Source code for hype.rsgd

#!/usr/bin/env python3
# Copyright (c) 2018-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

from torch.optim.optimizer import Optimizer, required


[docs]class RiemannianSGD(Optimizer): """Riemannian stochastic gradient descent. Args: rgrad (Function): Function to compute the Riemannian gradient from the Euclidean gradient retraction (Function): Function to update the retraction of the Riemannian gradient """ def __init__( self, params, lr=required, rgrad=required, expm=required, ): defaults = { 'lr': lr, 'rgrad': rgrad, 'expm': expm, } super(RiemannianSGD, self).__init__(params, defaults)
[docs] def step(self, lr=None, **kwargs): """Performs a single optimization step. Arguments: lr (float, optional): learning rate for the current update. """ loss = None for group in self.param_groups: for p in group['params']: lr = lr or group['lr'] rgrad = group['rgrad'] expm = group['expm'] if p.grad is None: continue d_p = p.grad.data # make sure we have no duplicates in sparse tensor if d_p.is_sparse: d_p = d_p.coalesce() d_p = rgrad(p.data, d_p) d_p.mul_(-lr) expm(p.data, d_p) return loss