Alink漫谈(四) : 模型的来龙去脉

0x00 摘要

Alink 是阿里巴巴基于实时计算引擎 Flink 研发的新一代机器学习算法平台,是业界首个同时支持批式算法、流式算法的机器学习平台。本文将从模型角度入手带领大家来再次深入Alink。

因为Alink的公开资料太少,所以以下均为自行揣测,肯定会有疏漏错误,希望大家指出,我会随时更新。

0x01 模型

之前的文章中,我们一直没有仔细说明Alink的模型,本篇我们就深入探究一下。套用下范伟的话:我既想知道模型是怎么来的,我又想知道模型是怎么没的。

1.1 模型包含内容

我们先想想,一个机器学习训练出来的模型,应该包含哪些内容。

  • 流水线:因为一个模型可能包括多个阶段,比如转化,预测等,这样构成了一个流水线。
  • 算法 :这个是具体机器学习平台绑定的。比如在Flink就是某一个java算法类。
  • 参数:这个是肯定要有的,机器学习很大一部分工作不就是做这个的嘛。
  • 数据:这个其实也应该算参数的一种,也是训练出来的。比如说KMeans算法训练出来的各个中心点。

1.2 Alink的模型文件

让我们打开Alink的模型文件来验证下:

-1,"{""schema"":["""",""model_id BIGINT,model_info VARCHAR""],""param"":[""{\""outputCol\"":\""\\\""features\\\""\"",\""selectedCols\"":\""[\\\""sepal_length\\\"",\\\""sepal_width\\\"",\\\""petal_length\\\"",\\\""petal_width\\\""]\""}"",""{\""vectorCol\"":\""\\\""features\\\""\"",\""maxIter\"":\""100\"",\""reservedCols\"":\""[\\\""category\\\""]\"",\""k\"":\""3\"",\""predictionCol\"":\""\\\""prediction_result\\\""\"",\""predictionDetailCol\"":\""\\\""prediction_detail\\\""\""}""],""clazz"":[""com.alibaba.alink.pipeline.dataproc.vector.VectorAssembler"",""com.alibaba.alink.pipeline.clustering.KMeansModel""]}"

1,"0^{""vectorCol"":""\""features\"""",""latitudeCol"":null,""longitudeCol"":null,""distanceType"":""\""EUCLIDEAN\"""",""k"":""3"",""vectorSize"":""4""}"
1,"1048576^{""clusterId"":0,""weight"":39.0,""vec"":{""data"":[6.8538461538461535,3.0769230769230766,5.7153846153846155,2.0538461538461545]}}"
1,"2097152^{""clusterId"":1,""weight"":61.0,""vec"":{""data"":[5.883606557377049,2.740983606557377,4.388524590163936,1.4344262295081969]}}"
1,"3145728^{""clusterId"":2,""weight"":50.0,""vec"":{""data"":[5.006,3.418,1.4640000000000002,0.24400000000000005]}}"

我们看到了两个类名字:

com.alibaba.alink.pipeline.dataproc.vector.VectorAssembler

com.alibaba.alink.pipeline.clustering.KMeansModel

这就是我们提到的算法,Alink在执行过程中,可以根据这两个类名字来生成java类。而两个算法类看起来是可以构建成一个流水线。我们也能看到参数和数据。

但是有几个地方很奇怪:

  • 1048576,2097152这些奇怪的数字是什么意思?
  • 为什么文件的第一个数值是-1?然后第二行第一个数字是 1?怎么没有 中间的 0 ?
  • 具体Alink是如何生成和加载模型的?

