ad_pai_models/widedeep_v13_3.py

#! /usr/bin/env python
# -*- coding: utf-8 -*-
# vim:fenc=utf-8
#
# Copyright © 2025 StrayWarrior <i@straywarrior.com>
"""
1.删除容易导致偏差的viewall特征
2.删除分桶不均匀的cpa特征
3.减少dense特征
4.增加U-I交叉统计
5.增加线性部分dense
6.减少wide部分embedding
7.减少部分bucket size
8.使用protobuf
9.调整embedding variable (PAI平台目前还不支持)
"""

import re
from easy_rec.python.protos.pipeline_pb2 import EasyRecConfig
from easy_rec.python.protos.train_pb2 import TrainConfig
from easy_rec.python.protos.eval_pb2 import EvalConfig, AUC, EvalMetrics
from easy_rec.python.protos.dataset_pb2 import DatasetConfig
from easy_rec.python.protos.feature_config_pb2 import FeatureConfig, FeatureGroupConfig, WideOrDeep, EVParams
from easy_rec.python.protos.easy_rec_model_pb2 import EasyRecModel
from easy_rec.python.protos.deepfm_pb2 import DeepFM
from easy_rec.python.protos.dnn_pb2 import DNN
from easy_rec.python.protos.export_pb2 import ExportConfig
from easy_rec.python.protos.optimizer_pb2 import Optimizer, AdamOptimizer, LearningRate, ConstantLearningRate
from google.protobuf import text_format

raw_input = open("data_fields_v3.config").readlines()
input_fields = dict(
    map(lambda x: (x[0], x[1]),
        map(lambda x: x.strip().split(' '), raw_input)))

def read_features(filename, excludes=None):
    features = open(filename).readlines()
    features = [name.strip().lower() for name in features]
    if excludes:
        for x in excludes:
            if x in features:
                features.remove(x)
    return features

exclude_features = ['viewall', 'cpa']

dense_features = read_features("features_top300.config", exclude_features)
top_dense_features = read_features('features_top50.config', exclude_features)

sparse_features = [
    "cid", "adid", "adverid",
    "region", "city", "brand",
    "vid", "cate1", "cate2",
    "apptype", "hour", "hour_quarter", "root_source_scene", "root_source_channel", "is_first_layer", "title_split", "user_has_conver_1y",
    "user_adverid_view_3d", "user_adverid_view_7d", "user_adverid_view_30d",
    "user_adverid_click_3d", "user_adverid_click_7d", "user_adverid_click_30d",
    "user_adverid_conver_3d", "user_adverid_conver_7d", "user_adverid_conver_30d",
    "user_skuid_view_3d", "user_skuid_view_7d", "user_skuid_view_30d",
    "user_skuid_click_3d", "user_skuid_click_7d", "user_skuid_click_30d",
    "user_skuid_conver_3d", "user_skuid_conver_7d", "user_skuid_conver_30d"
]
tag_features = [
    "user_vid_return_tags_2h", "user_vid_return_tags_1d", "user_vid_return_tags_3d",
    "user_vid_return_tags_7d", "user_vid_return_tags_14d"
]
seq_features = [
    "user_cid_click_list", "user_cid_conver_list"
]

input_type_map = {
    'BIGINT': DatasetConfig.FieldType.INT64,
    'DOUBLE': DatasetConfig.FieldType.DOUBLE,
    'STRING': DatasetConfig.FieldType.STRING
}

use_ev_features = [
    "cid", "adid", "adverid", "vid"
]

bucket_size_map = {
    'adverid': 100000,
    'region': 1000,
    'city': 10000,
    'brand': 10000,
    'cate1': 10000,
    'cate2': 10000,
    'apptype': 1000,
    'hour': 1000, # 实际上可以直接指定词表
    'hour_quarter': 4000,
    'root_source_scene': 100,
    'root_source_channel': 1000,
    'is_first_layer': 100,
    'user_has_conver_1y': 100,
}

def create_config():
    config = EasyRecConfig()

    # 训练配置
    train_config = TrainConfig()

    # 配置多个优化器
    optimizers = [
        (0.0010, False), # wide参数
        (0.0006, False), # dense参数
        (0.002, False)   # deep embedding参数
    ]

    for lr, use_moving_avg in optimizers:
        optimizer = Optimizer()
        adam_optimizer = AdamOptimizer()
        learning_rate = LearningRate()
        constant_lr = ConstantLearningRate()
        constant_lr.learning_rate = lr
        learning_rate.constant_learning_rate.CopyFrom(constant_lr)
        adam_optimizer.learning_rate.CopyFrom(learning_rate)
        optimizer.adam_optimizer.CopyFrom(adam_optimizer)
        optimizer.use_moving_average = use_moving_avg
        train_config.optimizer_config.append(optimizer)

    train_config.num_steps = 200000
    train_config.sync_replicas = True
    train_config.save_checkpoints_steps = 1100
    train_config.log_step_count_steps = 100
    train_config.save_summary_steps = 100
    config.train_config.CopyFrom(train_config)

