamax

    func(
        params,
        grads,
        exp_avgs,
        exp_infs,
        state_steps,
        eps=eps,
        beta1=beta1,
        beta2=beta2,
        lr=lr,
        weight_decay=weight_decay,
        maximize=maximize,
        differentiable=differentiable,
        has_complex=has_complex,
        capturable=capturable,
    )