下面我们就一一排查。

0x02 流程图

我们首先给出一个流程图便于大家理解。这个图只是逻辑上的大致概念,和真实运行有区别。因为实际场景上是先生成执行计划,再具体操作。

* 下面只是逻辑上的大致概念,和真实运行有区别,因为实际场景上是先生成执行计划,再具体操作.
* 所以只是给大家一个概念。
*
*
* Pipeline.fit 训练
* |
* |
* +-----> KMeansTrainModelData [ centroids, params -- 中心点数据,参数]
* | // KMeansOutputModel.calc()中执行,生成中心点数据和参数
* |
* |
* +-----> Tuple2<Params, Iterable<String>> [ "Params"是模型元数据,Iterable<String>是模型具体数据 ]
* | // KMeansModelDataConverter.serializeModel(),进行序列化操作,包括 把数据转换成json,调用KMeansTrainModelData.toParams设置各种参数
* |
* |
* +-----> Collector<Row> [ Row可以有任意的field,基于position(zero-based)访问field ]
* | // ModelConverterUtils.appendMetaRow,ModelConverterUtils.appendDataRows
* |
* |
* +-----> List<Row> model [ collector.getRows() ]
* | // List<Row> model = completeResult.calc(context);
* |
* |
* +-----> DataSet<Row> [ 序列化算子计算结果 ]
* | // BaseComQueue.exec --- serializeModel(clearObjs(loopEnd))
* |
* |
* +-----> Table output [ AlgoOperator.output,就是算子组件的输出表 ]
* | // KMeansTrainBatchOp.linkFrom --- setOutput
* |
* |
* +-----> KMeansModel [ 模型,Find the closest cluster center for every point ]
* | // createModel(train(input).getOutputTable()) 这里设定模型参数
* | // KMeansModel.setModelData(Table modelData) 这里设定模型数据
* |
* |
* +-----> TransformerBase[] [ PipelineModel.transformers ]
* | // 这就是最终训练出来的流水线模型,KMeansModel是其中一个,KMeansModelMapper是KMeansModel的业务组件
* |
* |
* PipelineModel.save 存储
* |
* |
* +-----> BatchOperator [ 把transformers数组压缩成BatchOperator ]
* | // ModelExporterUtils.packTransformersArray
* |
* |
* +-----> 存储的模型文件 [ csv文件 ]
* | // PipelineModel.save --- CsvSinkBatchOp(path)
* |
* PipelineModel.load 加载
* |
* |
* +-----> 存储的模型文件 [ csv文件 ]
* | // PipelineModel.load --- CsvSourceBatchOp(path)
* |
* |
* +-----> KMeansModel [ 模型,Find the closest cluster center for every point ]
* | // 依据文件生成模型,(TransformerBase) clazz.getConstructor(Params.class)
* | // 设置数据((ModelBase) transformers[i]).setModelData(data.getOutputTable())
* |
* +-----> TransformerBase[] [ 从csv文件读取并恢复的transformers ]
* | // ModelExporterUtils.unpackTransformersArray(batchOp)
* |
* |
* +-----> PipelineModel [ 流水线模型 ]
* | // new PipelineModel(ModelExporterUtils.unpackTransformersArray(batchOp));
* |
* |
* PipelineModel.transform(data) 预测
* |
* |
* |
* +-----> ModelSource [ Load model data from ModelSource when open() ]
* | // ModelMapperAdapter.open --- List<Row> modelRows = this.modelSource.getModelRows(getRuntimeContext());
* |
* +-----> Tuple2<Params, Iterable<String>> [ metaAndData ]
* | // SimpleModelDataConverter.load
* |
* |
* +-----> KMeansTrainModelData [ 反序列化 ]
* | // KMeansModelDataConverter.deserializeModel(Params params, Iterable<String> data)
* |
* |
* +-----> KMeansTrainModelData [ Load KMeansTrainModelData from saved model ]
* | // KMeansModelMapper.loadModel
* | // KMeansTrainModelData.loadModelForTrain(Params params, Iterable<String> data)
* |
* |
* +-----> KMeansPredictModelData [ Model data for KMeans trainData ]
* | // 将训练模型数据转换为预测模型数据,里面包含centroids
* | // KMeansUtil.transformTrainDataToPredictData(trainModelData);
* |
* |
* +-----> Row row [ "5.0,3.2,1.2,0.2,Iris-setosa,5.0 3.2 1.2 0.2" ]
* | // row是预测目标数据,ModelMapperAdapter.map
* |
* |
* +-----> Row row [ "0|0.4472728134421832 0.35775115900088217 0.19497602755693455" ]
* | // 预测结果,KMeansModelMapper.map
* |
* |

