recompute_optimizer.py 2.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and

from paddle.fluid.optimizer import RecomputeOptimizer as RO
from .meta_optimizer_base import MetaOptimizerBase


class RecomputeOptimizer(MetaOptimizerBase):
    def __init__(self, optimizer):
        super(RecomputeOptimizer, self).__init__(optimizer)
        #self.inner_opt = RO(optimizer)
        self.inner_opt = optimizer
        self.wrapped_opt = RO(optimizer)
        # we do not allow meta optimizer to be inner optimizer currently
25 26 27 28 29 30 31
        self.meta_optimizers_white_list = [
            "LarsOptimizer",
            "LambOptimizer",
            "GradientMergeOptimizer",
            "GraphExecutionOptimizer",
        ]
        self.meta_optimizers_black_list = []
32 33 34 35 36

    def _set_basic_info(self, loss, role_maker, user_defined_optimizer,
                        user_defined_strategy):
        super(RecomputeOptimizer, self)._set_basic_info(
            loss, role_maker, user_defined_optimizer, user_defined_strategy)
37 38
        self.wrapped_opt._set_checkpoints(
            list(user_defined_strategy.recompute_configs["checkpoints"]))
39 40

    def _can_apply(self):
M
mapingshuo 已提交
41
        if not self.role_maker._is_collective:
42 43
            return False

44
        if self.user_defined_strategy.recompute == True:
D
Dong Daxiang 已提交
45 46
            if len(self.user_defined_strategy.recompute_configs[
                    "checkpoints"]) == 0:
47 48 49 50
                return False
            else:
                return True

D
Dong Daxiang 已提交
51 52
    def _disable_strategy(self, dist_strategy):
        dist_strategy.recompute = False
53
        dist_strategy.recompute_configs = {}
D
Dong Daxiang 已提交
54

55
    def _enable_strategy(self, dist_strategy, context):
56 57 58
        # we do not support automatically recompute checkpoints currently
        return

59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
    def backward(self,
                 loss,
                 startup_program=None,
                 parameter_list=None,
                 no_grad_set=None,
                 callbacks=None):
        return self.wrapped_opt.backward(loss, startup_program, parameter_list,
                                         no_grad_set, callbacks)

    def minimize_impl(self,
                      loss,
                      startup_program=None,
                      parameter_list=None,
                      no_grad_set=None):
        optimize_ops, params_grads = \
            self.wrapped_opt.minimize(loss, startup_program,
                                      parameter_list, no_grad_set)
        return optimize_ops, params_grads