dist_train.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 argparse
import time
import os
import traceback
import functools
import subprocess

import numpy as np

import paddle
import paddle.fluid as fluid
import six
import sys
sys.path.append("..")
import models
import utils
from reader import train, val
from utility import add_arguments, print_arguments
from batch_merge import copyback_repeat_bn_params, append_bn_repeat_init_op
from dist_utils import pserver_prepare, nccl2_prepare
from env import dist_env


def parse_args():
    parser = argparse.ArgumentParser(description=__doc__)
    add_arg = functools.partial(add_arguments, argparser=parser)
    # yapf: disable
    add_arg('batch_size',       int,   256,                  "Minibatch size.")
    add_arg('use_gpu',          bool,  True,                 "Whether to use GPU or not.")
    add_arg('total_images',     int,   1281167,              "Training image number.")
    add_arg('num_epochs',       int,   120,                  "number of epochs.")
    add_arg('class_dim',        int,   1000,                 "Class number.")
    add_arg('image_shape',      str,   "3,224,224",          "input image size")
    add_arg('model_save_dir',   str,   "output",             "model save directory")
    add_arg('pretrained_model', str,   None,                 "Whether to use pretrained model.")
    add_arg('checkpoint',       str,   None,                 "Whether to resume checkpoint.")
    add_arg('lr',               float, 0.1,                  "set learning rate.")
    add_arg('lr_strategy',      str,   "piecewise_decay",    "Set the learning rate decay strategy.")
    add_arg('model',            str,   "DistResNet",         "Set the network to use.")
    add_arg('enable_ce',        bool,  False,                "If set True, enable continuous evaluation job.")
    add_arg('data_dir',         str,   "./data/ILSVRC2012",  "The ImageNet dataset root dir.")
    add_arg('model_category',   str,   "models",             "Whether to use models_name or not, valid value:'models','models_name'" )
    add_arg('fp16',             bool,  False,                "Enable half precision training with fp16." )
    add_arg('scale_loss',       float, 1.0,                  "Scale loss for fp16." )
    add_arg('reduce_master_grad', bool, False,               "Whether to allreduce fp32 gradients." )
    # for distributed
    add_arg('update_method',      str,  "local",            "Can be local, pserver, nccl2.")
    add_arg('multi_batch_repeat', int,  1,                  "Batch merge repeats.")
    add_arg('start_test_pass',    int,  0,                  "Start test after x passes.")
    add_arg('num_threads',        int,  8,                  "Use num_threads to run the fluid program.")
    add_arg('split_var',          bool, True,               "Split params on pserver.")
    add_arg('async_mode',         bool, False,              "Async distributed training, only for pserver mode.")
    add_arg('reduce_strategy',    str,  "allreduce",        "Choose from reduce or allreduce.")
    add_arg('enable_sequential_execution', bool, False,            "Skip data not if data not balanced on nodes.")
    #for dgc
    add_arg('enable_dgc', bool, False,            "Skip data not if data not balanced on nodes.")
    add_arg('rampup_begin_step', int, 5008,            "Skip data not if data not balanced on nodes.")
    # yapf: enable
    args = parser.parse_args()
    return args


def get_device_num():
    if os.getenv("CPU_NUM"):
        return int(os.getenv("CPU_NUM"))
    return fluid.core.get_cuda_device_count()


def prepare_reader(is_train, pyreader, args, pass_id=1):
    # NOTE: always use infinite reader for dist training
    if is_train:
        reader = train(
            data_dir=args.data_dir, pass_id_as_seed=pass_id, infinite=True)
    else:
        reader = val(data_dir=args.data_dir)
    if is_train:
        bs = args.batch_size / get_device_num()
    else:
        bs = 16
    pyreader.decorate_paddle_reader(paddle.batch(reader, batch_size=bs))


def build_program(is_train, main_prog, startup_prog, args):
    pyreader = None
    class_dim = args.class_dim
    image_shape = [int(m) for m in args.image_shape.split(",")]