0x03 生成模型

我们还是用KMeans算法来做示例,看看模型数据是什么样子,如何转换成Alink需要的样子。

VectorAssembler va = new VectorAssembler()
.setSelectedCols(new String[]{"sepal_length", "sepal_width", "petal_length", "petal_width"})
.setOutputCol("features"); KMeans kMeans = new KMeans().setVectorCol("features").setK(3)
.setPredictionCol("prediction_result")
.setPredictionDetailCol("prediction_detail")
.setReservedCols("category")
.setMaxIter(100); Pipeline pipeline = new Pipeline().add(va).add(kMeans);
pipeline.fit(data);

从之前文章中大家可以知道,KMeans训练最重要的类是KMeansTrainBatchOp。KMeansTrainBatchOp在算法迭代结束时候,使用.closeWith(new KMeansOutputModel(distanceType, vectorColName, latitudeColName, longitudeColName))来输出模型。

3.1 生成模型

所以我们重点就要看 KMeansOutputModel 类。其calc函数的作用就是把中心点和参数转化为模型。

  • 首先是调用serializeModel将中心点序列化成json。这里记做 (1),下面代码注释会对应指出。
  • 其次save函数会进行序列化,生成了一个Tuple2 <Params, Iterable>。Params是参数,Iterable是模型的具体数据,就是中心点的集合。这里记做 (2),下面代码注释会对应指出。
  • 然后save函数把参数和数据分开存储。这里记做 (3),下面注释会对应指出。
  • 最后collector就是模型数据。这里记做 (4),下面注释会对应指出。
/**
* Tranform the centroids to KmeansModel.
*/
public class KMeansOutputModel extends CompleteResultFunction {
private DistanceType distanceType;
private String vectorColName;
private String latitudeColName;
private String longtitudeColName;
@Override
public List <Row> calc(ComContext context) {
KMeansTrainModelData modelData = new KMeansTrainModelData();
... 各种赋值操作
modelData.params = new KMeansTrainModelData.ParamSummary();
modelData.params.k = k;
modelData.params.vectorColName = vectorColName;
... // 我们可以看出来,在此处,计算出来的中心点和各种参数已经被添加到KMeansTrainModelData之中。
modelData = {KMeansTrainModelData@11319}
centroids = {ArrayList@11327} size = 3
0 = {KMeansTrainModelData$ClusterSummary@11330}
clusterId = 0
weight = 38.0
vec = {DenseVector@11333} "6.849999999999999 3.0736842105263156 5.742105263157895 2.071052631578947"
1 = {KMeansTrainModelData$ClusterSummary@11331}
2 = {KMeansTrainModelData$ClusterSummary@11332}
params = {KMeansTrainModelData$ParamSummary@11328}
k = 3
vectorSize = 4
distanceType = {DistanceType@11287} "EUCLIDEAN"
vectorColName = "features"
latitudeColName = null
longtitudeColName = null RowCollector collector = new RowCollector();
// save函数中将进行(1)(2)(3),后续代码中会具体给出(1)(2)(3)的位置
new KMeansModelDataConverter().save(modelData, collector); // KMeansModelDataConverter实现了SimpleModelDataConverter,所以save就调用到了KMeansModelDataConverter.save,其调用serializeModel将中心点转换jason。最后生成了一个Tuple2 <Params, Iterable<String>> // (4) 这时候collector就是模型数据。
return collector.getRows(); // 我们能看出来,模型数据已经和模型文件的内容有几分相似了。里面有之前提到的奇怪数字。
collector = {RowCollector@11321}
rows = {ArrayList@11866} size = 4
0 = {Row@11737} "0,{"vectorCol":"\"features\"","latitudeCol":null,"longitudeCol":null,"distanceType":"\"EUCLIDEAN\"","k":"3","vectorSize":"4"}"
1 = {Row@11801} "1048576,{"clusterId":0,"weight":38.0,"vec":{"data":[6.849999999999999,3.0736842105263156,5.742105263157895,2.071052631578947]}}"
2 = {Row@11868} "2097152,{"clusterId":1,"weight":50.0,"vec":{"data":[5.006,3.4179999999999997,1.4640000000000002,0.24400000000000002]}}"
3 = {Row@11869} "3145728,{"clusterId":2,"weight":62.0,"vec":{"data":[5.901612903225806,2.7483870967741937,4.393548387096773,1.4338709677419355]}}"
}
}

