人工智能平台 PAI：EasyRec Processor（推荐打分服务）

背景信息

基于EasyRec Processor的推荐引擎的架构图如下所示：

其中EasyRec Processor主要包含以下模块：

• Item Feature Cache：将FeatureStore里面的特征缓存到内存中，可以减少请求FeatureStore带来的网络开销和压力。此外，Item特征缓存支持增量更新，例如实时特征的更新。

• Feature Generator：特征工程模块（FG）采用相同的实现保证了离线和在线特征处理的一致性。 特征工程的实现借鉴于淘宝沉淀的特征工程方案。

• TFModel：TensorFlow模型加载EasyRec导出的Saved_Model，并结合Blade做模型在CPU和GPU上的推理优化。

• 特征埋点和模型增量更新模块：通常应用于实时训练场景，详情请参见实时训练。

使用限制

仅支持使用通用型实例规格族g6、g7或g8机型（仅支持Intel系列的CPU），支持T4、A10、3090或4090等GPU型号，详情请参见通用型（g系列）。

版本列表

EasyRec Processor仍然在迭代中，建议使用最新的版本部署推理服务，新的版本将提供更多的功能和更高的推理性能，已经发布的版本：

步骤一：部署服务

使用eascmd客户端部署EasyRec模型服务时，您需要指定Processor种类为easyrec-{version}，关于如何使用客户端工具部署服务，详情请参见服务部署：EASCMD或DSW。服务配置文件示例如下：

使用FG的示例

bizdate=$1
cat << EOF > echo.json
{
  "name":"ali_rec_rnk_with_fg",
  "metadata": {
    "instance": 2,
    "rpc": {
      "enable_jemalloc": 1,
      "max_queue_size": 100
    }
  },
  "cloud": {
    "computing": {
      "instance_type": "ecs.g7.large"",
      "instances": null
    }
  },
  "model_config": {
    "remote_type": "hologres",
    "url": "postgresql://<AccessKeyID>:<AccessKeySecret>@<域名>:<port>/<database>",
    "tables": [{"name":"<schema>.<table_name>","key":"<index_column_name>","value": "<column_name>"}],
    "period": 2880,
    "fg_mode": "tf",
    "outputs":"probs_ctr,probs_cvr",
  },
  "model_path": "",
  "processor": "easyrec-1.9",
  "storage": [
    {
      "mount_path": "/home/admin/docker_ml/workspace/model/",
      "oss": {
        "path": "oss://easyrec/ali_rec_sln_acc_rnk/20221122/export/final_with_fg"
      }
    }
  ]
}

EOF
# 执行部署命令。
eascmd  create echo.json
# eascmd -i <AccessKeyID>  -k  <AccessKeySecret>   -e <endpoint> create echo.json
# 执行更新命令
eascmd update ali_rec_rnk_with_fg -s echo.json

不使用FG的示例

bizdate=$1
cat << EOF > echo.json
{
  "name":"ali_rec_rnk_no_fg",
  "metadata": {
    "instance": 2,
    "rpc": {
      "enable_jemalloc": 1,
      "max_queue_size": 100
    }
  },
  "cloud": {
    "computing": {
      "instance_type": "ecs.g7.large"",
      "instances": null
    }
  },
  "model_config": {
    "fg_mode": "bypass"
  },
  "processor": "easyrec-1.9",
  "processor_envs": [
    {
      "name": "INPUT_TILE",
      "value": "2"
    }
  ],
  "storage": [
    {
      "mount_path": "/home/admin/docker_ml/workspace/model/",
      "oss": {
        "path": "oss://easyrec/ali_rec_sln_acc_rnk/20221122/export/final/"
      }
    }
  ],
  "warm_up_data_path": "oss://easyrec/ali_rec_sln_acc_rnk/rnk_warm_up.bin"
}

EOF
# 执行部署命令。
eascmd  create echo.json
# eascmd -i <AccessKeyID>  -k  <AccessKeySecret>   -e <endpoint> create echo.json
# 执行更新命令
eascmd update ali_rec_rnk_no_fg -s echo.json

其中关键参数说明如下，其他参数说明，请参见服务模型所有相关参数说明。

EasyRecProcessor的推理优化参数

步骤二：调用服务

EasyRec模型服务部署完成后，在模型在线服务（EAS）页面，单击待调用服务服务方式列下的调用信息，查看服务的访问地址和Token信息。

EasyRec模型服务的输入输出格式为Protobuf格式，根据是否包含FG，分为以下两种调用方法：

包含FG：fg_mode=tf

import com.aliyun.openservices.eas.predict.http.*;
import com.aliyun.openservices.eas.predict.request.EasyRecRequest;