    trainer_count = args.dist_env["num_trainers"]
    device_num_per_worker = get_device_num()
    with fluid.program_guard(main_prog, startup_prog):
        pyreader = fluid.layers.py_reader(
            capacity=16,
            shapes=([-1] + image_shape, (-1, 1)),
            dtypes=('float32', 'int64'),
            name="train_reader" if is_train else "test_reader",
            use_double_buffer=True)
        with fluid.unique_name.guard():
            image, label = fluid.layers.read_file(pyreader)
            if args.fp16:
                image = fluid.layers.cast(image, "float16")
            model_def = models.__dict__[args.model](layers=50,
                                                    is_train=is_train)
            predict = model_def.net(image, class_dim=class_dim)
            cost, pred = fluid.layers.softmax_with_cross_entropy(
                predict, label, return_softmax=True)
            if args.scale_loss > 1:
                avg_cost = fluid.layers.mean(x=cost) * float(args.scale_loss)
            else:
                avg_cost = fluid.layers.mean(x=cost)

            batch_acc1 = fluid.layers.accuracy(input=pred, label=label, k=1)
            batch_acc5 = fluid.layers.accuracy(input=pred, label=label, k=5)

            optimizer = None
            if is_train:
                start_lr = args.lr
                end_lr = args.lr * trainer_count * args.multi_batch_repeat
                if os.getenv("FLAGS_selected_gpus"):
                    # in multi process mode, "trainer_count" will be total devices
                    # in the whole cluster, and we need to scale num_of nodes.
                    end_lr /= device_num_per_worker

                total_images = args.total_images / trainer_count
                if os.getenv("FLAGS_selected_gpus"):
                    step = int(total_images /
                               (args.batch_size / device_num_per_worker *
                                args.multi_batch_repeat) + 1)
                else:
                    step = int(total_images / (args.batch_size *
                                               args.multi_batch_repeat) + 1)
                warmup_steps = step * 5  # warmup 5 passes
                epochs = [30, 60, 80]
                bd = [step * e for e in epochs]
                base_lr = end_lr
                lr = []
                lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
                print("start lr: %s, end lr: %s, decay boundaries: %s" %
                      (start_lr, end_lr, bd))

                # NOTE: we put weight decay in layers config, and remove
                # weight decay on bn layers, so don't add weight decay in
                # optimizer config.
                optimizer = fluid.optimizer.Momentum(
                    learning_rate=utils.learning_rate.lr_warmup(
                        fluid.layers.piecewise_decay(
                            boundaries=bd, values=lr),
                        warmup_steps,
                        start_lr,
                        end_lr),
                    momentum=0.9)

                if args.enable_dgc:
                    optimizer = fluid.optimizer.DGCMomentumOptimizer(
                        learning_rate=utils.learning_rate.lr_warmup(
                            fluid.layers.piecewise_decay(
                                boundaries=bd, values=lr),
                            warmup_steps,
                            start_lr,
                            end_lr),
                        momentum=0.9,
                        sparsity=[0.999, 0.999],
                        rampup_begin_step=args.rampup_begin_step)

                if args.fp16:
                    params_grads = optimizer.backward(avg_cost)
                    master_params_grads = utils.create_master_params_grads(
                        params_grads,
                        main_prog,
                        startup_prog,
                        args.scale_loss,
                        reduce_master_grad=args.reduce_master_grad)
                    optimizer.apply_gradients(master_params_grads)
                    utils.master_param_to_train_param(master_params_grads,
                                                      params_grads, main_prog)
                else:
                    optimizer.minimize(avg_cost)

    # prepare reader for current program
    prepare_reader(is_train, pyreader, args)

    return pyreader, avg_cost, batch_acc1, batch_acc5


def test_single(exe, test_prog, args, pyreader, fetch_list):
    acc1 = fluid.metrics.Accuracy()
    acc5 = fluid.metrics.Accuracy()
    test_losses = []
    pyreader.start()
    while True:
        try:
            acc_rets = exe.run(program=test_prog, fetch_list=fetch_list)
            test_losses.append(acc_rets[0])
            acc1.update(value=np.array(acc_rets[1]), weight=args.batch_size)
            acc5.update(value=np.array(acc_rets[2]), weight=args.batch_size)
        except fluid.core.EOFException:
            pyreader.reset()
            break
    test_avg_loss = np.mean(np.array(test_losses))
    return test_avg_loss, np.mean(acc1.eval()), np.mean(acc5.eval())