具体转化是在KMeansModelDataConverter和其基类SimpleModelDataConverter中完成。首先是调用serializeModel将中心点序列化成json,形成了一个json列表。

/**
* KMeans Model.
* Save the id, center point and point number of clusters.
*/
public class KMeansModelDataConverter extends SimpleModelDataConverter<KMeansTrainModelData, KMeansPredictModelData> {
public KMeansModelDataConverter() {} @Override
public Tuple2<Params, Iterable<String>> serializeModel(KMeansTrainModelData modelData) {
List <String> data = new ArrayList <>();
for (ClusterSummary centroid : modelData.centroids) {
data.add(JsonConverter.toJson(centroid)); // (1),把中心点转换生成json
}
return Tuple2.of(modelData.params.toParams(), data);
} @Override
public KMeansPredictModelData deserializeModel(Params params, Iterable<String> data) {
KMeansTrainModelData trainModelData = KMeansUtil.loadModelForTrain(params, data);
return KMeansUtil.transformTrainDataToPredictData(trainModelData);
}
}

其次进行序列化操作,生成Tuple2<Params, Iterable>。

/**
* The abstract class for a kind of {@link ModelDataConverter} where the model data can serialize to
* "Tuple2&jt;Params, Iterable&jt;String&gt;&gt;". Here "Params" is the meta data of the model, and "Iterable&jt;String&gt;" is
* concrete data of the model.
*/
public abstract class SimpleModelDataConverter<M1, M2> implements ModelDataConverter<M1, M2> {
@Override
public M2 load(List<Row> rows) {
Tuple2<Params, Iterable<String>> metaAndData = ModelConverterUtils.extractModelMetaAndData(rows);
return deserializeModel(metaAndData.f0, metaAndData.f1);
} @Override
public void save(M1 modelData, Collector<Row> collector) {
// (2),序列化生成Tuple2
Tuple2<Params, Iterable<String>> model = serializeModel(modelData); // 此时模型数据是一个元祖Tuple2<Params, Iterable<String>>
model = {Tuple2@11504} "(Params {vectorCol="features", latitudeCol=null, longitudeCol=null, distanceType="EUCLIDEAN", k=3, vectorSize=4},[{"clusterId":0,"weight":38.0,"vec":{"data":[6.849999999999999,3.0736842105263156,5.742105263157895,2.071052631578947]}}, {"clusterId":1,"weight":50.0,"vec":{"data":[5.006,3.4179999999999997,1.4640000000000002,0.24400000000000002]}}, {"clusterId":2,"weight":62.0,"vec":{"data":[5.901612903225806,2.7483870967741937,4.393548387096773,1.4338709677419355]}}])" // (3) 分开发送参数和数据
ModelConverterUtils.appendMetaRow(model.f0, collector, 2);
ModelConverterUtils.appendDataRows(model.f1, collector, 2);
}
}

然后分开存储参数和数据。

/**
* Collector of Row type data.
*/
public class RowCollector implements Collector<Row> {
private List<Row> rows;
@Override
public void collect(Row row) {
rows.add(row); // 把数据存储起来
}
}
// 调用栈是
collect:37, RowCollector (com.alibaba.alink.common.utils)
collect:12, RowCollector (com.alibaba.alink.common.utils)
appendStringData:270, ModelConverterUtils (com.alibaba.alink.common.model)
appendMetaRow:35, ModelConverterUtils (com.alibaba.alink.common.model)
save:57, SimpleModelDataConverter (com.alibaba.alink.common.model)
calc:76, KMeansOutputModel (com.alibaba.alink.operator.common.clustering.kmeans)
mapPartition:287, BaseComQueue$4 (com.alibaba.alink.common.comqueue)

3.2 转换DataSet

模型数据是要转换成 DataSet,即 a collection of rows。其转换目的是为了让模型数据在Alink中更好的传输和被利用。

把模型数据中的string转换为 row数据的时候,可能会遇到string过长的问题,所以Alink就将String分割转存为多行row。这时候就用ModelConverterUtils的getModelId,getStringIndex函数来分割。

