torch/optim/sparse_adam.py

import torch
from . import _functional as F
from .optimizer import Optimizer, _maximize_doc

__all__ = ['SparseAdam']

class SparseAdam(Optimizer):
    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, maximize: bool = False):
        if not 0.0 < lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 < eps:
            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]))

        params = list(params)

        sparse_params = []
        for index, param in enumerate(params):
            if isinstance(param, dict):
                # given param group, convert given params to a list first before iterating
                param['params'] = list(param.get("params", []))
                for d_index, d_param in enumerate(param['params']):
                    if d_param.is_sparse:
                        sparse_params.append([index, d_index])
            elif param.is_sparse:
                sparse_params.append(index)
        if sparse_params:
            raise ValueError(
                f"Sparse params at indices {sparse_params}: SparseAdam requires dense parameter tensors"
            )

        defaults = dict(lr=lr, betas=betas, eps=eps, maximize=maximize)
        super(SparseAdam, self).__init__(params, defaults)

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

        Args:
            closure (Callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            params_with_grad = []
            grads = []
            exp_avgs = []
            exp_avg_sqs = []
            state_steps = []
            eps = group['eps']
            lr = group['lr']
            beta1, beta2 = group['betas']
            maximize = group.get('maximize', False)

            for p in group['params']:
                if p.grad is not None:
                    params_with_grad.append(p)
                    if not p.grad.is_sparse:
                        raise RuntimeError('SparseAdam does not support dense gradients, please consider Adam instead')
                    grads.append(p.grad)

                    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, memory_format=torch.preserve_format)
                        # Exponential moving average of squared gradient values
                        state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)

                    exp_avgs.append(state['exp_avg'])
                    exp_avg_sqs.append(state['exp_avg_sq'])

                    # update the steps for each param group update
                    state['step'] += 1
                    # record the step after step update
                    state_steps.append(state['step'])

            F.sparse_adam(params_with_grad,
                          grads,
                          exp_avgs,
                          exp_avg_sqs,
                          state_steps,
                          beta1=beta1,
                          beta2=beta2,
                          lr=group['lr'],
                          eps=group['eps'],
                          maximize=maximize)

        return loss

SparseAdam.__doc__ = r"""Implements lazy version of Adam algorithm suitable for sparse tensors.

    In this variant, only moments that show up in the gradient get updated, and
    only those portions of the gradient get applied to the parameters.

    Args:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float, optional): learning rate (default: 1e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its square (default: (0.9, 0.999))
        eps (float, optional): term added to the denominator to improve
            numerical stability (default: 1e-8)
        {maximize}

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

    """.format(maximize=_maximize_doc)