dist_fleet_heter_ctr.py 7.3 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 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179
#   Copyright (c) 2020 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.
"""
Distribute CTR model for test fleet api
"""

from __future__ import print_function

import shutil
import tempfile
import time

import paddle
import paddle.fluid as fluid
import os
import numpy as np

import ctr_dataset_reader
from test_dist_fleet_heter_base import runtime_main, FleetDistHeterRunnerBase
from dist_fleet_ctr import TestDistCTR2x2, fake_ctr_reader
from paddle.distributed.fleet.base.util_factory import fleet_util

# Fix seed for test
fluid.default_startup_program().random_seed = 1
fluid.default_main_program().random_seed = 1


class TestHeterPsCTR2x2(FleetDistHeterRunnerBase):
    """
    For test CTR model, using Fleet api
    """

    def net(self, args, batch_size=4, lr=0.01):
        """
        network definition

        Args:
            batch_size(int): the size of mini-batch for training
            lr(float): learning rate of training
        Returns:
            avg_cost: LoDTensor of cost.
        """
        dnn_input_dim, lr_input_dim = int(1e5), int(1e5)

        dnn_data = fluid.layers.data(
            name="dnn_data",
            shape=[-1, 1],
            dtype="int64",
            lod_level=1,
            append_batch_size=False)
        lr_data = fluid.layers.data(
            name="lr_data",
            shape=[-1, 1],
            dtype="int64",
            lod_level=1,
            append_batch_size=False)
        label = fluid.layers.data(
            name="click",
            shape=[-1, 1],
            dtype="float32",
            lod_level=0,
            append_batch_size=False)

        datas = [dnn_data, lr_data, label]

        if args.reader == "pyreader":
            self.reader = fluid.io.PyReader(
                feed_list=datas,
                capacity=64,
                iterable=False,
                use_double_buffer=False)

        # build dnn model
        dnn_layer_dims = [128, 64, 32, 1]
        dnn_embedding = fluid.layers.embedding(
            is_distributed=False,
            input=dnn_data,
            size=[dnn_input_dim, dnn_layer_dims[0]],
            param_attr=fluid.ParamAttr(
                name="deep_embedding",
                initializer=fluid.initializer.Constant(value=0.01)),
            is_sparse=True)
        dnn_pool = fluid.layers.sequence_pool(
            input=dnn_embedding, pool_type="sum")
        dnn_out = dnn_pool

        # build lr model
        lr_embbding = fluid.layers.embedding(
            is_distributed=False,
            input=lr_data,
            size=[lr_input_dim, 1],
            param_attr=fluid.ParamAttr(
                name="wide_embedding",
                initializer=fluid.initializer.Constant(value=0.01)),
            is_sparse=True)
        lr_pool = fluid.layers.sequence_pool(input=lr_embbding, pool_type="sum")

        with fluid.device_guard("gpu"):
            for i, dim in enumerate(dnn_layer_dims[1:]):
                fc = fluid.layers.fc(
                    input=dnn_out,
                    size=dim,
                    act="relu",
                    param_attr=fluid.ParamAttr(
                        initializer=fluid.initializer.Constant(value=0.01)),
                    name='dnn-fc-%d' % i)
                dnn_out = fc

            merge_layer = fluid.layers.concat(input=[dnn_out, lr_pool], axis=1)
            label = fluid.layers.cast(label, dtype="int64")
            predict = fluid.layers.fc(input=merge_layer, size=2, act='softmax')

            cost = fluid.layers.cross_entropy(input=predict, label=label)
            avg_cost = fluid.layers.mean(x=cost)
            fluid.layers.Print(avg_cost, message="avg_cost")

        self.feeds = datas
        self.train_file_path = ["fake1", "fake2"]
        self.avg_cost = avg_cost
        self.predict = predict

        return avg_cost

    def check_model_right(self, dirname):
        model_filename = os.path.join(dirname, "__model__")

        with open(model_filename, "rb") as f:
            program_desc_str = f.read()

        program = fluid.Program.parse_from_string(program_desc_str)
        with open(os.path.join(dirname, "__model__.proto"), "w") as wn:
            wn.write(str(program))

    def do_pyreader_training(self, fleet):
        """
        do training using dataset, using fetch handler to catch variable
        Args:
            fleet(Fleet api): the fleet object of Parameter Server, define distribute training role
        """

        exe = fluid.Executor(fluid.CPUPlace())
        fleet.init_worker()
        exe.run(fluid.default_startup_program())
        batch_size = 4
        train_reader = paddle.batch(fake_ctr_reader(), batch_size=batch_size)
        self.reader.decorate_sample_list_generator(train_reader)

        for epoch_id in range(1):
            self.reader.start()
            try:
                pass_start = time.time()
                while True:
                    exe.run(program=fluid.default_main_program())

                pass_time = time.time() - pass_start
            except fluid.core.EOFException:
                self.reader.reset()

        fleet.stop_worker()

    def do_dataset_training(self, fleet):
        train_file_list = ctr_dataset_reader.prepare_fake_data()

        exe = fluid.Executor(fluid.CPUPlace())

        fleet.init_worker()
        exe.run(fluid.default_startup_program())

C
Chengmo 已提交
180
        thread_num = int(os.getenv("CPU_NUM", 2))
181 182 183 184 185
        batch_size = 128
        filelist = fleet_util.get_file_shard(train_file_list)
        print("filelist: {}".format(filelist))

        # config dataset
186 187 188
        dataset = paddle.distributed.QueueDataset()
        dataset._set_batch_size(batch_size)
        dataset._set_use_var(self.feeds)
189
        pipe_command = 'python ctr_dataset_reader.py'
190
        dataset._set_pipe_command(pipe_command)
191 192

        dataset.set_filelist(filelist)
193
        dataset._set_thread(thread_num)
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

        for epoch_id in range(1):
            pass_start = time.time()
            dataset.set_filelist(filelist)
            exe.train_from_dataset(
                program=fluid.default_main_program(),
                dataset=dataset,
                fetch_list=[self.avg_cost],
                fetch_info=["cost"],
                print_period=2,
                debug=int(os.getenv("Debug", "0")))
            pass_time = time.time() - pass_start
            print("do_dataset_training done. using time {}".format(pass_time))
        if os.getenv("SAVE_MODEL") == "1":
            model_dir = tempfile.mkdtemp()
            fleet.save_inference_model(exe, model_dir,
                                       [feed.name for feed in self.feeds],
                                       self.avg_cost)
            self.check_model_right(model_dir)
            shutil.rmtree(model_dir)

        fleet.stop_worker()
        print("do_dataset_training stop worker.")


if __name__ == "__main__":
    runtime_main(TestHeterPsCTR2x2)