这时候得到的model Id就是计算出来的1048576,就是模型文件中的那个奇怪数字

后续load模型时候也会用同样思路从row转换回模型string。

// A utility class for converting model data to a collection of rows.
class ModelConverterUtils {
/**
* Maximum number of slices a string can split to.
*/
static final long MAX_NUM_SLICES = 1024L * 1024L; private static long getModelId(int stringIndex, int sliceIndex) {
return MAX_NUM_SLICES * stringIndex + sliceIndex;
} private static int getStringIndex(long modelId) {
return (int) ((modelId) / MAX_NUM_SLICES);
}
} row = {Row@11714} "1048576,{"clusterId":0,"weight":62.0,"vec":{"data":[5.901612903225806,2.7483870967741932,4.393548387096773,1.4338709677419355]}}"
fields = {Object[2]@11724}
0 = {Long@11725} 1048576
1 = "{"clusterId":0,"weight":62.0,"vec":{"data":[5.901612903225806,2.7483870967741932,4.393548387096773,1.4338709677419355]}}" // 相关调用栈如下
appendStringData:270, ModelConverterUtils (com.alibaba.alink.common.model)
appendDataRows:52, ModelConverterUtils (com.alibaba.alink.common.model)
save:58, SimpleModelDataConverter (com.alibaba.alink.common.model)
calc:76, KMeansOutputModel (com.alibaba.alink.operator.common.clustering.kmeans)
mapPartition:287, BaseComQueue$4 (com.alibaba.alink.common.comqueue)
run:103, MapPartitionDriver (org.apache.flink.runtime.operators)
...
run:748, Thread (java.lang)

3.3 存储为Table

前面KMeansOutputModel最终返回的是一个DataSet,这里将把这个DataSet转化为Table存储在流水线中。

public final class KMeansTrainBatchOp extends BatchOperator <KMeansTrainBatchOp>

	public KMeansTrainBatchOp linkFrom(BatchOperator <?>... inputs) {
DataSet <Row> finalCentroid = iterateICQ(initCentroid, data,
vectorSize, maxIter, tol, distance, distanceType, vectorColName, null, null); // 这里存储为Table
this.setOutput(finalCentroid, new KMeansModelDataConverter().getModelSchema());
return this;
} this = {KMeansTrainBatchOp@5130} "UnnamedTable$1"
params = {Params@5143} "Params {vectorCol="features", maxIter=100, reservedCols=["category"], k=3, predictionCol="prediction_result", predictionDetailCol="prediction_detail"}"
output = {TableImpl@5188} "UnnamedTable$1"
tableEnvironment = {BatchTableEnvironmentImpl@5190}
operationTree = {DataSetQueryOperation@5191}
operationTreeBuilder = {OperationTreeBuilder@5192}
lookupResolver = {LookupCallResolver@5193}
tableName = "UnnamedTable$1"
sideOutputs = null

我们可以看到,在Alink运行时候,模型数据都统一转化为Table类型。这部分原因可能是因为Alink想要统一处理DataSet和DataStream,即批和流都要用一个思路或者代码来处理。而Flink目前已经用Table来统一整合二者,所以Alink就针对此统一用Table。参见如下:

public abstract class ModelBase<M extends ModelBase<M>> extends TransformerBase<M>
implements Model<M> {
protected Table modelData;
} public abstract class AlgoOperator<T extends AlgoOperator<T>>
implements WithParams<T>, HasMLEnvironmentId<T>, Serializable {
// Params for algorithms.
private Params params; // The table held by operator.
private Table output = null; // The side outputs of operator that be similar to the stream's side outputs.
private Table[] sideOutputs = null;
}

0x04 存储模型

4.1 存储代码

我们修改一下代码,调用save函数把流水线模型存储起来。Alink目前是把模型文件存储成特殊格式的csv文件。

Pipeline pipeline = new Pipeline().add(va).add(kMeans);
pipeline.fit(data).save("./kmeans.csv");

流水线存储代码如下:

public class PipelineModel extends ModelBase<PipelineModel> implements LocalPredictable {
// Pack the pipeline model to a BatchOperator.
public BatchOperator save() {
return ModelExporterUtils.packTransformersArray(transformers);
}
}

我们可以看到,流水线最终调用到 ModelExporterUtils.packTransformersArray,所以我们就重点看看这个函数。这里可以解答模型文件中的问题:为什么第一个数值是-1?然后是 1?怎么没有 中间的 0 ?

