Model prediction-Platform For AI(PAI)-阿里云帮助中心

The Model Prediction component runs offline batch inference on a trained EasyRec model in Machine Learning Designer. It reads records from a MaxCompute table, scores each record using the model, and writes the predictions to a new output table.

Use this component when you need to:

Score a large dataset without a persistent online endpoint
Generate embeddings or predicted scores for downstream ranking or filtering steps

Prerequisites

Before you begin, ensure that you have:

A trained EasyRec model produced by the Model Training component. See Model Training.
An input MaxCompute table containing the feature data to score

Configure the component in Machine Learning Designer

Input ports

Port (left to right)	Data type	Supported upstream components	PAI command parameter	Required
Input model	Model stored in OSS	Model Training	`saved_model_dir`	Yes
Input table	MaxCompute table	SQL Script, Read Table	`input_table`	Yes

Parameters setting tab

Parameter	Required	Default	Description	PAI command parameter
Input Feature Columns	No	—	Feature columns selected from the input table for inference. Mutually exclusive with Excluded Columns.	`selected_cols`
Excluded Columns	No	—	Columns to drop from the input table before inference. Mutually exclusive with Input Feature Columns.	`excluded_cols`
Reserved Columns	No	—	Columns copied as-is from the input table into the output table. Use this to preserve ID columns (such as `user_id` or `item_id`) so you can join predictions back to the original dataset in downstream pipeline steps.	`reserved_cols`
Output Columns	No	`probs double`	Names and data types of the prediction columns written to the output table. Separate multiple columns with commas.	`output_cols`
miniBatch size	No	1024	Number of samples processed in each mini-batch. Increase for higher throughput on large datasets; decrease if workers run out of memory.	`batch_size`
Specify the algorithm version	Yes (under Advanced Options)	—	OSS path of the EasyRec TAR package to run. To set this parameter: (1) Generate a TAR package — see Release & Upgrade in the EasyRec documentation. (2) Upload the TAR package to OSS — see Upload objects. (3) Select the uploaded TAR package.	`script`

Tuning tab

All four tuning parameters are combined into the cluster parameter in the PAI command.

Parameter	Required	Default	Description	When to adjust
Worker Count	No	1	Number of worker nodes.	Increase for larger datasets to parallelize inference across multiple workers.
Worker CPU	No	6	CPU cores per worker (1 = one core).	Increase if CPU utilization is consistently high during a run.
Worker Memory	No	30000	Memory per worker in MB (100 = 100 MB).	Increase if workers are killed due to out-of-memory errors.
Worker GPU	No	0	GPUs per worker.	GPUs are not required for most EasyRec inference jobs; set to 0 unless the model was trained with GPU-specific operators.

Output ports

Port (left to right)	Data type	PAI command parameter
Output table	MaxCompute table	`output_table`

PAI command and parameters

The following command shows a complete example. All worker resource settings are encoded in the cluster parameter as a JSON string.

PAI -project algo_public -name easy_rec_ext
    -Darn="acs:ram::xxx:role/aliyunodpspaidefaultrole"
    -Dbatch_size="1024"
    -Dbuckets="oss://rec_sln_demo/"
    -Dcluster="{\"worker\": {\"count\": 1, \"cpu\": 600, \"gpu\": 0, \"memory\": 30000}}"
    -Dcmd="predict"
    -Dinput_table="odps://pai_hangzhou/tables/pai_temp_flow_inpwi02on49ooub78p_node_dn3y3lvucm862jr71n_outputTable"
    -Dlifecycle="28"
    -DossHost="oss-cn-hangzhou-internal.aliyuncs.com"
    -Doutput_cols="item_emb string"
    -Doutput_table="odps://pai_hangzhou/tables/pai_temp_flow_da1nuzwmbdfyw5kajy_node_5jgko0vlrjiwawp6y8_outputTable"
    -Dreserved_cols="item_id"
    -Dsaved_model_dir="oss://rec_sln_demo/EasyRec/deploy/rec_sln_demo_dssm_recall_v1/20230425/export/final/item"
    -Dscript="oss://rec_sln_demo/easy_rec_ext_0.6.1_res.tar.gz";

Parameter	Required	Description
`saved_model_dir`	Yes	Directory of the exported model files in OSS.
`input_table`	Yes	Full path of the input MaxCompute table.
`arn`	Yes	RAM role ARN that grants access to OSS and MaxCompute. To get the value, go to the PAI console, choose Activation & Authorization > Dependent Services, find the Designer section, then click View Authorization next to OSS.
`buckets`	Yes	OSS bucket(s) containing the model files. Separate multiple buckets with commas, for example: `oss://bucket-a/,oss://bucket-b/`.
`ossHost`	Yes	OSS endpoint for the region where the model is stored. See Regions and endpoints.
`output_table`	No	Output MaxCompute table. Created automatically.
`reserved_cols`	No	Columns copied from the input table to the output table. Separate multiple columns with commas.
`output_cols`	No	Names and data types of the prediction columns, for example: `item_emb string`. Separate multiple columns with commas.
`batch_size`	No	Mini-batch size for inference. Default: `1024`.
`script`	No	OSS path of the EasyRec TAR package. See Release & Upgrade in the EasyRec documentation.

Example

Create a pipeline as shown in the following figure. d97c947a2a84a25c45582ba08920423c..png

Section	Description
① ② ③	Configure the upstream components. See Examples in the Model Training documentation.
④	Connect the Model Training output to the left input port of Model Prediction. Connect Read Table-2 to the right input port. Set Excluded Columns to `clk` and set Reserved Columns to `user_id, pid`.

After the pipeline run completes, right-click the Model Prediction component and choose View Data > Output Table.

What's next

For an end-to-end example that uses Model Prediction as part of a DSSM vector recall pipeline, see 13_rec_sln_demo_dssm_recall_item_embedding_v1 in Vector Recall.