dist_fleet_ctr.py 8.6 KB
Newer Older
T
tangwei12 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
#   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.
14 15 16
"""
Distribute CTR model for test fleet api
"""
T
tangwei12 已提交
17 18 19 20 21 22 23

from __future__ import print_function

import shutil
import tempfile
import time

1
123malin 已提交
24
import paddle
T
tangwei12 已提交
25 26
import paddle.fluid as fluid
import os
1
123malin 已提交
27
import numpy as np
T
tangwei12 已提交
28 29 30

import ctr_dataset_reader
from test_dist_fleet_base import runtime_main, FleetDistRunnerBase
31
from paddle.distributed.fleet.base.util_factory import fleet_util
T
tangwei12 已提交
32 33 34 35 36 37

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


38 39 40 41 42 43 44 45 46 47 48
def fake_ctr_reader():
    def reader():
        for _ in range(1000):
            deep = np.random.random_integers(0, 1e5 - 1, size=16).tolist()
            wide = np.random.random_integers(0, 1e5 - 1, size=8).tolist()
            label = np.random.random_integers(0, 1, size=1).tolist()
            yield [deep, wide, label]

    return reader


T
tangwei12 已提交
49
class TestDistCTR2x2(FleetDistRunnerBase):
50 51 52 53
    """
    For test CTR model, using Fleet api
    """

54
    def net(self, args, batch_size=4, lr=0.01):
55 56 57 58 59 60 61 62 63
        """
        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.
        """
64 65
        dnn_input_dim, lr_input_dim = int(1e5), int(1e5)

T
tangwei12 已提交
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
        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="int64",
            lod_level=0,
            append_batch_size=False)

        datas = [dnn_data, lr_data, label]

87 88 89 90 91 92 93
        if args.reader == "pyreader":
            self.reader = fluid.io.PyReader(
                feed_list=datas,
                capacity=64,
                iterable=False,
                use_double_buffer=False)

T
tangwei12 已提交
94
        # build dnn model
C
Chengmo 已提交
95
        dnn_layer_dims = [128, 128, 64, 32, 1]
T
tangwei12 已提交
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
        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
        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

        # 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")

        merge_layer = fluid.layers.concat(input=[dnn_out, lr_pool], axis=1)

        predict = fluid.layers.fc(input=merge_layer, size=2, act='softmax')
        acc = fluid.layers.accuracy(input=predict, label=label)
132

T
tangwei12 已提交
133 134
        auc_var, batch_auc_var, auc_states = fluid.layers.auc(input=predict,
                                                              label=label)
135

T
tangwei12 已提交
136 137 138 139
        cost = fluid.layers.cross_entropy(input=predict, label=label)
        avg_cost = fluid.layers.mean(x=cost)

        self.feeds = datas
140
        self.train_file_path = ["fake1", "fake2"]
T
tangwei12 已提交
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
        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))

1
123malin 已提交
156
    def do_pyreader_training(self, fleet):
157 158 159 160 161
        """
        do training using dataset, using fetch handler to catch variable
        Args:
            fleet(Fleet api): the fleet object of Parameter Server, define distribute training role
        """
T
tangwei12 已提交
162 163 164 165 166

        exe = fluid.Executor(fluid.CPUPlace())
        fleet.init_worker()
        exe.run(fleet.startup_program)

167 168
        batch_size = 4
        train_reader = paddle.batch(fake_ctr_reader(), batch_size=batch_size)
1
123malin 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184
        self.reader.decorate_sample_list_generator(train_reader)

        compiled_prog = fluid.compiler.CompiledProgram(
            fleet.main_program).with_data_parallel(
                loss_name=self.avg_cost.name,
                build_strategy=self.strategy.get_build_strategy(),
                exec_strategy=self.strategy.get_execute_strategy())

        for epoch_id in range(1):
            self.reader.start()
            try:
                pass_start = time.time()
                while True:
                    loss_val = exe.run(program=compiled_prog,
                                       fetch_list=[self.avg_cost.name])
                    loss_val = np.mean(loss_val)
185 186 187 188 189
                    # TODO(randomly fail)
                    #   reduce_output = fleet_util.all_reduce(
                    #       np.array(loss_val), mode="sum")
                    #   loss_all_trainer = fleet_util.all_gather(float(loss_val))
                    #   loss_val = float(reduce_output) / len(loss_all_trainer)
190 191 192 193
                    message = "TRAIN ---> pass: {} loss: {}\n".format(epoch_id,
                                                                      loss_val)
                    fleet_util.print_on_rank(message, 0)

1
123malin 已提交
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
                pass_time = time.time() - pass_start
            except fluid.core.EOFException:
                self.reader.reset()

        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()

    def do_dataset_training(self, fleet):
        dnn_input_dim, lr_input_dim, train_file_path = ctr_dataset_reader.prepare_data(
        )

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

        fleet.init_worker()
        exe.run(fleet.startup_program)

        thread_num = 2
        batch_size = 128
T
tangwei12 已提交
216 217 218 219 220
        filelist = []
        for _ in range(thread_num):
            filelist.append(train_file_path)

        # config dataset
221
        dataset = paddle.distributed.fleet.DatasetFactory().create_dataset()
1
123malin 已提交
222
        dataset.set_batch_size(batch_size)
T
tangwei12 已提交
223 224 225 226 227 228 229
        dataset.set_use_var(self.feeds)
        pipe_command = 'python ctr_dataset_reader.py'
        dataset.set_pipe_command(pipe_command)

        dataset.set_filelist(filelist)
        dataset.set_thread(thread_num)

230
        for epoch_id in range(1):
T
tangwei12 已提交
231 232 233 234 235 236 237
            pass_start = time.time()
            dataset.set_filelist(filelist)
            exe.train_from_dataset(
                program=fleet.main_program,
                dataset=dataset,
                fetch_list=[self.avg_cost],
                fetch_info=["cost"],
238
                print_period=2,
239
                debug=int(os.getenv("Debug", "0")))
240 241
            pass_time = time.time() - pass_start

242 243 244 245 246 247 248
        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)
249

T
tangwei12 已提交
250 251 252 253 254
        fleet.stop_worker()


if __name__ == "__main__":
    runtime_main(TestDistCTR2x2)