模型文件中每行第一个数字对应的是transformer的index。config是特殊的所以index设置为-1,下面代码中有指出。

模型文件中的1 就是说明第二个transformer KMeansModel具有数据,具体数据内容就在index 1对应这行 。

为什么模型文件没有 0 就是因为第一个transformer VectorAssembler没有自己的数据,所以就不包括了。

class ModelExporterUtils {
//Pack an array of transformers to a BatchOperator.
static BatchOperator packTransformersArray(TransformerBase[] transformers) {
int numTransformers = transformers.length;
String[] clazzNames = new String[numTransformers];
String[] params = new String[numTransformers];
String[] schemas = new String[numTransformers];
for (int i = 0; i < numTransformers; i++) {
clazzNames[i] = transformers[i].getClass().getCanonicalName();
params[i] = transformers[i].getParams().toJson();
schemas[i] = "";
if (transformers[i] instanceof PipelineModel) {
schemas[i] = CsvUtil.schema2SchemaStr(PIPELINE_MODEL_SCHEMA);
} else if (transformers[i] instanceof ModelBase) {
long envId = transformers[i].getMLEnvironmentId();
BatchOperator data = BatchOperator.fromTable(((ModelBase) transformers[i]).getModelData());
data.setMLEnvironmentId(envId);
data = data.link(new VectorSerializeBatchOp().setMLEnvironmentId(envId));
schemas[i] = CsvUtil.schema2SchemaStr(data.getSchema());
}
}
Map<String, Object> config = new HashMap<>();
config.put("clazz", clazzNames);
config.put("param", params);
config.put("schema", schemas);
// 这里就对应着模型文件的第一个数值 -1,就是config对应的index就是-1。
Row row = Row.of(-1L, JsonConverter.toJson(config)); // 这个时候我们可以看到,schema, param, clazz 就是对应着模型文件中的输出,我们距离目标更近了一步
config = {HashMap@5432} size = 3
"schema" -> {String[2]@5431}
key = "schema"
value = {String[2]@5431}
0 = ""
1 = "model_id BIGINT,model_info VARCHAR"
"param" -> {String[2]@5430}
key = "param"
value = {String[2]@5430}
0 = "{"outputCol":"\"features\"","selectedCols":"[\"sepal_length\",\"sepal_width\",\"petal_length\",\"petal_width\"]"}"
1 = "{"vectorCol":"\"features\"","maxIter":"100","reservedCols":"[\"category\"]","k":"3","predictionCol":"\"prediction_result\"","predictionDetailCol":"\"prediction_detail\""}"
"clazz" -> {String[2]@5429}
key = "clazz"
value = {String[2]@5429}
0 = "com.alibaba.alink.pipeline.dataproc.vector.VectorAssembler"
1 = "com.alibaba.alink.pipeline.clustering.KMeansModel" BatchOperator packed = new MemSourceBatchOp(Collections.singletonList(row), PIPELINE_MODEL_SCHEMA)
.setMLEnvironmentId(transformers.length > 0 ? transformers[0].getMLEnvironmentId() :
MLEnvironmentFactory.DEFAULT_ML_ENVIRONMENT_ID);
for (int i = 0; i < numTransformers; i++) {
BatchOperator data = null;
final long envId = transformers[i].getMLEnvironmentId();
if (transformers[i] instanceof PipelineModel) {
data = packTransformersArray(((PipelineModel) transformers[i]).transformers);
} else if (transformers[i] instanceof ModelBase) {
data = BatchOperator.fromTable(((ModelBase) transformers[i]).getModelData())
.setMLEnvironmentId(envId);
data = data.link(new VectorSerializeBatchOp().setMLEnvironmentId(envId));
}
if (data != null) {
// 这对应模型文件中的1, 为什么模型文件没有 0就是因为VectorAssembler没有自己的数据,所以就不包括了。
packed = new UnionAllBatchOp().setMLEnvironmentId(envId).linkFrom(packed, packBatchOp(data, i));
}
}
return packed;
}
}

