model.py 7.4 KB
Newer Older
Y
add din  
yaoxuefeng 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 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
import paddle.fluid as fluid
import math

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


class Model(ModelBase):
    def __init__(self, config):
        ModelBase.__init__(self, config)
    
    def config_read(self, config_path):
        with open(config_path, "r") as fin:
            user_count = int(fin.readline().strip())
            item_count = int(fin.readline().strip())
            cat_count = int(fin.readline().strip())
        return user_count, item_count, cat_count
    
    def din_attention(self, hist, target_expand, mask):
        """activation weight"""

        hidden_size = hist.shape[-1]

        concat = fluid.layers.concat(
            [hist, target_expand, hist - target_expand, hist * target_expand],
            axis=2)
        atten_fc1 = fluid.layers.fc(name="atten_fc1",
                                    input=concat,
                                    size=80,
                                    act=self.act,
                                    num_flatten_dims=2)
        atten_fc2 = fluid.layers.fc(name="atten_fc2",
                                    input=atten_fc1,
                                    size=40,
                                    act=self.act,
                                    num_flatten_dims=2)
        atten_fc3 = fluid.layers.fc(name="atten_fc3",
                                    input=atten_fc2,
                                    size=1,
                                    num_flatten_dims=2)
        atten_fc3 += mask
        atten_fc3 = fluid.layers.transpose(x=atten_fc3, perm=[0, 2, 1])
        atten_fc3 = fluid.layers.scale(x=atten_fc3, scale=hidden_size**-0.5)
        weight = fluid.layers.softmax(atten_fc3)
        out = fluid.layers.matmul(weight, hist)
        out = fluid.layers.reshape(x=out, shape=[0, hidden_size])
        return out
    
    def train_net(self):
        seq_len = -1
        self.item_emb_size = envs.get_global_env("hyper_parameters.item_emb_size", 64, self._namespace)
        self.cat_emb_size = envs.get_global_env("hyper_parameters.cat_emb_size", 64, self._namespace)
        self.act = envs.get_global_env("hyper_parameters.act", "sigmoid", self._namespace)
        #item_emb_size = 64
        #cat_emb_size = 64
        self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse", False, self._namespace)
        #significant for speeding up the training process
        self.config_path = envs.get_global_env("hyper_parameters.config_path", "data/config.txt", self._namespace)
        self.use_DataLoader = envs.get_global_env("hyper_parameters.use_DataLoader", False, self._namespace)
        user_count, item_count, cat_count = self.config_read(self.config_path)


        item_emb_attr = fluid.ParamAttr(name="item_emb")
        cat_emb_attr = fluid.ParamAttr(name="cat_emb")

        hist_item_seq = fluid.data(
            name="hist_item_seq", shape=[None, seq_len], dtype="int64")
        self._data_var.append(hist_item_seq)
        
        hist_cat_seq = fluid.data(
            name="hist_cat_seq", shape=[None, seq_len], dtype="int64")
        self._data_var.append(hist_cat_seq)
        
        target_item = fluid.data(name="target_item", shape=[None], dtype="int64")
        self._data_var.append(target_item)
        
        target_cat = fluid.data(name="target_cat", shape=[None], dtype="int64")
        self._data_var.append(target_cat)
        
        label = fluid.data(name="label", shape=[None, 1], dtype="float32")
        self._data_var.append(label)
        
        mask = fluid.data(name="mask", shape=[None, seq_len, 1], dtype="float32")
        self._data_var.append(mask)
        
        target_item_seq = fluid.data(
            name="target_item_seq", shape=[None, seq_len], dtype="int64")
        self._data_var.append(target_item_seq)
        
        target_cat_seq = fluid.data(
            name="target_cat_seq", shape=[None, seq_len], dtype="int64")
        self._data_var.append(target_cat_seq)

        if self.use_DataLoader:
            self._data_loader = fluid.io.DataLoader.from_generator(
                feed_list=self._data_var, capacity=10000, use_double_buffer=False, iterable=False)
        
        hist_item_emb = fluid.embedding(
            input=hist_item_seq,
            size=[item_count, self.item_emb_size],
            param_attr=item_emb_attr,
            is_sparse=self.is_sparse)

        hist_cat_emb = fluid.embedding(
            input=hist_cat_seq,
            size=[cat_count, self.cat_emb_size],
            param_attr=cat_emb_attr,
            is_sparse=self.is_sparse)

        target_item_emb = fluid.embedding(
            input=target_item,
            size=[item_count, self.item_emb_size],
            param_attr=item_emb_attr,
            is_sparse=self.is_sparse)

        target_cat_emb = fluid.embedding(
            input=target_cat,
            size=[cat_count, self.cat_emb_size],
            param_attr=cat_emb_attr,
            is_sparse=self.is_sparse)

        target_item_seq_emb = fluid.embedding(
            input=target_item_seq,
            size=[item_count, self.item_emb_size],
            param_attr=item_emb_attr,
            is_sparse=self.is_sparse)

        target_cat_seq_emb = fluid.embedding(
            input=target_cat_seq,
            size=[cat_count, self.cat_emb_size],
            param_attr=cat_emb_attr,
            is_sparse=self.is_sparse)

        item_b = fluid.embedding(
            input=target_item,
            size=[item_count, 1],
            param_attr=fluid.initializer.Constant(value=0.0))

        hist_seq_concat = fluid.layers.concat([hist_item_emb, hist_cat_emb], axis=2)
        target_seq_concat = fluid.layers.concat(
            [target_item_seq_emb, target_cat_seq_emb], axis=2)
        target_concat = fluid.layers.concat(
            [target_item_emb, target_cat_emb], axis=1)

        out = self.din_attention(hist_seq_concat, target_seq_concat, mask)
        out_fc = fluid.layers.fc(name="out_fc",
                                input=out,
                                size=self.item_emb_size + self.cat_emb_size,
                                num_flatten_dims=1)
        embedding_concat = fluid.layers.concat([out_fc, target_concat], axis=1)

        fc1 = fluid.layers.fc(name="fc1",
                            input=embedding_concat,
                            size=80,
                            act=self.act)
        fc2 = fluid.layers.fc(name="fc2", input=fc1, size=40, act=self.act)
        fc3 = fluid.layers.fc(name="fc3", input=fc2, size=1)
        logit = fc3 + item_b

        loss = fluid.layers.sigmoid_cross_entropy_with_logits(x=logit, label=label)
        
        avg_loss = fluid.layers.mean(loss)
        self._cost = avg_loss

        self.predict = fluid.layers.sigmoid(logit)
        predict_2d = fluid.layers.concat([1 - self.predict, self.predict], 1)
        label_int = fluid.layers.cast(label, 'int64')
        auc_var, batch_auc_var, _ = fluid.layers.auc(input=predict_2d,
                                                            label=label_int,
                                                            slide_steps=0)
        self._metrics["AUC"] = auc_var
        self._metrics["BATCH_AUC"] = batch_auc_var


    def optimizer(self):
        learning_rate = envs.get_global_env("hyper_parameters.learning_rate", None, self._namespace)
        optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
        return optimizer

    def infer_net(self, parameter_list):
        self.deepfm_net()