model.py 7.9 KB
Newer Older
F
frankwhzhang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
# 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.

import math
16
import numpy as np
F
frankwhzhang 已提交
17 18 19 20 21 22 23 24 25
import paddle.fluid as fluid

from paddlerec.core.utils import envs
from paddlerec.core.model import Model as ModelBase


class Model(ModelBase):
    def __init__(self, config):
        ModelBase.__init__(self, config)
F
frankwhzhang 已提交
26 27

    def _init_hyper_parameters(self):
F
frankwhzhang 已提交
28 29 30 31 32 33 34
        self.item_len = envs.get_global_env("hyper_parameters.self.item_len")
        self.hidden_size = envs.get_global_env("hyper_parameters.hidden_size")
        self.user_vocab = envs.get_global_env("hyper_parameters.user_vocab")
        self.item_vocab = envs.get_global_env("hyper_parameters.item_vocab")
        self.embed_size = envs.get_global_env("hyper_parameters.embed_size")

    def input_data(self, is_infer=False, **kwargs):
F
frankwhzhang 已提交
35 36 37 38 39 40 41
        user_slot_names = fluid.data(
            name='user_slot_names',
            shape=[None, 1],
            dtype='int64',
            lod_level=1)
        item_slot_names = fluid.data(
            name='item_slot_names',
42
            shape=[None, self.item_len],
F
frankwhzhang 已提交
43 44 45 46
            dtype='int64',
            lod_level=1)
        lens = fluid.data(name='lens', shape=[None], dtype='int64')
        labels = fluid.data(
47 48 49 50
            name='labels',
            shape=[None, self.item_len],
            dtype='int64',
            lod_level=1)
F
frankwhzhang 已提交
51 52

        inputs = [user_slot_names] + [item_slot_names] + [lens] + [labels]
53

F
frankwhzhang 已提交
54
        # demo: hot to use is_infer:
F
frankwhzhang 已提交
55
        if is_infer:
F
frankwhzhang 已提交
56
            return inputs
F
frankwhzhang 已提交
57
        else:
F
frankwhzhang 已提交
58
            return inputs
F
frankwhzhang 已提交
59 60

    def net(self, inputs, is_infer=False):
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77
        # user encode
        user_embedding = fluid.embedding(
            input=inputs[0],
            size=[self.user_vocab, self.embed_size],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Xavier(),
                regularizer=fluid.regularizer.L2Decay(1e-5)),
            is_sparse=True)

        user_feature = fluid.layers.fc(
            input=user_embedding,
            size=self.hidden_size,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.TruncatedNormal(
                    loc=0.0, scale=np.sqrt(1.0 / self.hidden_size))),
            bias_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant(
                value=0.0)),
F
frankwhzhang 已提交
78 79
            act='relu',
            name='user_feature_fc')
80 81 82 83 84 85 86 87
        # item encode
        item_embedding = fluid.embedding(
            input=inputs[1],
            size=[self.item_vocab, self.embed_size],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Xavier(),
                regularizer=fluid.regularizer.L2Decay(1e-5)),
            is_sparse=True)
F
frankwhzhang 已提交
88 89 90 91

        item_embedding = fluid.layers.sequence_unpad(
            x=item_embedding, length=inputs[2])

92 93 94 95 96 97 98 99
        item_fc = fluid.layers.fc(
            input=item_embedding,
            size=self.hidden_size,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.TruncatedNormal(
                    loc=0.0, scale=np.sqrt(1.0 / self.hidden_size))),
            bias_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant(
                value=0.0)),
F
frankwhzhang 已提交
100 101 102
            act='relu',
            name='item_fc')

103 104 105 106 107 108 109 110 111
        pos = self._fluid_sequence_get_pos(item_fc)
        pos_embed = fluid.embedding(
            input=pos,
            size=[self.user_vocab, self.embed_size],
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Xavier(),
                regularizer=fluid.regularizer.L2Decay(1e-5)),
            is_sparse=True)

F
frankwhzhang 已提交
112 113 114
        pos_embed = fluid.layers.squeeze(pos_embed, [1])

        # item gru
115 116 117
        gru_input = fluid.layers.fc(
            input=fluid.layers.concat([item_fc, pos_embed], 1),
            size=self.hidden_size * 3,
F
frankwhzhang 已提交
118 119
            name='item_gru_fc')

