Java 类名：com.alibaba.alink.operator.batch.timeseries.LSTNetTrainBatchOp
Python 类名：LSTNetTrainBatchOp

功能介绍

使用 LSTNet 进行时间序列训练和预测。

使用方式

参考文档 https://www.yuque.com/pinshu/alink_guide/xbp5ky

参数说明

| 名称 | 中文名称 | 描述 | 类型 | 是否必须？ | 取值范围 | 默认值 | | —- | —- | —- | —- | —- | —- | —- |

| checkpointFilePath | 保存 checkpoint 的路径 | 用于保存中间结果的路径，将作为 TensorFlow 中 Estimator 的 model_dir 传入，需要为所有 worker 都能访问到的目录 | String | ✓ | | |

| timeCol | 时间戳列(TimeStamp) | 时间戳列(TimeStamp) | String | ✓ | 所选列类型为 [TIMESTAMP] | |

| batchSize | 数据批大小 | 数据批大小 | Integer | | | 128 |

| horizon | horizon大小 | horizon大小 | Integer | | [1, +inf) | 12 |

| intraOpParallelism | Op 间并发度 | Op 间并发度 | Integer | | | 4 |

| learningRate | 学习率 | 学习率 | Double | | | 0.001 |

| numEpochs | epoch数 | epoch数 | Integer | | | 10 |

| numPSs | PS 角色数 | PS 角色的数量。值未设置时，如果 Worker 角色数也未设置，则为作业总并发度的 1/4（需要取整），否则为总并发度减去 Worker 角色数。 | Integer | | | null |

| numWorkers | Worker 角色数 | Worker 角色的数量。值未设置时，如果 PS 角色数也未设置，则为作业总并发度的 3/4（需要取整），否则为总并发度减去 PS 角色数。 | Integer | | | null |

| pythonEnv | Python 环境路径 | Python 环境路径，一般情况下不需要填写。如果是压缩文件，需要解压后得到一个目录，且目录名与压缩文件主文件名一致，可以使用 http://, https://, oss://, hdfs:// 等路径；如果是目录，那么只能使用本地路径，即 file://。 | String | | | “” |

| removeCheckpointBeforeTraining | 是否在训练前移除 checkpoint 相关文件 | 是否在训练前移除 checkpoint 相关文件用于重新训练，只会删除必要的文件 | Boolean | | | null |

| selectedCol | 计算列对应的列名 | 计算列对应的列名, 默认值是null | String | | | null |

| vectorCol | 向量列名 | 向量列对应的列名，默认值是null | String | | 所选列类型为 [DENSE_VECTOR, SPARSE_VECTOR, STRING, VECTOR] | null |