PredictClient client = new PredictClient(new HttpConfig());
// 通过普通网关访问时，需要使用以用户UID开头的Endpoint，在EAS控制台服务的调用信息中可以获得该信息。
client.setEndpoint("xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com");
client.setModelName("ali_rec_rnk_with_fg");
// 替换为服务Token信息。
client.setToken("******");

EasyRecRequest easyrecRequest = new EasyRecRequest(separator);
// userFeatures: 用户特征, 特征之间用\u0002(CTRL_B)分隔, 特征名和特征值之间用:分隔。
//  user_fea0:user_fea0_val\u0002user_fea1:user_fea1_val
// 特征值的格式请参考: https://easyrec.readthedocs.io/en/latest/feature/rtp_fg.html
easyrecRequest.appendUserFeatureString(userFeatures);
// 也可以每次添加一个user特征:
// easyrecRequest.addUserFeature(String userFeaName, T userFeaValue)。
// 特征值的类型T: String, float, long, int。

// contextFeatures: context特征, 特征之间用\u0002(CTRL_B)分隔, 特征名和特征值之间用:分割, 特征值和特征值之间用:分隔。
//   ctxt_fea0:ctxt_fea0_ival0:ctxt_fea0_ival1:ctxt_fea0_ival2\u0002ctxt_fea1:ctxt_fea1_ival0:ctxt_fea1_ival1:ctxt_fea1_ival2
easyrecRequest.appendContextFeatureString(contextFeatures);
// 也可以每次添加一个context特征：
// easyrecRequest.addContextFeature(String ctxtFeaName, List<Object> ctxtFeaValue)。
// ctxtFeaValue的类型: String, Float, Long, Integer。

// itemIdStr: 要预测的itemId的列表，以半角逗号（,）分割。
easyrecRequest.appendItemStr(itemIdStr, ",");
// 也可以每次添加一个itemId:
// easyrecRequest.appendItemId(String itemId)

PredictProtos.PBResponse response = client.predict(easyrecRequest);

for (Map.Entry<String, PredictProtos.Results> entry : response.getResultsMap().entrySet()) {
    String key = entry.getKey();
    PredictProtos.Results value = entry.getValue();
    System.out.print("key: " + key);
    for (int i = 0; i < value.getScoresCount(); i++) {
        System.out.format("value: %.6g\n", value.getScores(i));
    }
}

// 获取FG之后的特征，以便和离线的特征对比一致性。
// 将DebugLevel设置成1，即可返回生成的特征。
easyrecRequest.setDebugLevel(1);
PredictProtos.PBResponse response = client.predict(easyrecRequest);
Map<String, String> genFeas = response.getGenerateFeaturesMap();
for(String itemId: genFeas.keySet()) {
    System.out.println(itemId);
    System.out.println(genFeas.get(itemId));
}

from eas_prediction import PredictClient

from eas_prediction.easyrec_request import EasyRecRequest
from eas_prediction.easyrec_predict_pb2 import PBFeature
from eas_prediction.easyrec_predict_pb2 import PBRequest

if __name__ == '__main__':
    endpoint = 'http://xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com'
    service_name = 'ali_rec_rnk_with_fg'
    token = '******'

    client = PredictClient(endpoint, service_name)
    client.set_token(token)
    client.init()

    req = PBRequest()
    uid = PBFeature()
    uid.string_feature = 'u0001'
    req.user_features['user_id'] = uid
    age = PBFeature()
    age.int_feature = 12
    req.user_features['age'] = age
    weight = PBFeature()
    weight.float_feature = 129.8
    req.user_features['weight'] = weight

    req.item_ids.extend(['item_0001', 'item_0002', 'item_0003'])
    
    easyrec_req = EasyRecRequest()
    easyrec_req.add_feed(req, debug_level=0)
    res = client.predict(easyrec_req)
    print(res)

不包含FG：fg_mode=bypass

import java.util.List;

import com.aliyun.openservices.eas.predict.http.PredictClient;
import com.aliyun.openservices.eas.predict.http.HttpConfig;
import com.aliyun.openservices.eas.predict.request.TFDataType;
import com.aliyun.openservices.eas.predict.request.TFRequest;
import com.aliyun.openservices.eas.predict.response.TFResponse;

public class TestEasyRec {
    public static TFRequest buildPredictRequest() {
        TFRequest request = new TFRequest();
 
        request.addFeed("user_id", TFDataType.DT_STRING, 
                        new long[]{5}, new String []{ "u0001", "u0001", "u0001"});
      	request.addFeed("age", TFDataType.DT_FLOAT, 
                        new long[]{5}, new float []{ 18.0f, 18.0f, 18.0f});
        // 注意: 如果设置了INPUT_TILE=2，那么上述值都相同的feature可以只传一次:
        //    request.addFeed("user_id", TFDataType.DT_STRING,
        //            new long[]{1}, new String []{ "u0001" });
        //    request.addFeed("age", TFDataType.DT_FLOAT, 
        //            new long[]{1}, new float []{ 18.0f});
      	request.addFeed("item_id", TFDataType.DT_STRING, 
                        new long[]{5}, new String []{ "i0001", "i0002", "i0003"});  
        request.addFetch("probs");
      	return request;
    }

    public static void main(String[] args) throws Exception {
        PredictClient client = new PredictClient(new HttpConfig());

