Platform For AI:EasyRec

背景情報

次の図は、EasyRecプロセッサに基づくレコメンデーションエンジンのアーキテクチャを示しています。

EasyRecプロセッサには、次のモジュールが含まれます。

• Item Feature Cache: このモジュールはFeatureStoreの機能をメモリにキャッシュします。これにより、頻繁なリクエスト操作によるFeatureStoreの負荷が軽減されます。 アイテム機能キャッシュは、リアルタイム機能更新などの増分更新をサポートします。

• Feature Generator (FG): このモジュールでは、リアルタイムおよびオフラインの機能エンジニアリングに同じ実装を使用して、一貫性を確保します。 FGは淘宝網からの豊富な経験に基づいて設計されています。

• TFModel: このモジュールは、TensorFlowを使用してEasyRecプロセッサによってエクスポートされたSavedModelをロードし、Bladeを使用してCPUとGPUの両方で推論の最適化を行います。

• 機能追跡と増分更新: ほとんどの場合、これらのモジュールはリアルタイムトレーニングに使用されます。 詳細については、「Online Deep Learning」をご参照ください。

制限事項

EasyRecプロセッサは、T4、A10、3090、4090タイプのGPUデバイス、およびIntel CPUを使用するg6、g7、g8などの汎用Elastic Compute Service (ECS) インスタンスファミリーで使用できます。

プロセッサのバージョン

EasyRecプロセッサは継続的に改善されています。 それ以降のバージョンでは、機能と推論パフォーマンスが強化されます。 最適な結果を得るには、最新バージョンを使用して推論サービスをデプロイすることを推奨します。 次の表に、リリースされたバージョンとその基本情報を示します。

手順1: サービスのデプロイ

EASCMDクライアントを使用してEasyRecモデルサービスをデプロイする場合、プロセッサタイプをeasyrec-{version} に設定する必要があります。 詳細については、「EASCMDまたはDSWを使用したモデルサービスのデプロイ」をご参照ください。 次のコードでは、構成ファイルの例を示します。

サンプルコード (FGが有効)

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>@<DomainName>:<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
# Run the deployment command. 
eascmd  create echo.json
# eascmd -i <AccessKeyID>  -k  <AccessKeySecret>   -e <endpoint> create echo.json
# Run the update command.
eascmd update ali_rec_rnk_with_fg -s echo.json

サンプルコード (FGが無効の場合)

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
# Run the deployment command. 
eascmd  create echo.json
# eascmd -i <AccessKeyID>  -k  <AccessKeySecret>   -e <endpoint> create echo.json
# Run the update command.
eascmd update ali_rec_rnk_no_fg -s echo.json

次の表に、主要なパラメーターを示します。 その他のパラメーターについては、「モデルサービスのパラメーター」をご参照ください。

EasyRecプロセッサの推論最適化に使用するパラメータ

ステップ2: コールサービス

EasyRecモデルサービスをデプロイした後、Elastic Algorithm service (EAS) ページに移動します。 このページで、[サービスの種類] 列の [呼び出し方法] をクリックして、サービスのエンドポイントとトークンを表示します。

EasyRecモデルサービスの入力と出力は、プロトコルバッファ (protobuf) 形式です。 FGが有効かどうかに基づいて、次の方法でサービスを呼び出すことができます。

FGが有効な場合のサンプルコード (fg_mode=tf)

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

PredictClient client = new PredictClient(new HttpConfig());
// Specify the endpoint of the service that you want to call. The endpoint starts with your user ID. 
client.setEndpoint("xxxxxxx.vpc.cn-hangzhou.pai-eas.aliyuncs.com");
client.setModelName("ali_rec_rnk_with_fg");
// Specify the token of the service. 
client.setToken("******");

EasyRecRequest easyrecRequest = new EasyRecRequest(separator);
// userFeatures: Specify multiple user features at the same time. Separate multiple user features with \u0002 (CTRL_B). For each user feature, separate the feature name and feature value with a colon (:). 
//  user_fea0:user_fea0_val\u0002user_fea1:user_fea1_val
// For more information about the feature value format, visit https://easyrec.readthedocs.io/en/latest/feature/rtp_fg.html.
easyrecRequest.appendUserFeatureString(userFeatures);
// Alternatively, add one user feature at a time.
// easyrecRequest.addUserFeature(String userFeaName, T userFeaValue). 
// T: the type of the feature value. Valid values: String, float, long, and int. 

// contextFeatures: Specify multiple context features at the same time. Separate multiple context features with \u0002 (CTRL_B). For each context feature, separate the feature name and feature value with a colon (:). 
//   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);
// Alternatively, add one context feature at a time.
// easyrecRequest.addContextFeature(String ctxtFeaName, List<Object> ctxtFeaValue). 
// Valid data types of ctxtFeaValue: String, Float, Long, and Integer. 

// itemIdStr: the list of item IDs to be predicted. Separate multiple item IDs with commas (,). 
easyrecRequest.appendItemStr(itemIdStr, ",");
// Alternatively, add one item ID at a time.
// 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));
    }
}

// Obtain the features processed by FG to compare with the offline features. 
// Set DebugLevel to 1 to return the generated features. 
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)

FGが無効の場合のサンプルコード (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});
        // Note: If you set INPUT_TILE to 2, you can simplify the code in the following way:
        //    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());

        // Call setDirectEndpoint to access the service by using a virtual private cloud (VPC) direct connection channel. 
        //   client.setDirectEndpoint("pai-eas-vpc.cn-shanghai.aliyuncs.com");
        // You need to create a VPC direct connection channel on the EAS page of the PAI console. 
        // Compared with using a gateway, using the direct connection channel improves stability and performance. 
        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: the name of the output field. You can run the cURL command to view the input and output of the model.
                //   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') # Replace server_dafault with the signature_name of the actual model. For more information, see https://www.alibabacloud.com/help/en/pai/user-guide/sdk-for-python 
    req.add_feed('user_id', [3], TFRequest.DT_STRING, ['u0001'] * 3)
    req.add_feed('age', [3], TFRequest.DT_FLOAT, [18.0] * 3)
    # Note: If you set INPUT_TILE to 2, you can simplify the code in the following way:
    #   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)

カスタムサービス要求を作成することもできます。 詳細については、「リクエスト構文」をご参照ください。

リクエスト構文

Python以外のクライアントの場合は、. protoファイルから手動で予測コードを生成する必要があります。次のprotobuf定義を使用して、カスタムサービス要求のコードを生成します。

• tf_predict.proto: TensorFlowモデルのprotobuf定義

  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モデルとFGのprotobuf定義

  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;
  }