def test_parallel(exe, test_prog, args, pyreader, fetch_list):
    acc1 = fluid.metrics.Accuracy()
    acc5 = fluid.metrics.Accuracy()
    test_losses = []
    pyreader.start()
    while True:
        try:
            acc_rets = exe.run(fetch_list=fetch_list)
            test_losses.append(acc_rets[0])
            acc1.update(value=np.array(acc_rets[1]), weight=args.batch_size)
            acc5.update(value=np.array(acc_rets[2]), weight=args.batch_size)
        except fluid.core.EOFException:
            pyreader.reset()
            break
    test_avg_loss = np.mean(np.array(test_losses))
    return test_avg_loss, np.mean(acc1.eval()), np.mean(acc5.eval())


def run_pserver(train_prog, startup_prog):
    server_exe = fluid.Executor(fluid.CPUPlace())
    server_exe.run(startup_prog)
    server_exe.run(train_prog)


def train_parallel(args):
    train_prog = fluid.Program()
    test_prog = fluid.Program()
    startup_prog = fluid.Program()

    train_pyreader, train_cost, train_acc1, train_acc5 = build_program(
        True, train_prog, startup_prog, args)
    test_pyreader, test_cost, test_acc1, test_acc5 = build_program(
        False, test_prog, startup_prog, args)

    if args.update_method == "pserver":
        train_prog, startup_prog = pserver_prepare(args, train_prog,
                                                   startup_prog)
    elif args.update_method == "nccl2":
        nccl2_prepare(args, startup_prog, main_prog=train_prog)

    if args.dist_env["training_role"] == "PSERVER":
        run_pserver(train_prog, startup_prog)
        exit(0)

    if args.use_gpu:
        # NOTE: for multi process mode: one process per GPU device.        
        gpu_id = 0
        if os.getenv("FLAGS_selected_gpus"):
            gpu_id = int(os.getenv("FLAGS_selected_gpus"))
    place = fluid.CUDAPlace(gpu_id) if args.use_gpu else fluid.CPUPlace()

    startup_exe = fluid.Executor(place)
    if args.multi_batch_repeat > 1:
        append_bn_repeat_init_op(train_prog, startup_prog,
                                 args.multi_batch_repeat)
    startup_exe.run(startup_prog)

    if args.checkpoint:
        fluid.io.load_persistables(
            startup_exe, args.checkpoint, main_program=train_prog)

    strategy = fluid.ExecutionStrategy()
    strategy.num_threads = args.num_threads
    # num_iteration_per_drop_scope indicates how
    # many iterations to clean up the temp variables which
    # is generated during execution. It may make the execution faster,
    # because the temp variable's shape are the same between two iterations.
    strategy.num_iteration_per_drop_scope = 30

    build_strategy = fluid.BuildStrategy()
    build_strategy.enable_inplace = False
    build_strategy.memory_optimize = False
    build_strategy.enable_sequential_execution = bool(
        args.enable_sequential_execution)

    if args.reduce_strategy == "reduce":
        build_strategy.reduce_strategy = fluid.BuildStrategy(
        ).ReduceStrategy.Reduce
    else:
        build_strategy.reduce_strategy = fluid.BuildStrategy(
        ).ReduceStrategy.AllReduce

    if args.update_method == "pserver" or args.update_method == "local":
        # parameter server mode distributed training, merge
        # gradients on local server, do not initialize
        # ParallelExecutor with multi server all-reduce mode.
        num_trainers = 1
        trainer_id = 0
    else:
        num_trainers = args.dist_env["num_trainers"]
        trainer_id = args.dist_env["trainer_id"]
        # Set this to let build_strategy to add "allreduce_deps_pass" automatically
        build_strategy.num_trainers = num_trainers
        build_strategy.trainer_id = trainer_id

    if args.multi_batch_repeat > 1:
        pass_builder = build_strategy._finalize_strategy_and_create_passes()
        mypass = pass_builder.insert_pass(
            len(pass_builder.all_passes()) - 4, "multi_batch_merge_pass")
        mypass.set("num_repeats", args.multi_batch_repeat)

    exe = fluid.ParallelExecutor(
        True,
        train_cost.name,
        main_program=train_prog,
        exec_strategy=strategy,
        build_strategy=build_strategy,
        num_trainers=num_trainers,
        trainer_id=trainer_id)