0x05 读取模型

下面代码作用是:读取模型,然后进行转换。

BatchOperator data = new CsvSourceBatchOp().setFilePath(URL).setSchemaStr(SCHEMA_STR);
PipelineModel pipeline = PipelineModel.load("./kmeans.csv");
pipeline.transform(data).print();

读取模型文件,然后转换成PipelineModel。

public class PipelineModel extends ModelBase<PipelineModel> implements LocalPredictable {
//Load the pipeline model from a path.
public static PipelineModel load(String path) {
return load(new CsvSourceBatchOp(path, PIPELINE_MODEL_SCHEMA));
} //Load the pipeline model from a BatchOperator.
public static PipelineModel load(BatchOperator batchOp) {
return new PipelineModel(ModelExporterUtils.unpackTransformersArray(batchOp));
} public PipelineModel(TransformerBase[] transformers) {
super(null);
if (null == transformers) {
this.transformers = new TransformerBase[]{};
} else {
List<TransformerBase> flattened = new ArrayList<>();
flattenTransformers(transformers, flattened);
this.transformers = flattened.toArray(new TransformerBase[0]);
}
}
}
// 相关调用栈如下
unpackTransformersArray:91, ModelExporterUtils (com.alibaba.alink.pipeline)
load:149, PipelineModel (com.alibaba.alink.pipeline)
load:142, PipelineModel (com.alibaba.alink.pipeline)
main:22, KMeansExample2 (com.alibaba.alink)

以下是为导入导出用到的功能类,比如导入导出transformer。我们能够看到大致功能如下:

  • 从index为-1处获取配置信息。
  • 从配置信息中获取了算法类,参数,shema等信息。
  • 根据算法类,生成所有transformer。
  • 每次生成一个新transformer时候,会读取文件中对应行内容,unpack该行内容,生成模型对应的数据,然后赋值给transformer。注意的是,解析出来的数据被包装成一个BatchOperator。
class ModelExporterUtils {
// Unpack transformers array from a BatchOperator.
static TransformerBase[] unpackTransformersArray(BatchOperator batchOp) {
String configStr;
try {
// 从index为-1处获取配置信息。
List<Row> rows = batchOp.as(new String[]{"f1", "f2"}).where("f1=-1").collect();
Preconditions.checkArgument(rows.size() == 1, "Invalid model.");
configStr = (String) rows.get(0).getField(1);
} catch (Exception e) {
throw new RuntimeException("Fail to collect model config.");
}
// 这里从配置信息中获取了算法类,参数,shema等信息
String[] clazzNames = JsonConverter.fromJson(JsonPath.read(configStr, "$.clazz").toString(), String[].class);
String[] params = JsonConverter.fromJson(JsonPath.read(configStr, "$.param").toString(), String[].class);
String[] schemas = JsonConverter.fromJson(JsonPath.read(configStr, "$.schema").toString(), String[].class); // 遍历,生成所有transformer。
int numTransformers = clazzNames.length;
TransformerBase[] transformers = new TransformerBase[numTransformers];
for (int i = 0; i < numTransformers; i++) {
try {
Class clazz = Class.forName(clazzNames[i]);
transformers[i] = (TransformerBase) clazz.getConstructor(Params.class).newInstance(
Params.fromJson(params[i])
.set(HasMLEnvironmentId.ML_ENVIRONMENT_ID, batchOp.getMLEnvironmentId()));
} catch (Exception e) {
throw new RuntimeException("Fail to re construct transformer.", e);
} BatchOperator packed = batchOp.as(new String[]{"f1", "f2"}).where("f1=" + i);
if (transformers[i] instanceof PipelineModel) {
BatchOperator data = unpackBatchOp(packed, CsvUtil.schemaStr2Schema(schemas[i]));
transformers[i] = new PipelineModel(unpackTransformersArray(data))
.setMLEnvironmentId(batchOp.getMLEnvironmentId());
} else if (transformers[i] instanceof ModelBase) {
BatchOperator data = unpackBatchOp(packed, CsvUtil.schemaStr2Schema(schemas[i]));
// 这里会设置模型数据。
((ModelBase) transformers[i]).setModelData(data.getOutputTable());
}
}
return transformers;
} }