        // 如果要使用网络直连功能，需使用setDirectEndpoint方法, 如: 
        //   client.setDirectEndpoint("pai-eas-vpc.cn-shanghai.aliyuncs.com");
        // 网络直连需打通在EAS控制台开通，提供用于访问EAS服务的源vswitch，
        // 网络直连具有更好的稳定性和性能。
        client.setEndpoint("xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com");
        client.setModelName("ali_rec_rnk_no_fg");
        client.setToken("");
        long startTime = System.currentTimeMillis();
        for (int i = 0; i < 100; i++) {
            try {
                TFResponse response = client.predict(buildPredictRequest());
                // probs为模型的输出的字段名, 可以使用curl命令查看模型的输入输出:
                //   curl xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com -H "Authorization:{token}"
                List<Float> result = response.getFloatVals("probs");
                System.out.print("Predict Result: [");
                for (int j = 0; j < result.size(); j++) {
                    System.out.print(result.get(j).floatValue());
                    if (j != result.size() - 1) {
                        System.out.print(", ");
                    }
                }
                System.out.print("]\n");
            } catch (Exception e) {
                e.printStackTrace();
            }
        }
        long endTime = System.currentTimeMillis();
        System.out.println("Spend Time: " + (endTime - startTime) + "ms");
        client.shutdown();
    }
}

#!/usr/bin/env python

from eas_prediction import PredictClient
from eas_prediction import StringRequest
from eas_prediction import TFRequest

if __name__ == '__main__':
    client = PredictClient('http://xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com', 'ali_rec_rnk_no_fg')
    client.set_token('')
    client.init()

    req = TFRequest('server_default') # 注意请将 server_dafault 替换为真实模型的 signature_name，详细见上文的使用说明文档
    req.add_feed('user_id', [3], TFRequest.DT_STRING, ['u0001'] * 3)
    req.add_feed('age', [3], TFRequest.DT_FLOAT, [18.0] * 3)
    # 注意: 开启INPUT_TILE=2的优化之后, 上述特征可以只传一个值
    #   req.add_feed('user_id', [1], TFRequest.DT_STRING, ['u0001'])
    #   req.add_feed('age', [1], TFRequest.DT_FLOAT, [18.0])
    req.add_feed('item_id', [3], TFRequest.DT_STRING, 
        ['i0001', 'i0002', 'i0003'])
    for x in range(0, 100):
        resp = client.predict(req)
        print(resp)

您也可以自行构建服务请求，详情请参见请求格式。

请求格式

除Python外，使用其他语言客户端调用服务都需要根据.proto文件手动生成预测的请求代码文件。如果您希望自行构建服务请求，则可以参考如下protobuf的定义来生成相关的代码：

• tf_predict.proto: tensorflow模型的请求定义

  syntax = "proto3";
  
  option cc_enable_arenas = true;
  option go_package = ".;tf";
  option java_package = "com.aliyun.openservices.eas.predict.proto";
  option java_outer_classname = "PredictProtos";
  
  enum ArrayDataType {
    // Not a legal value for DataType. Used to indicate a DataType field
    // has not been set.
    DT_INVALID = 0;
  
    // Data types that all computation devices are expected to be
    // capable to support.
    DT_FLOAT = 1;
    DT_DOUBLE = 2;
    DT_INT32 = 3;
    DT_UINT8 = 4;
    DT_INT16 = 5;
    DT_INT8 = 6;
    DT_STRING = 7;
    DT_COMPLEX64 = 8;  // Single-precision complex
    DT_INT64 = 9;
    DT_BOOL = 10;
    DT_QINT8 = 11;     // Quantized int8
    DT_QUINT8 = 12;    // Quantized uint8
    DT_QINT32 = 13;    // Quantized int32
    DT_BFLOAT16 = 14;  // Float32 truncated to 16 bits.  Only for cast ops.
    DT_QINT16 = 15;    // Quantized int16
    DT_QUINT16 = 16;   // Quantized uint16
    DT_UINT16 = 17;
    DT_COMPLEX128 = 18;  // Double-precision complex
    DT_HALF = 19;
    DT_RESOURCE = 20;
    DT_VARIANT = 21;  // Arbitrary C++ data types
  }
  
