parallel_executor.py

#   Copyright (c) 2018 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
# limitations under the License.

import core
import multiprocessing
import framework
import executor
import warnings
import sys

__all__ = ['ParallelExecutor']


class ParallelExecutor(object):
    def __init__(self,
                 use_cuda,
                 loss_name=None,
                 main_program=None,
                 num_threads=None,
                 allow_op_delay=False,
                 share_vars_from=None,
                 use_default_grad_scale=True,
                 balance_parameter_opt_between_cards=False,
                 num_trainers=0,
                 trainer_id=0):
        """
        ParallelExecutor can run program in parallel.

        Args:
            use_cuda(bool): Whether to use CUDA or not.
            loss_name(str, default None): The loss name must set in training.
            main_program(Program, default None): The program that need to run,
                if not provided, then default_main_program will be used.
            num_threads(int, default None): How many threads are used for
                training.
            allow_op_delay(bool, default False): Whether to delay and buffer
                some operators together for scheduling or not, which may
                improve performance in some cases, default False.
            share_vars_from(ParallelExecutor, default None): If provied,
                it will share variables from the specified ParallelExecutor.
            use_default_grad_scale(bool, default True): If set True, a default
                scale value equal to `1./device_count` would be multiplied to
                gradients of each device and scaled gradients would be
                aggregated. Otherwise, a customized scale value should be fed
                to the network.
            balance_parameter_opt_between_cards(bool, default True): Whether
                updating different gradients on different cards. Currently, it
                is not recommended.
            num_trainers(int, default 0): If greater than 0, NCCL will be
                initialized with multpile rank of nodes, each node should have
                same number of GPUs. Distributed training will be enabled then.
            trainer_id(int, default 0): Must use together with num_trainers.
                trainer_id is the "rank" of current node starts from 0.

        Returns:
            A ParallelExecutor object.

        Raises:
            TypeError: If share_vars_from is provided, but not ParallelExecutor
                object.

        Examples:
            .. code-block:: python

              train_exe = fluid.ParallelExecutor(
                  use_cuda=True, loss_name=loss.name)
              test_exe = fluid.ParallelExecutor(
                  use_cuda=True,
                  main_program=test_program,
                  share_vars_from=train_exe)

              train_loss, = train_exe.run([loss.name], feed=feed_dict)
              test_loss, = test_exe.run([loss.name], feed=feed_dict)
        """

        self._places = []
        self._act_places = []
        if use_cuda:
            for i in xrange(core.get_cuda_device_count()):
                p = core.Place()
                self._act_places.append(core.CUDAPlace(i))
                p.set_place(self._act_places[-1])
                self._places.append(p)
        else:
            for i in xrange(multiprocessing.cpu_count()):
                p = core.Place()
                self._act_places.append(core.CPUPlace())
                p.set_place(self._act_places[-1])
                self._places.append(p)
        assert self._places, "no place for execution"

        if num_threads is None:
            if use_cuda:
                # Experiments on se-resnext shows that too many threads hurt
                # performance. Worth tunning for other models in the future.
                num_threads = len(self._places) * 2
            else:
                num_threads = min(
                    len(self._places) * 2, multiprocessing.cpu_count())

        main = main_program
        main = main if main else framework.default_main_program()
        scope = executor.global_scope()

        if share_vars_from and not isinstance(share_vars_from,
                                              ParallelExecutor):
            raise TypeError("share_vars_from must be ParallelExecutor.")
        local_scopes = share_vars_from.executor.local_scopes(
        ) if share_vars_from else []

        self.persistable_vars = [
            v.name
            for v in filter(
                lambda var: var.persistable and var.type != core.VarDesc.VarType.RAW,
                main.list_vars())
        ]

        self.executor = core.ParallelExecutor(
            num_threads,
            True if use_cuda else False,  # use_event
            self._places,
            set([
                p.name for p in main.global_block().iter_parameters()
                if not p.stop_gradient
            ]),
            set(self.persistable_vars),
            main.desc,
            loss_name if loss_name else '',
            scope,
            local_scopes,
            allow_op_delay,
            use_default_grad_scale,
            balance_parameter_opt_between_cards,
            num_trainers,
            trainer_id)
        self.scope = scope

    def run(self, fetch_list, feed=None, feed_dict=None):
        """
        Run a parallel executor with fetch_list.

        The feed parameter can be a dict or a list. If feed is a dict, the
        feed data will be split into multiple devices. If feed is a list, we
        assume the data has been splitted into multiple devices, the each
        element in the list will be copied to each device directly.

        For example, if the feed is a dict:
        >>> exe = ParallelExecutor()
        >>> # the image will be splitted into devices. If there is two devices
        >>> # each device will process an image with shape (24, 1, 28, 28)
        >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))})

        For example, if the feed is a list:
        >>> exe = ParallelExecutor()
        >>> # each device will process each element in the list.
        >>> # the 1st device will process an image with shape (48, 1, 28, 28)
        >>> # the 2nd device will process an image with shape (32, 1, 28, 28)
        >>> #
        >>> # you can use exe.device_count to get the device number.
        >>> exe.run(feed=[{"image": numpy.random.random(size=(48, 1, 28, 28))},
        >>>               {"image": numpy.random.random(size=(32, 1, 28, 28))},
        >>>              ])


        Args:
            fetch_list(list): The fetched variable names
            feed(list|dict|None): The feed variables. If the feed is a dict,
                tensors in that dict will be splitted into each devices. If
                the feed is a list, each element of the list will be copied
                to each device.
            feed_dict: Alias for feed parameter, for backward compatibility.
                This parameter is deprecated.

        Returns: fetched result list.

        """
        if feed is None and feed_dict is not None:
            feed = feed_dict
            print >> sys.stderr, "`feed_dict` is deprecated. Please use `feed=`"

        if isinstance(feed, dict):
            feed_tensor_dict = dict()
            for feed_name in feed:
                feed_tensor = feed[feed_name]
                if not isinstance(feed_tensor, core.LoDTensor):
                    feed_tensor = core.LoDTensor()
                    # always set to CPU place, since the tensor need to be splitted
                    # it is fast in CPU
                    feed_tensor.set(feed[feed_name], core.CPUPlace())
                feed_tensor_dict[feed_name] = feed_tensor

            self.executor.feed_and_split_tensor_into_local_scopes(
                feed_tensor_dict)
        elif isinstance(feed, list) or isinstance(feed, tuple):
            if len(feed) != len(self._act_places):
                raise ValueError(
                    "Feed a list of tensor, the list should be the same size as places"
                )

            res = list()

            for i, each in enumerate(feed):
                if not isinstance(each, dict):
                    raise TypeError(
                        "Each element of feed list should be a dict")
                res_dict = dict()
                for feed_name in each:
                    tensor = each[feed_name]
                    if not isinstance(tensor, core.LoDTensor):
                        tmp = core.LoDTensor()
                        tmp.set(tensor, self._act_places[i])
                        tensor = tmp
                    res_dict[feed_name] = tensor
                res.append(res_dict)
            self.executor.feed_tensors_into_local_scopes(res)

        fetch_var_name = '@FETCHED_VAR_NAME@'
        self.executor.run(fetch_list, fetch_var_name)
        arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array()
        return [arr[i] for i in range(len(arr))]

    def bcast_params(self):
        self.executor.bcast_params(set(self.persistable_vars))

    @property
    def device_count(self):
        return len(self._act_places)