    # 评估配置
    eval_config = EvalConfig()
    metrics_set = EvalMetrics()
    metrics_set.auc.SetInParent()
    eval_config.metrics_set.append(metrics_set)
    eval_config.eval_online = True
    eval_config.eval_interval_secs = 120
    config.eval_config.CopyFrom(eval_config)

    # 数据配置
    data_config = DatasetConfig()
    data_config.batch_size = 512
    data_config.num_epochs = 1
    data_config.prefetch_size = 32
    data_config.input_type = DatasetConfig.InputType.OdpsInputV2

    # 添加输入字段
    for name in input_fields:
        input_field = DatasetConfig.Field()
        input_field.input_name = name
        input_field.input_type = input_type_map[input_fields[name]]
        if name in dense_features:
            input_field.default_val = "0"
        data_config.input_fields.append(input_field)

    # 添加标签字段
    data_config.label_fields.append("has_conversion")
    config.data_config.CopyFrom(data_config)

    # 特征配置
    feature_configs = []

    # Dense特征配置
    boundaries = [ x / 100 for x in range(0, 101)]

    for name in dense_features:
        feature_config = FeatureConfig()
        feature_config.input_names.append(name)
        feature_config.feature_type = FeatureConfig.RawFeature
        feature_config.boundaries.extend(boundaries)
        feature_config.embedding_dim = 6
        feature_configs.append(feature_config)

    # Sparse特征配置
    for name in sparse_features:
        feature_config = FeatureConfig()
        feature_config.input_names.append(name)
        feature_config.feature_type = FeatureConfig.IdFeature
        # 只有INT64类型的特征才能使用embedding variable特性
        if name in use_ev_features:
            if input_type_map[input_fields[name]] != DatasetConfig.FieldType.INT64:
                raise ValueError(f"Feature {name} must be of type INT64 to use embedding variable.")
            feature_config.ev_params.filter_freq = 2
        else:
            feature_config.hash_bucket_size = bucket_size_map.get(name, 1000000)
        feature_config.embedding_dim = 6
        feature_configs.append(feature_config)

    # Tag特征配置
    for name in tag_features + seq_features:
        feature_config = FeatureConfig()
        feature_config.input_names.append(name)
        feature_config.feature_type = FeatureConfig.TagFeature
        feature_config.hash_bucket_size = bucket_size_map.get(name, 1000000)
        feature_config.embedding_dim = 6
        feature_config.separator = ','
        feature_configs.append(feature_config)

    config.feature_configs.extend(feature_configs)

    # 模型配置
    model_config = EasyRecModel()
    model_config.model_class = "DeepFM"

    # Wide特征组
    wide_group = FeatureGroupConfig()
    wide_group.group_name = 'wide'
    wide_group.feature_names.extend(dense_features + sparse_features)
    wide_group.wide_deep = WideOrDeep.WIDE
    model_config.feature_groups.append(wide_group)

    # Deep特征组
    deep_group = FeatureGroupConfig()
    deep_group.group_name = 'deep'
    deep_group.feature_names.extend(top_dense_features + sparse_features + tag_features + seq_features)
    deep_group.wide_deep = WideOrDeep.DEEP
    model_config.feature_groups.append(deep_group)

    # DeepFM配置
    deepfm = DeepFM()
    deepfm.wide_output_dim = 2

    # DNN配置
    dnn = DNN()
    dnn.hidden_units.extend([256, 128, 64])
    deepfm.dnn.CopyFrom(dnn)

    # Final DNN配置
    final_dnn = DNN()
    final_dnn.hidden_units.extend([64, 32])
    deepfm.final_dnn.CopyFrom(final_dnn)

    deepfm.l2_regularization = 1e-5

    model_config.deepfm.CopyFrom(deepfm)
    model_config.embedding_regularization = 1e-6
    config.model_config.CopyFrom(model_config)

    # 导出配置
    export_config = ExportConfig()
    export_config.exporter_type = "final"
    config.export_config.CopyFrom(export_config)

    return config

def merge_repeated_elements(msg_str, field_name):
    msg_str = re.sub(
        fr'( +{field_name}: [^\n]+\n)+',
        lambda m: '{}{}: [{}]\n'.format(
            m.group(0).split(field_name)[0],
            field_name,
            ', '.join(re.findall(fr'{field_name}: ([^\n]+)', m.group(0)))
        ),
        msg_str
    )
    return msg_str

def main():
    config = create_config()
    msg_str = text_format.MessageToString(config)
    msg_str = merge_repeated_elements(msg_str, 'boundaries')
    msg_str = merge_repeated_elements(msg_str, 'hidden_units')
    print(msg_str)

if __name__ == '__main__':
    main()