  // Dimensions of an array
  message ArrayShape {
    repeated int64 dim = 1 [packed = true];
  }
  
  // Protocol buffer representing an array
  message ArrayProto {
    // Data Type.
    ArrayDataType dtype = 1;
  
    // Shape of the array.
    ArrayShape array_shape = 2;
  
    // DT_FLOAT.
    repeated float float_val = 3 [packed = true];
  
    // DT_DOUBLE.
    repeated double double_val = 4 [packed = true];
  
    // DT_INT32, DT_INT16, DT_INT8, DT_UINT8.
    repeated int32 int_val = 5 [packed = true];
  
    // DT_STRING.
    repeated bytes string_val = 6;
  
    // DT_INT64.
    repeated int64 int64_val = 7 [packed = true];
  
    // DT_BOOL.
    repeated bool bool_val = 8 [packed = true];
  }
  
  // PredictRequest specifies which TensorFlow model to run, as well as
  // how inputs are mapped to tensors and how outputs are filtered before
  // returning to user.
  message PredictRequest {
    // A named signature to evaluate. If unspecified, the default signature
    // will be used
    string signature_name = 1;
  
    // Input tensors.
    // Names of input tensor are alias names. The mapping from aliases to real
    // input tensor names is expected to be stored as named generic signature
    // under the key "inputs" in the model export.
    // Each alias listed in a generic signature named "inputs" should be provided
    // exactly once in order to run the prediction.
    map<string, ArrayProto> inputs = 2;
  
    // Output filter.
    // Names specified are alias names. The mapping from aliases to real output
    // tensor names is expected to be stored as named generic signature under
    // the key "outputs" in the model export.
    // Only tensors specified here will be run/fetched and returned, with the
    // exception that when none is specified, all tensors specified in the
    // named signature will be run/fetched and returned.
    repeated string output_filter = 3;
    
    // Debug flags
    // 0: just return prediction results, no debug information
    // 100: return prediction results, and save request to model_dir 
    // 101: save timeline to model_dir
    int32 debug_level = 100;
  }
  
  // Response for PredictRequest on successful run.
  message PredictResponse {
    // Output tensors.
    map<string, ArrayProto> outputs = 1;
  }

• easyrec_predict.proto: Tensorflow模型+FG的请求定义

  syntax = "proto3";
  
  option cc_enable_arenas = true;
  option go_package = ".;easyrec";
  option java_package = "com.aliyun.openservices.eas.predict.proto";
  option java_outer_classname = "EasyRecPredictProtos";
  
  import "tf_predict.proto";
  
  // context features
  message ContextFeatures {
    repeated PBFeature features = 1;
  }
  
  message PBFeature {
    oneof value {
      int32 int_feature = 1;
      int64 long_feature = 2;
      string string_feature = 3;
      float float_feature = 4;
    }
  }
  
  // PBRequest specifies the request for aggregator
  message PBRequest {
    // Debug flags
    // 0: just return prediction results, no debug information
    // 3: return features generated by FG module, string format, feature values are separated by \u0002, 
    //    could be used for checking feature consistency check and generating online deep learning samples 
    // 100: return prediction results, and save request to model_dir 
    // 101: save timeline to model_dir
    // 102: for recall models such as DSSM and MIND, only only return Faiss retrieved results
    //      but also return user embedding vectors.
    int32 debug_level = 1;
  
    // user features
    map<string, PBFeature> user_features = 2;
  
    // item ids, static(daily updated) item features 
    // are fetched from the feature cache resides in 
    // each processor node by item_ids
    repeated string item_ids = 3;
  
    // context features for each item, realtime item features
    //    could be passed as context features.
    map<string, ContextFeatures> context_features = 4;
  
    // embedding retrieval neighbor number.
    int32 faiss_neigh_num = 5;
  }
  
  // return results
  message Results {
    repeated double scores = 1 [packed = true];
  }
  
  enum StatusCode {
    OK = 0;
    INPUT_EMPTY = 1;
    EXCEPTION = 2;
  }
  
  // PBResponse specifies the response for aggregator
  message PBResponse {
    // results
    map<string, Results> results = 1;
  
    // item features
    map<string, string> item_features = 2;
  
    // fg generate features
    map<string, string> generate_features = 3;
  
    // context features
    map<string, ContextFeatures> context_features = 4;
  
    string error_msg = 5;
  
    StatusCode status_code = 6;
  
    // item ids
    repeated string item_ids = 7;
  
    repeated string outputs = 8;
  
    // all fg input features
    map<string, string> raw_features = 9;
  
    // output tensors
    map<string, ArrayProto> tf_outputs = 10;
  }