120 121 122 123 124 125 126 127 128 129 130 131
        # forward gru
        item_gru_forward = fluid.layers.dynamic_gru(
            input=gru_input,
            size=self.hidden_size,
            is_reverse=False,
            h_0=user_feature)
        # backward gru
        item_gru_backward = fluid.layers.dynamic_gru(
            input=gru_input,
            size=self.hidden_size,
            is_reverse=True,
            h_0=user_feature)
F
frankwhzhang 已提交
132 133 134 135

        item_gru = fluid.layers.concat(
            [item_gru_forward, item_gru_backward], axis=1)

136 137 138 139 140 141 142 143
        out_click_fc1 = fluid.layers.fc(
            input=item_gru,
            size=self.hidden_size,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.TruncatedNormal(
                    loc=0.0, scale=np.sqrt(1.0 / self.hidden_size))),
            bias_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant(
                value=0.0)),
F
frankwhzhang 已提交
144 145 146
            act='relu',
            name='out_click_fc1')

147 148 149 150
        click_prob = fluid.layers.fc(input=out_click_fc1,
                                     size=2,
                                     act='softmax',
                                     name='out_click_fc2')
F
frankwhzhang 已提交
151 152

        labels = fluid.layers.sequence_unpad(x=inputs[3], length=inputs[2])
153

F
frankwhzhang 已提交
154 155
        auc_val, batch_auc, auc_states = fluid.layers.auc(input=click_prob,
                                                          label=labels)
156

F
frankwhzhang 已提交
157 158 159
        if is_infer:
            self._infer_results["AUC"] = auc_val
            return
160

F
frankwhzhang 已提交
161 162 163 164 165 166
        loss = fluid.layers.reduce_mean(
            fluid.layers.cross_entropy(
                input=click_prob, label=labels))
        self._cost = loss
        self._metrics['auc'] = auc_val

F
frankwhzhang 已提交
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
    def _fluid_sequence_pad(self, input, pad_value, maxlen=None):
        """
        args:
            input: (batch*seq_len, dim)
        returns:
            (batch, max_seq_len, dim)
        """
        pad_value = fluid.layers.cast(
            fluid.layers.assign(input=np.array([pad_value], 'float32')),
            input.dtype)
        input_padded, _ = fluid.layers.sequence_pad(
            input, pad_value,
            maxlen=maxlen)  # (batch, max_seq_len, 1), (batch, 1)
        # TODO, maxlen=300, used to solve issues: https://github.com/PaddlePaddle/Paddle/issues/14164
        return input_padded

    def _fluid_sequence_get_pos(self, lodtensor):
        """
        args:
            lodtensor: lod = [[0,4,7]]
        return:
            pos: lod = [[0,4,7]]
                 data = [0,1,2,3,0,1,3]
                 shape = [-1, 1]
        """
        lodtensor = fluid.layers.reduce_sum(lodtensor, dim=1, keep_dim=True)
        assert lodtensor.shape == (-1, 1), (lodtensor.shape())
        ones = fluid.layers.cast(lodtensor * 0 + 1,
                                 'float32')  # (batch*seq_len, 1)
        ones_padded = self._fluid_sequence_pad(ones,
                                               0)  # (batch, max_seq_len, 1)
        ones_padded = fluid.layers.squeeze(ones_padded,
                                           [2])  # (batch, max_seq_len)
        seq_len = fluid.layers.cast(
            fluid.layers.reduce_sum(
                ones_padded, 1, keep_dim=True), 'int64')  # (batch, 1)
        seq_len = fluid.layers.squeeze(seq_len, [1])

        pos = fluid.layers.cast(
            fluid.layers.cumsum(
                ones_padded, 1, exclusive=True), 'int64')
        pos = fluid.layers.sequence_unpad(pos, seq_len)  # (batch*seq_len, 1)
        pos.stop_gradient = True
        return pos

    #def train_net(self):
    #    input_data = self.input_data()
    #    self.net(input_data)
F
frankwhzhang 已提交
215

F
frankwhzhang 已提交
216 217 218
    #def infer_net(self):
    #    input_data = self.input_data(is_infer=True)
    #    self.net(input_data, is_infer=True)