| window | 窗口大小 | 窗口大小 | Integer | | | 5 |

代码示例

Python 代码

  1. from pyalink.alink import *
  3. import pandas as pd
  5. useLocalEnv(1)
  7. import time, datetime
  8. import numpy as np
  9. import pandas as pd
  11. data = pd.DataFrame([
  12.     [0, datetime.datetime.fromisoformat("2021-11-01 00:00:00"), 100.0],
  13.     [0, datetime.datetime.fromisoformat("2021-11-02 00:00:00"), 200.0],
  14.     [0, datetime.datetime.fromisoformat("2021-11-03 00:00:00"), 300.0],
  15.     [0, datetime.datetime.fromisoformat("2021-11-04 00:00:00"), 400.0],
  16.     [0, datetime.datetime.fromisoformat("2021-11-06 00:00:00"), 500.0],
  17.     [0, datetime.datetime.fromisoformat("2021-11-07 00:00:00"), 600.0],
  18.     [0, datetime.datetime.fromisoformat("2021-11-08 00:00:00"), 700.0],
  19.     [0, datetime.datetime.fromisoformat("2021-11-09 00:00:00"), 800.0],
  20.     [0, datetime.datetime.fromisoformat("2021-11-10 00:00:00"), 900.0],
  21.     [0, datetime.datetime.fromisoformat("2021-11-11 00:00:00"), 800.0],
  22.     [0, datetime.datetime.fromisoformat("2021-11-12 00:00:00"), 700.0],
  23.     [0, datetime.datetime.fromisoformat("2021-11-13 00:00:00"), 600.0],
  24.     [0, datetime.datetime.fromisoformat("2021-11-14 00:00:00"), 500.0],
  25.     [0, datetime.datetime.fromisoformat("2021-11-15 00:00:00"), 400.0],
  26.     [0, datetime.datetime.fromisoformat("2021-11-16 00:00:00"), 300.0],
  27.     [0, datetime.datetime.fromisoformat("2021-11-17 00:00:00"), 200.0],
  28.     [0, datetime.datetime.fromisoformat("2021-11-18 00:00:00"), 100.0],
  29.     [0, datetime.datetime.fromisoformat("2021-11-19 00:00:00"), 200.0],
  30.     [0, datetime.datetime.fromisoformat("2021-11-20 00:00:00"), 300.0],
  31.     [0, datetime.datetime.fromisoformat("2021-11-21 00:00:00"), 400.0],
  32.     [0, datetime.datetime.fromisoformat("2021-11-22 00:00:00"), 500.0],
  33.     [0, datetime.datetime.fromisoformat("2021-11-23 00:00:00"), 600.0],
  34.     [0, datetime.datetime.fromisoformat("2021-11-24 00:00:00"), 700.0],
  35.     [0, datetime.datetime.fromisoformat("2021-11-25 00:00:00"), 800.0],
  36.     [0, datetime.datetime.fromisoformat("2021-11-26 00:00:00"), 900.0],
  37.     [0, datetime.datetime.fromisoformat("2021-11-27 00:00:00"), 800.0],
  38.     [0, datetime.datetime.fromisoformat("2021-11-28 00:00:00"), 700.0],
  39.     [0, datetime.datetime.fromisoformat("2021-11-29 00:00:00"), 600.0],
  40.     [0, datetime.datetime.fromisoformat("2021-11-30 00:00:00"), 500.0],
  41.     [0, datetime.datetime.fromisoformat("2021-12-01 00:00:00"), 400.0],
  42.     [0, datetime.datetime.fromisoformat("2021-12-02 00:00:00"), 300.0],
  43.     [0, datetime.datetime.fromisoformat("2021-12-03 00:00:00"), 200.0]
  44. ])
  46. source = dataframeToOperator(data, schemaStr='id int, ts timestamp, series double', op_type='batch')
  48. lstNetTrainBatchOp = LSTNetTrainBatchOp()\
  49.     .setTimeCol("ts")\
  50.     .setSelectedCol("series")\
  51.     .setNumEpochs(10)\
  52.     .setWindow(24)\
  53.     .setHorizon(1)
  55. groupByBatchOp = GroupByBatchOp()\
  56.     .setGroupByPredicate("id")\
  57.     .setSelectClause("mtable_agg(ts, series) as mtable_agg_series")
  59. lstNetPredictBatchOp = LSTNetPredictBatchOp()\
  60.     .setPredictNum(1)\
  61.     .setPredictionCol("pred")\
  62.     .setReservedCols([])\
  63.     .setValueCol("mtable_agg_series")\
  65. lstNetPredictBatchOp\
  66.     .linkFrom(
  67.         lstNetTrainBatchOp.linkFrom(source),
  68.         groupByBatchOp.linkFrom(source.filter("ts >= TO_TIMESTAMP('2021-11-10 00:00:00')"))
  69.     )\
  70.     .print()

