test_dist_fleet_ps11.py 8.2 KB
Newer Older
T
Thunderbrook 已提交
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
#   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 os
import unittest

import paddle
import paddle.fluid as fluid
import paddle.distributed.fleet.base.role_maker as role_maker
import paddle.distributed.fleet as fleet

paddle.enable_static()

# For Net
base_lr = 0.2
emb_lr = base_lr * 3
dict_dim = 1500
emb_dim = 128
hid_dim = 128
margin = 0.1
sample_rate = 1
batch_size = 4


class TestPSPassWithBow(unittest.TestCase):
37

T
Thunderbrook 已提交
38
    def net(self):
39

T
Thunderbrook 已提交
40 41 42 43
        def get_acc(cos_q_nt, cos_q_pt, batch_size):
            cond = fluid.layers.less_than(cos_q_nt, cos_q_pt)
            cond = fluid.layers.cast(cond, dtype='float64')
            cond_3 = fluid.layers.reduce_sum(cond)
44 45 46 47 48 49
            acc = fluid.layers.elementwise_div(cond_3,
                                               fluid.layers.fill_constant(
                                                   shape=[1],
                                                   value=batch_size * 1.0,
                                                   dtype='float64'),
                                               name="simnet_acc")
T
Thunderbrook 已提交
50 51 52 53
            return acc

        def get_loss(cos_q_pt, cos_q_nt):
            loss_op1 = fluid.layers.elementwise_sub(
54 55 56 57
                fluid.layers.fill_constant_batch_size_like(input=cos_q_pt,
                                                           shape=[-1, 1],
                                                           value=margin,
                                                           dtype='float32'),
T
Thunderbrook 已提交
58 59 60
                cos_q_pt)
            loss_op2 = fluid.layers.elementwise_add(loss_op1, cos_q_nt)
            loss_op3 = fluid.layers.elementwise_max(
61 62 63 64
                fluid.layers.fill_constant_batch_size_like(input=loss_op2,
                                                           shape=[-1, 1],
                                                           value=0.0,
                                                           dtype='float32'),
T
Thunderbrook 已提交
65
                loss_op2)
66
            avg_cost = paddle.mean(loss_op3)
T
Thunderbrook 已提交
67 68 69 70 71 72
            return avg_cost

        is_distributed = False
        is_sparse = True

        # query
73
        q = fluid.layers.data(name="1", shape=[1], dtype="int64", lod_level=1)
T
Thunderbrook 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96
        # embedding
        q_emb = fluid.contrib.layers.sparse_embedding(
            input=q,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__emb__",
                learning_rate=emb_lr))
        q_emb = fluid.layers.reshape(q_emb, [-1, emb_dim])
        # vsum
        q_sum = fluid.layers.sequence_pool(input=q_emb, pool_type='sum')
        q_ss = fluid.layers.softsign(q_sum)
        # fc layer after conv
        q_fc = fluid.layers.fc(
            input=q_ss,
            size=hid_dim,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__q_fc__",
                learning_rate=base_lr))
        # label data
        label = fluid.layers.data(name="label", shape=[1], dtype="int64")
        # pt
97
        pt = fluid.layers.data(name="2", shape=[1], dtype="int64", lod_level=1)
T
Thunderbrook 已提交
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
        # embedding
        pt_emb = fluid.contrib.layers.sparse_embedding(
            input=pt,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__emb__",
                learning_rate=emb_lr))
        pt_emb = fluid.layers.reshape(pt_emb, [-1, emb_dim])
        # vsum
        pt_sum = fluid.layers.sequence_pool(input=pt_emb, pool_type='sum')
        pt_ss = fluid.layers.softsign(pt_sum)
        # fc layer
        pt_fc = fluid.layers.fc(
            input=pt_ss,
            size=hid_dim,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__fc__",
                learning_rate=base_lr),
            bias_attr=fluid.ParamAttr(name="__fc_b__"))
        # nt
120
        nt = fluid.layers.data(name="3", shape=[1], dtype="int64", lod_level=1)
T
Thunderbrook 已提交
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
        # embedding
        nt_emb = fluid.contrib.layers.sparse_embedding(
            input=nt,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__emb__",
                learning_rate=emb_lr))
        nt_emb = fluid.layers.reshape(nt_emb, [-1, emb_dim])
        # vsum
        nt_sum = fluid.layers.sequence_pool(input=nt_emb, pool_type='sum')
        nt_ss = fluid.layers.softsign(nt_sum)
        # fc layer
        nt_fc = fluid.layers.fc(
            input=nt_ss,
            size=hid_dim,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.01),
                name="__fc__",
                learning_rate=base_lr),
            bias_attr=fluid.ParamAttr(name="__fc_b__"))
        cos_q_pt = fluid.layers.cos_sim(q_fc, pt_fc)
        cos_q_nt = fluid.layers.cos_sim(q_fc, nt_fc)
        # loss
        avg_cost = get_loss(cos_q_pt, cos_q_nt)
        # acc
        acc = get_acc(cos_q_nt, cos_q_pt, batch_size)
        return [avg_cost, acc, cos_q_pt]

