# 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. from __future__ import print_function import multiprocessing from . import core from . import framework from . import executor from .. import compat as cpt import warnings import sys import six import os __all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): """ ParallelExecutor is designed for data parallelism, which focuses on distributing the data across different nodes and every node operates on the data in parallel. If you use ParallelExecutor to run the current program on GPU, the node means GPU device, and ParallelExecutor will get the available GPU device automatically on the current machine. If you use ParallelExecutor to run the current program on CPU, the node means the CPU device, and you can specify the CPU device number by adding 'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number of CPUs in the system. Args: use_cuda (bool): Whether to use CUDA or not. loss_name (str): The loss name must set in training. Default None. main_program (Program): The program that need to run, if not provided, then default_main_program will be used. Default None. share_vars_from(ParallelExecutor): If provide, it will share variables from the specified ParallelExecutor. Default None. exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run the program in ParallelExecutor, for example how many threads are used to execute the program, how many iterations to clean up the temp variables which is generated during execution. For more information, please refer to fluid.ExecutionStrategy. Default None. build_strategy(BuildStrategy): build_strategy is used to control how to build the SSA Graph in ParallelExecutor by setting the property, for example reduce_strategy, gradient_scale_strategy. For more information, please refer to fluid.BuildStrategy. Default None. num_trainers(int): If greater than 1, NCCL will be initialized with multiple rank of nodes, each node should have same number of GPUs. Distributed training will be enabled then. Default 1. trainer_id(int): Must use together with num_trainers. trainer_id is the "rank" of current node starts from 0. Default 0. scope(Scope): scope to run with, default use fluid.global_scope(). Returns: ParallelExecutor: The initialized 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) """ def __init__(self, use_cuda, loss_name=None, main_program=None, share_vars_from=None, exec_strategy=None, build_strategy=None, num_trainers=1, trainer_id=0, scope=None): self._places = [] self._act_places = [] if use_cuda: gpus = [] gpus_env = os.getenv("FLAGS_selected_gpus") if gpus_env: gpus = [int(s) for s in gpus_env.split(",")] else: for i in six.moves.range(core.get_cuda_device_count()): gpus.append(i) for i in gpus: p = core.Place() self._act_places.append(core.CUDAPlace(i)) p.set_place(self._act_places[-1]) self._places.append(p) else: cpu_num = int( os.environ.get('CPU_NUM', multiprocessing.cpu_count())) for i in six.moves.range(cpu_num): 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 exec_strategy is None: exec_strategy = ExecutionStrategy() exec_strategy.use_cuda = use_cuda if exec_strategy.num_threads == 0: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. exec_strategy.num_threads = len(self._places) * 4 else: cpu_num = int( os.environ.get('CPU_NUM', multiprocessing.cpu_count())) exec_strategy.num_threads = cpu_num * 2 if build_strategy is None: build_strategy = BuildStrategy() build_strategy.num_trainers = num_trainers build_strategy.trainer_id = trainer_id main = main_program main = main if main else framework.default_main_program() trainers_endpoints = main._trainers_endpoints if num_trainers > 1 and trainers_endpoints: assert num_trainers == len( trainers_endpoints), "num_trainers == len(end_points)" build_strategy.trainers_endpoints = trainers_endpoints if scope == None: 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 [ var for var in main.list_vars() if var.persistable and var.type != core.VarDesc.VarType.RAW ] ] self.executor = core.ParallelExecutor( self._places, set([ cpt.to_text(p.name) for p in main.global_block().iter_parameters() if not p.stop_gradient ]), set(cpt.to_text(var) for var in self.persistable_vars), main.desc, cpt.to_text(loss_name) if loss_name else six.u(''), scope, local_scopes, exec_strategy, build_strategy, num_trainers, trainer_id) self.scope = scope def run(self, fetch_list, feed=None, feed_dict=None, return_numpy=True): """ 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. Default None. feed_dict: Alias for feed parameter, for backward compatibility. This parameter has been deprecated. Default None. return_numpy(bool): Whether converts the fetched tensor to numpy. Default: True. Returns: List: The fetched result list. Raises: ValueError: If the feed is a list, but its length is not equal the length of active places, or its element's is not dict. NOTES: 1. If the feed's type is dict, the number of data that feeds to ParallelExecutor must be bigger than active places. Otherwise, it will throw exception from C++ side. Special attention should be paid to check whether the last batch of the dataset is bigger than active places. 2. If active places are more than one, the fetch results for each variable is a list, and each element of this list is the variable of respective active place. Examples: .. code-block:: python pe = fluid.ParallelExecutor(use_cuda=use_cuda, loss_name=avg_cost.name, main_program=fluid.default_main_program()) loss = pe.run(feed=feeder.feed(cur_batch), fetch_list=[avg_cost.name])) """ if feed is None and feed_dict is not None: feed = feed_dict print( "`feed_dict` is deprecated. Please use `feed=`", file=sys.stderr) 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() if return_numpy: return executor.as_numpy(arr) return [arr[i] for i in range(len(arr))] @property def device_count(self): return len(self._act_places)