Java 代码

  1. import org.apache.flink.types.Row;
  3. import com.alibaba.alink.operator.batch.BatchOperator;
  4. import com.alibaba.alink.operator.batch.source.MemSourceBatchOp;
  5. import com.alibaba.alink.operator.batch.sql.GroupByBatchOp;
  6. import com.alibaba.alink.operator.batch.timeseries.LSTNetPredictBatchOp;
  7. import com.alibaba.alink.operator.batch.timeseries.LSTNetTrainBatchOp;
  8. import org.junit.Test;
  10. import java.sql.Timestamp;
  11. import java.util.Arrays;
  12. import java.util.List;
  14. public class LSTNetTrainBatchOpTest {
  16.     @Test
  17.     public void testLSTNetTrainBatchOp() throws Exception {
  18.         BatchOperator.setParallelism(1);
  20.         List <Row> data = Arrays.asList(
  21.             Row.of(0, Timestamp.valueOf("2021-11-01 00:00:00"), 100.0),
  22.             Row.of(0, Timestamp.valueOf("2021-11-02 00:00:00"), 200.0),
  23.             Row.of(0, Timestamp.valueOf("2021-11-03 00:00:00"), 300.0),
  24.             Row.of(0, Timestamp.valueOf("2021-11-04 00:00:00"), 400.0),
  25.             Row.of(0, Timestamp.valueOf("2021-11-06 00:00:00"), 500.0),
  26.             Row.of(0, Timestamp.valueOf("2021-11-07 00:00:00"), 600.0),
  27.             Row.of(0, Timestamp.valueOf("2021-11-08 00:00:00"), 700.0),
  28.             Row.of(0, Timestamp.valueOf("2021-11-09 00:00:00"), 800.0),
  29.             Row.of(0, Timestamp.valueOf("2021-11-10 00:00:00"), 900.0),
  30.             Row.of(0, Timestamp.valueOf("2021-11-11 00:00:00"), 800.0),
  31.             Row.of(0, Timestamp.valueOf("2021-11-12 00:00:00"), 700.0),
  32.             Row.of(0, Timestamp.valueOf("2021-11-13 00:00:00"), 600.0),
  33.             Row.of(0, Timestamp.valueOf("2021-11-14 00:00:00"), 500.0),
  34.             Row.of(0, Timestamp.valueOf("2021-11-15 00:00:00"), 400.0),
  35.             Row.of(0, Timestamp.valueOf("2021-11-16 00:00:00"), 300.0),
  36.             Row.of(0, Timestamp.valueOf("2021-11-17 00:00:00"), 200.0),
  37.             Row.of(0, Timestamp.valueOf("2021-11-18 00:00:00"), 100.0),
  38.             Row.of(0, Timestamp.valueOf("2021-11-19 00:00:00"), 200.0),
  39.             Row.of(0, Timestamp.valueOf("2021-11-20 00:00:00"), 300.0),
  40.             Row.of(0, Timestamp.valueOf("2021-11-21 00:00:00"), 400.0),
  41.             Row.of(0, Timestamp.valueOf("2021-11-22 00:00:00"), 500.0),
  42.             Row.of(0, Timestamp.valueOf("2021-11-23 00:00:00"), 600.0),
  43.             Row.of(0, Timestamp.valueOf("2021-11-24 00:00:00"), 700.0),
  44.             Row.of(0, Timestamp.valueOf("2021-11-25 00:00:00"), 800.0),
  45.             Row.of(0, Timestamp.valueOf("2021-11-26 00:00:00"), 900.0),
  46.             Row.of(0, Timestamp.valueOf("2021-11-27 00:00:00"), 800.0),
  47.             Row.of(0, Timestamp.valueOf("2021-11-28 00:00:00"), 700.0),
  48.             Row.of(0, Timestamp.valueOf("2021-11-29 00:00:00"), 600.0),
  49.             Row.of(0, Timestamp.valueOf("2021-11-30 00:00:00"), 500.0),
  50.             Row.of(0, Timestamp.valueOf("2021-12-01 00:00:00"), 400.0),
  51.             Row.of(0, Timestamp.valueOf("2021-12-02 00:00:00"), 300.0),
  52.             Row.of(0, Timestamp.valueOf("2021-12-03 00:00:00"), 200.0)
  53.         );
  55.         MemSourceBatchOp memSourceBatchOp = new MemSourceBatchOp(data, "id int, ts timestamp, series double");
  57.         LSTNetTrainBatchOp lstNetTrainBatchOp = new LSTNetTrainBatchOp()
  58.             .setTimeCol("ts")
  59.             .setSelectedCol("series")
  60.             .setNumEpochs(10)
  61.             .setWindow(24)
  62.             .setHorizon(1);
  64.         GroupByBatchOp groupByBatchOp = new GroupByBatchOp()
  65.             .setGroupByPredicate("id")
  66.             .setSelectClause("mtable_agg(ts, series) as mtable_agg_series");
  68.         LSTNetPredictBatchOp lstNetPredictBatchOp = new LSTNetPredictBatchOp()
  69.             .setPredictNum(1)
  70.             .setPredictionCol("pred")
  71.             .setReservedCols()
  72.             .setValueCol("mtable_agg_series");
  74.         lstNetPredictBatchOp
  75.             .linkFrom(
  76.                 lstNetTrainBatchOp.linkFrom(memSourceBatchOp),
  77.                 groupByBatchOp.linkFrom(memSourceBatchOp.filter("ts >= TO_TIMESTAMP('2021-11-10 00:00:00')"))
  78.             )
  79.             .print();
  80.     }
  81. }

运行结果

| pred | | —- |

| {“data”:{“ts”:[“2021-12-04 00:00:00.0”],”series”:[441.76019287109375]},”schema”:”ts TIMESTAMP,series DOUBLE”} |