test_dist_fleet_spmt.py 9.0 KB
Newer Older
L
lxsbupt 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 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
#   Copyright (c) 2022 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

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 TestSPMT(unittest.TestCase):
    def net(self):
        def get_acc(cos_q_nt, cos_q_pt, batch_size):
            cond = paddle.less_than(cos_q_nt, cos_q_pt)
            cond = fluid.layers.cast(cond, dtype='float64')
            cond_3 = paddle.sum(cond)
            acc = paddle.divide(
                cond_3,
                fluid.layers.fill_constant(
                    shape=[1], value=batch_size * 1.0, dtype='float64'
                ),
                name="simnet_acc",
            )
            return acc

        def get_loss(cos_q_pt, cos_q_nt):
            loss_op1 = paddle.subtract(
                fluid.layers.fill_constant_batch_size_like(
                    input=cos_q_pt, shape=[-1, 1], value=margin, dtype='float32'
                ),
                cos_q_pt,
            )
            loss_op2 = paddle.add(loss_op1, cos_q_nt)
            loss_op3 = paddle.maximum(
                fluid.layers.fill_constant_batch_size_like(
                    input=loss_op2, shape=[-1, 1], value=0.0, dtype='float32'
                ),
                loss_op2,
            )
            avg_cost = paddle.mean(loss_op3)
            return avg_cost

        is_distributed = False
        is_sparse = True

        # query
G
GGBond8488 已提交
70 71 72
        q = paddle.static.data(
            name="1", shape=[-1, 1], dtype="int64", lod_level=1
        )
L
lxsbupt 已提交
73 74 75 76 77
        # embedding
        q_emb = fluid.contrib.layers.sparse_embedding(
            input=q,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
78
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
79 80 81 82 83 84
                name="__emb__",
                learning_rate=emb_lr,
            ),
        )
        q_emb = paddle.reshape(q_emb, [-1, emb_dim])
        # vsum
85 86 87
        q_sum = paddle.static.nn.sequence_lod.sequence_pool(
            input=q_emb, pool_type='sum'
        )
L
lxsbupt 已提交
88 89
        q_ss = paddle.nn.functional.softsign(q_sum)
        # fc layer after conv
C
Charles-hit 已提交
90 91
        q_fc = paddle.static.nn.fc(
            x=q_ss,
L
lxsbupt 已提交
92
            size=hid_dim,
C
Charles-hit 已提交
93
            weight_attr=fluid.ParamAttr(
94
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
95 96 97 98 99
                name="__q_fc__",
                learning_rate=base_lr,
            ),
        )
        # label data
G
GGBond8488 已提交
100
        label = paddle.static.data(name="label", shape=[-1, 1], dtype="int64")
L
lxsbupt 已提交
101
        # pt
G
GGBond8488 已提交
102 103 104
        pt = paddle.static.data(
            name="2", shape=[-1, 1], dtype="int64", lod_level=1
        )
L
lxsbupt 已提交
105 106 107 108 109
        # embedding
        pt_emb = fluid.contrib.layers.sparse_embedding(
            input=pt,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
110
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
111 112 113 114 115 116
                name="__emb__",
                learning_rate=emb_lr,
            ),
        )
        pt_emb = paddle.reshape(pt_emb, [-1, emb_dim])
        # vsum
117 118 119
        pt_sum = paddle.static.nn.sequence_lod.sequence_pool(
            input=pt_emb, pool_type='sum'
        )
L
lxsbupt 已提交
120 121
        pt_ss = paddle.nn.functional.softsign(pt_sum)
        # fc layer
C
Charles-hit 已提交
122 123
        pt_fc = paddle.static.nn.fc(
            x=pt_ss,
L
lxsbupt 已提交
124
            size=hid_dim,
C
Charles-hit 已提交
125
            weight_attr=fluid.ParamAttr(
126
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
127 128 129 130 131 132
                name="__fc__",
                learning_rate=base_lr,
            ),
            bias_attr=fluid.ParamAttr(name="__fc_b__"),
        )
        # nt
G
GGBond8488 已提交
133 134 135
        nt = paddle.static.data(
            name="3", shape=[-1, 1], dtype="int64", lod_level=1
        )
L
lxsbupt 已提交
136 137 138 139 140
        # embedding
        nt_emb = fluid.contrib.layers.sparse_embedding(
            input=nt,
            size=[dict_dim, emb_dim],
            param_attr=fluid.ParamAttr(
141
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
142 143 144 145 146 147
                name="__emb__",
                learning_rate=emb_lr,
            ),
        )
        nt_emb = paddle.reshape(nt_emb, [-1, emb_dim])
        # vsum
148 149 150
        nt_sum = paddle.static.nn.sequence_lod.sequence_pool(
            input=nt_emb, pool_type='sum'
        )
L
lxsbupt 已提交
151 152
        nt_ss = paddle.nn.functional.softsign(nt_sum)
        # fc layer
C
Charles-hit 已提交
153 154
        nt_fc = paddle.static.nn.fc(
            x=nt_ss,
L
lxsbupt 已提交
155
            size=hid_dim,
C
Charles-hit 已提交
156
            weight_attr=fluid.ParamAttr(
157
                initializer=paddle.nn.initializer.Constant(value=0.01),
L
lxsbupt 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
                name="__fc__",
                learning_rate=base_lr,
            ),
            bias_attr=fluid.ParamAttr(name="__fc_b__"),
        )
        cos_q_pt = paddle.nn.functional.cosine_similarity(q_fc, pt_fc)
        cos_q_nt = paddle.nn.functional.cosine_similarity(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)

    def get_dist_env(self):
        trainer_id = int(os.getenv('PADDLE_TRAINER_ID', '0'))
        trainer_endpoints = ''
        current_endpoint = ''
        num_trainers = 0
        if os.getenv('PADDLE_TRAINER_ENDPOINTS'):
            trainer_endpoints = os.getenv('PADDLE_TRAINER_ENDPOINTS')
            current_endpoint = trainer_endpoints.split(',')[trainer_id]
            num_trainers = len(trainer_endpoints.split(','))

        return {
            'trainer_id': trainer_id,
            'num_trainers': num_trainers,
            'current_endpoint': current_endpoint,
            'trainer_endpoints': trainer_endpoints,
        }

    def test_SingleProcessMultiThread(self):
        """
        Testcase for SingleProcessMultiThread
        """
        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"
        os.environ["PADDLE_FUSE_ALLREDUCE"] = "1"
        os.environ["PADDLE_LOSS_SCALE"] = "1"

        startup_program = fluid.Program()
        main_program = fluid.Program()
        with fluid.program_guard(main_program, startup_program):
            with fluid.unique_name.guard():
                loss, acc, _ = self.net()
        optimizer = paddle.fluid.optimizer.Adam(learning_rate=0.01)
        optimizer.minimize(loss)
        print("===main_program====")
        print(main_program)
        print("===main_program====")
        from paddle.fluid.transpiler.collective import SingleProcessMultiThread

        t = SingleProcessMultiThread()
        env = self.get_dist_env()
        t.transpile(
            startup_program=startup_program,
            main_program=main_program,
            rank=env["trainer_id"],
            endpoints=env["trainer_endpoints"],
            current_endpoint=env['current_endpoint'],
            wait_port=False,
        )
        param_cnt = t._get_update_param_count()
        print("param_cnt:", param_cnt)


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