PAI-Rec engine fine-rank configuration - Platform For AI

Model configuration

In the fine rank stage, PAI-Rec obtains scores by calling an algorithm model service on Elastic Algorithm Service (EAS). EAS allows you to deploy various model types that PAI-Rec can call, including EasyRec, TorchEasyRec, TensorFlow, PMML, PS, and Alink.

The code for EasyRec is open-source. To use PAI-Rec for scoring, follow the EasyRec documentation to train a model, export the model, and deploy it as an EAS service by using the EasyRec Processor.

TorchEasyRec is the Torch version of EasyRec. To use it for scoring with PAI-Rec, follow the TorchEasyRec documentation to train a model, export the model, and deploy it as an EAS service by using the TorchEasyRec Processor.

Configuration example

{
    "AlgoConfs": [
        {
            "Name": "room3_mt_v2",
            "Type": "EAS",
            "EasConf": {
                "Processor": "EasyRec",
                "Timeout": 300,
                "ResponseFuncName": "easyrecMutValResponseFunc",
                "Url": "http:xxx.pai-eas.aliyuncs.com/api/predict/xxx",
                "EndpointType": "DIRECT",
                "Auth": ""
            }
        },
        {
            "Name": "room3_mt_v2_torch",
            "Type": "EAS",
            "EasConf": {
                "Processor": "EasyRec",
                "Timeout": 300,
                "ResponseFuncName": "torchrecMutValResponseFunc",
                "Url": "http:xxx.pai-eas.aliyuncs.com/api/predict/xxx",
                "EndpointType": "DIRECT",
                "Auth": ""
            }
        },
        {
            "Name": "tf",
            "Type": "EAS",
            "EasConf": {
                "Processor": "TensorFlow",
                "ResponseFuncName": "tfMutValResponseFunc",
                "Url": "http:xxx.pai-eas.aliyuncs.com/api/predict/xxx",
                "Auth": ""
            }
        },
        {
            "Name": "ps_smart",
            "Type": "EAS",
            "EasConf": {
                "Processor": "PMML",
                "ResponseFuncName": "pssmartResponseFunc",
                "Url": "http:xxx.pai-eas.aliyuncs.com/api/predict/xxx",
                "Auth": ""
            }
        },
        {
            "Name": "fm",
            "Type": "EAS",
            "EasConf": {
                "Processor": "ALINK_FM",
                "ResponseFuncName": "alinkFMResponseFunc",
                "Url": "http:xxx.pai-eas.aliyuncs.com/api/predict/xxx",
                "Auth": ""
            }
        }
    ]
}

Parameter	Type	Required	Description
Name	string	Yes	A custom name for the model information. This name is referenced in `RankConf`.
Type	string	Yes	The model deployment type.

EasConf

Parameter	Type	Required	Description
Processor	string	Yes	The model type. Valid values: EasyRec TensorFlow ALINK_FM PMML TFServing
Timeout	int	No	The timeout period for model requests, in milliseconds.
ResponseFuncName	string	Yes	The function used to parse the response data. This function must correspond to the `Processor` type. EasyRec For the TensorFlow version, use `easyrecMutValResponseFunc`. For the Torch version, use `torchrecMutValResponseFunc`. `TensorFlow`: `tfMutValResponseFunc` `ALINK_FM`: `alinkFMResponseFunc` `PMML`: `pssmartResponseFunc` `TFServing`: `tfServingResponseFunc`
Url	string	Yes	The endpoint URL of the model.
EndpointType	string	No	The endpoint type. Valid values: `DOCKER` and `DIRECT`. If both the engine and the model are deployed on EAS, set this parameter to `DIRECT` to connect directly to the model for higher performance.
Auth	string	Yes	The token for model authentication.

The easyrecMutValResponseFunc, torchrecMutValResponseFunc, and tfMutValResponseFunc functions can return multiple target prediction values. You can configure operations in PAI-Rec to process these values. For example, you can sum all returned values to generate the final ranking score.

If the model is deployed on EAS, you can find the Url and Auth values in the PAI-EAS console.

In the PAI-EAS console, go to the Inference Service page. Find the target service and click the Invocation Information link. In the dialog box that appears, select the VPC Address Call tab to obtain the Url. The Auth value corresponds to the token field.

Algorithm configuration

The algorithm configuration corresponds to RankConf in the configuration overview. RankConf is a Map[string]object structure. The key is the scenario name, such as ${scene_name1} and ${scene_name2} in the example. You can select different models for each scenario by using the custom model names defined in AlgoConfs.

Configuration example

{
    "RankConf": {
        "${scene_name1}": {
            "RankAlgoList": [
                "fm"
            ],
            "RankScore": "${fm}"
        },
        "${scene_name2}": {
            "RankAlgoList": [
                "dbmtl_v1"
            ],
            "RankScore": "${dbmtl_v1_probs_ctr} * (${dbmtl_v1_probs_view} + ${dbmtl_v1_y_view_time} + 0.5 * ${dbmtl_v1_probs_on_wheat} + 0.5 * ${dbmtl_v1_y_on_wheat_time} + 0.25 * ${dbmtl_v1_y_gift} + 0.25 * ${dbmtl_v1_probs_add_friend} + 0.1)",
            "Processor": "EasyRec"
        }
    }
}

Parameter	Type	Required	Description
RankAlgoList	[]string	Yes	An array of model names defined in `AlgoConfs`. Specifies which models to use for scoring in this scenario.
RankScore	string	Yes	The model fusion expression.
Processor	string	No	Required and must be set to `EasyRec` if any model in `RankAlgoList` uses the `EasyRec` processor.
BatchCount	int	No	The number of items to send in a single batch request to the model. The default value is 100.
ContextFeatures	[]string	No	A list of context features to pass to the processor. This parameter applies only to `EasyRec` processors.
ItemFeatures	[]string	No	A list of item features to pass to the EasyRec processor, which overrides the cached item features in the processor. This parameter applies only to `EasyRec` processors. To pass all item features, set the value to `["*"]`. To pass specific features, list their names, such as `["feature1", "feature2"]`. This parameter can be used in conjunction with `ContextFeatures`.

RankScore is a model fusion expression for multiple target prediction values. For example, in dbmtl_v1_probs_ctr, dbmtl_v1 refers to the model configuration name, and probs_ctr refers to one of the objectives of the multi-objective model.

RankScore supports the following arithmetic operators: addition (+), subtraction (-), multiplication (*), division (/), exponentiation (^), and modulo (%). For example, ${dbmtl_v1_probs_ctr}%2 performs a modulo operation, and ${dbmtl_v1_probs_ctr}^2 performs an exponentiation operation.

Fusing scores from multiple models follows a similar pattern. For example, if you have two models named dbmtl_v1 and dbmtl_v2, set RankAlgoList to ["dbmtl_v1","dbmtl_v2"]. You can then set RankScore to an expression like ${dbmtl_v1_probs_ctr}+${dbmtl_v2_probs_y_gift}. Each part of the expression combines a model name with one of its target objectives. In this example, probs_ctr and y_gift are used for demonstration purposes.

BatchCount specifies the number of items PAI-Rec sends in a single batch request. For example, if a candidate set for ranking contains 1,000 items and you set BatchCount to 100, PAI-Rec sends 10 separate batch requests, each containing 100 items. Adjust this parameter based on the resources of your model inference service.