    # Uncomment below lines to use ParallelExecutor to run test.
    # test_exe = fluid.ParallelExecutor(
    #     True,
    #     main_program=test_prog,
    #     share_vars_from=exe,
    #     scope=fluid.global_scope().new_scope()
    # )

    over_all_start = time.time()
    fetch_list = [train_cost.name, train_acc1.name, train_acc5.name]
    # 1. MP mode, batch size for current process should be args.batch_size / GPUs
    # 2. SP/PG mode, batch size for each process should be original args.batch_size
    if os.getenv("FLAGS_selected_gpus"):
        steps_per_pass = args.total_images / (
            args.batch_size / get_device_num()) / args.dist_env["num_trainers"]
    else:
        steps_per_pass = args.total_images / args.batch_size / args.dist_env[
            "num_trainers"]

    for pass_id in range(args.num_epochs):
        num_samples = 0
        start_time = time.time()
        batch_id = 1
        if pass_id == 0:
            train_pyreader.start()
        while True:
            try:
                if batch_id % 30 == 0:
                    fetch_ret = exe.run(fetch_list)
                    fetched_data = [np.mean(np.array(d)) for d in fetch_ret]
                    print(
                        "Pass [%d/%d], batch [%d/%d], loss %s, acc1: %s, acc5: %s, avg batch time %.4f"
                        % (pass_id, args.num_epochs, batch_id, steps_per_pass,
                           fetched_data[0], fetched_data[1], fetched_data[2],
                           (time.time() - start_time) / batch_id))
                else:
                    fetch_ret = exe.run([])
            except fluid.core.EOFException:
                break
            except fluid.core.EnforceNotMet:
                traceback.print_exc()
                break
            num_samples += args.batch_size
            batch_id += 1
            if batch_id >= steps_per_pass:
                break

        print_train_time(start_time, time.time(), num_samples)
        if pass_id >= args.start_test_pass:
            if args.multi_batch_repeat > 1:
                copyback_repeat_bn_params(train_prog)
            test_fetch_list = [test_cost.name, test_acc1.name, test_acc5.name]
            test_ret = test_single(startup_exe, test_prog, args, test_pyreader,
                                   test_fetch_list)
            # NOTE: switch to below line if you use ParallelExecutor to run test.
            # test_ret = test_parallel(test_exe, test_prog, args, test_pyreader,test_fetch_list)
            print("Pass: %d, Test Loss %s, test acc1: %s, test acc5: %s\n" %
                  (pass_id, test_ret[0], test_ret[1], test_ret[2]))
            model_path = os.path.join(args.model_save_dir + '/' + args.model,
                                      str(pass_id))
            print("saving model to ", model_path)
            if not os.path.isdir(model_path):
                os.makedirs(model_path)
            fluid.io.save_persistables(
                startup_exe, model_path, main_program=train_prog)
    train_pyreader.reset()
    startup_exe.close()
    print("total train time: ", time.time() - over_all_start)


def print_train_time(start_time, end_time, num_samples):
    train_elapsed = end_time - start_time
    examples_per_sec = num_samples / train_elapsed
    print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
          (num_samples, train_elapsed, examples_per_sec))


def print_paddle_envs():
    print('----------- Configuration envs -----------')
    for k in os.environ:
        if "PADDLE_" in k:
            print("ENV %s:%s" % (k, os.environ[k]))
    print('------------------------------------------------')


def main():
    args = parse_args()
    print_arguments(args)
    print_paddle_envs()
    args.dist_env = dist_env()
    train_parallel(args)


if __name__ == "__main__":
    main()