LBFGS wrapper for Pytorch

lbfgs.py

"""
Copyright (c) 2017 Arthur Mensch

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""

import torch
from functools import reduce

from scipy.optimize import fmin_l_bfgs_b
import numpy as np
eps = np.finfo('double').eps

from torch.optim import Optimizer


class LBFGSScipy(Optimizer):
    """Wrap L-BFGS algorithm, using scipy routines.

    .. warning::
        This optimizer doesn't support per-parameter options and parameter
        groups (there can be only one).

    .. warning::
        Right now CPU only

    .. note::
        This is a very memory intensive optimizer (it requires additional
        ``param_bytes * (history_size + 1)`` bytes). If it doesn't fit in memory
        try reducing the history size, or use a different algorithm.

    Arguments:
        max_iter (int): maximal number of iterations per optimization step
            (default: 20)
        max_eval (int): maximal number of function evaluations per optimization
            step (default: max_iter * 1.25).
        tolerance_grad (float): termination tolerance on first order optimality
            (default: 1e-5).
        tolerance_change (float): termination tolerance on function
            value/parameter changes (default: 1e-9).
        history_size (int): update history size (default: 100).
    """

    def __init__(self, params, max_iter=20, max_eval=None,
                 tolerance_grad=1e-5, tolerance_change=1e-9, history_size=10,
                 ):
        if max_eval is None:
            max_eval = max_iter * 5 // 4
        defaults = dict(max_iter=max_iter, max_eval=max_eval,
                        tolerance_grad=tolerance_grad, tolerance_change=tolerance_change,
                        history_size=history_size)
        super(LBFGSScipy, self).__init__(params, defaults)

        if len(self.param_groups) != 1:
            raise ValueError("LBFGS doesn't support per-parameter options "
                             "(parameter groups)")

        self._params = self.param_groups[0]['params']
        self._numel_cache = None

        self._n_iter = 0
        self._last_loss = None

    def _numel(self):
        if self._numel_cache is None:
            self._numel_cache = reduce(lambda total, p: total + p.numel(), self._params, 0)
        return self._numel_cache

    def _gather_flat_grad(self):
        views = []
        for p in self._params:
            if p.grad is None:
                view = p.data.new(p.data.numel()).zero_()
            elif p.grad.data.is_sparse:
                view = p.grad.data.to_dense().view(-1)
            else:
                view = p.grad.data.view(-1)
            views.append(view)
        return torch.cat(views, 0)

    def _gather_flat_params(self):
        views = []
        for p in self._params:
            if p.data.is_sparse:
                view = p.data.to_dense().view(-1)
            else:
                view = p.data.view(-1)
            views.append(view)
        return torch.cat(views, 0)

    def _distribute_flat_params(self, params):
        offset = 0
        for p in self._params:
            numel = p.numel()
            # view as to avoid deprecated pointwise semantics
            p.data = params[offset:offset + numel].view_as(p.data)
            offset += numel
        assert offset == self._numel()

    def step(self, closure):
        """Performs a single optimization step.

        Arguments:
            closure (callable): A closure that reevaluates the model
                and returns the loss.
        """
        assert len(self.param_groups) == 1

        group = self.param_groups[0]
        max_iter = group['max_iter']
        max_eval = group['max_eval']
        tolerance_grad = group['tolerance_grad']
        tolerance_change = group['tolerance_change']
        history_size = group['history_size']

        def wrapped_closure(flat_params):
            """closure must call zero_grad() and backward()"""
            flat_params = torch.from_numpy(flat_params)
            self._distribute_flat_params(flat_params)
            loss = closure()
            self._last_loss = loss
            loss = loss.data[0]
            flat_grad = self._gather_flat_grad().numpy()
            return loss, flat_grad

        def callback(flat_params):
            self._n_iter += 1
            print('Iter %i Loss %.5f' % (self._n_iter, self._last_loss.data[0]))

        initial_params = self._gather_flat_params()

        fmin_l_bfgs_b(wrapped_closure, initial_params, maxiter=max_iter,
                      maxfun=max_eval,
                      factr=tolerance_change / eps, pgtol=tolerance_grad, epsilon=0,
                      m=history_size,
                      callback=callback)

Hello @arthurmensch ,

Amazing work, I used and it just works. Would it be possible to add a license on top of the file (e.g. MIT) so we know we can use it in our projects?
Thanks in advance!

Done ;) May the loss go down

Thanks so much for taking the time!