    def test(self):
        os.environ["PADDLE_PSERVER_NUMS"] = "2"
        os.environ["PADDLE_TRAINERS_NUM"] = "2"
        os.environ["POD_IP"] = "127.0.0.1"
        os.environ["PADDLE_PORT"] = "36001"
        os.environ["PADDLE_TRAINER_ID"] = "0"
        os.environ["PADDLE_TRAINERS_NUM"] = "2"
        os.environ[
            "PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001,127.0.0.2:36001"
        os.environ[
            "PADDLE_PSERVERS_IP_PORT_LIST"] = "127.0.0.1:36002,127.0.0.2:36002"
        os.environ["TRAINING_ROLE"] = "TRAINER"
        os.environ["FLAGS_selected_gpus"] = "0"
        role = role_maker.PaddleCloudRoleMaker()
        fleet.init(role)
        loss, acc, _ = self.net()

        strategy = paddle.distributed.fleet.DistributedStrategy()
        configs = {"use_ps_gpu": 1, "launch_barrier": False}
        strategy.a_sync_configs = configs
        strategy.a_sync = True
        optimizer = paddle.fluid.optimizer.Adam(learning_rate=0.01)
        optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
        optimizer.minimize(loss)

175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
    def test_gpups_dataset(self):
        """
        Testcase for GPUPS InMemoryDataset .
        """
        with open("test_in_memory_dataset_run_a.txt", "w") as f:
            data = "1 1 2 3 3 4 5 5 5 5 1 1\n"
            data += "1 2 2 3 4 4 6 6 6 6 1 2\n"
            data += "1 3 2 3 5 4 7 7 7 7 1 3\n"
            f.write(data)
        with open("test_in_memory_dataset_run_b.txt", "w") as f:
            data = "1 4 2 3 3 4 5 5 5 5 1 4\n"
            data += "1 5 2 3 4 4 6 6 6 6 1 5\n"
            data += "1 6 2 3 5 4 7 7 7 7 1 6\n"
            data += "1 7 2 3 6 4 8 8 8 8 1 7\n"
            f.write(data)

        slots = ["slot1", "slot2", "slot3", "slot4"]
        slots_vars = []
        for slot in slots:
            var = fluid.layers.data(name=slot,
                                    shape=[1],
                                    dtype="int64",
                                    lod_level=1)
            slots_vars.append(var)

        dataset = paddle.distributed.InMemoryDataset()
        dataset._set_use_ps_gpu(True)
        dataset.init(batch_size=32,
                     thread_num=3,
                     pipe_command="cat",
                     use_var=slots_vars)
        dataset.set_filelist([
            "test_in_memory_dataset_run_a.txt",
            "test_in_memory_dataset_run_b.txt"
        ])

        os.remove("./test_in_memory_dataset_run_a.txt")
        os.remove("./test_in_memory_dataset_run_b.txt")

T
Thunderbrook 已提交
214 215 216

if __name__ == '__main__':
    unittest.main()