最后生成的transformers如下:

transformers = {TransformerBase[2]@9340}
0 = {VectorAssembler@9383}
mapperBuilder = {VectorAssembler$lambda@9385}
params = {Params@9386} "Params {outputCol="features", selectedCols=["sepal_length","sepal_width","petal_length","petal_width"], MLEnvironmentId=0}"
1 = {KMeansModel@9384}
mapperBuilder = {KMeansModel$lambda@9388}
modelData = {TableImpl@9389} "UnnamedTable$1"
params = {Params@9390} "Params {vectorCol="features", maxIter=100, reservedCols=["category"], k=3, MLEnvironmentId=0, predictionCol="prediction_result", predictionDetailCol="prediction_detail"}"

0x06 预测

pipeline.transform(data).print();是预测的代码。

6.1 生成runtime rapper

预测算法需要被包装成RichMapFunction,才能够被Flink引用。

VectorAssembler是起到转换csv文件作用。KMeansModel是用来预测。预测时候会调用到KMeansModel.transform,其又会调用到linkFrom,这里生成了runtime rapper。

public abstract class MapModel<T extends MapModel<T>>
extends ModelBase<T> implements LocalPredictable {
@Override
public BatchOperator transform(BatchOperator input) {
return new ModelMapBatchOp(this.mapperBuilder, this.params)
.linkFrom(BatchOperator.fromTable(this.getModelData())
.setMLEnvironmentId(input.getMLEnvironmentId()), input);
}
} // this.getModelData()是模型数据,对应linkFrom的输入参数inputs[0]
// input 这个是待处理的数据。,对应linkFrom的输入参数inputs[1] // 模型数据就是之前从csv中取出来设置的。
public abstract class ModelBase<M extends ModelBase<M>> extends TransformerBase<M>
implements Model<M> {
public Table getModelData() {
return this.modelData;
}
}

ModelMapBatchOp.linkFrom 代码中,会生成ModelMapperAdapter。此时会把模型信息作为广播变量存起来。这样在后续预测时候就可以先load模型数据。

public class ModelMapBatchOp<T extends ModelMapBatchOp<T>> extends BatchOperator<T> {

   private static final String BROADCAST_MODEL_TABLE_NAME = "broadcastModelTable";

   // (modelScheme, dataSchema, params) -> ModelMapper
private final TriFunction<TableSchema, TableSchema, Params, ModelMapper> mapperBuilder; public ModelMapBatchOp(TriFunction<TableSchema, TableSchema, Params, ModelMapper> mapperBuilder, Params params) {
super(params);
this.mapperBuilder = mapperBuilder;
} @Override
public T linkFrom(BatchOperator<?>... inputs) {
BroadcastVariableModelSource modelSource = new BroadcastVariableModelSource(BROADCAST_MODEL_TABLE_NAME);
ModelMapper mapper = this.mapperBuilder.apply(
inputs[0].getSchema(),
inputs[1].getSchema(),
this.getParams());
DataSet<Row> modelRows = inputs[0].getDataSet().rebalance();
// 这里会广播变量
DataSet<Row> resultRows = inputs[1].getDataSet()
.map(new ModelMapperAdapter(mapper, modelSource))
.withBroadcastSet(modelRows, BROADCAST_MODEL_TABLE_NAME); TableSchema outputSchema = mapper.getOutputSchema();
this.setOutput(resultRows, outputSchema);
return (T) this;
}
}

6.2 加载模型

当预测时候,ModelMapperAdapter会在open函数先加载模型。

public class ModelMapperAdapter extends RichMapFunction<Row, Row> implements Serializable {
@Override
public void open(Configuration parameters) throws Exception {
List<Row> modelRows = this.modelSource.getModelRows(getRuntimeContext());
this.mapper.loadModel(modelRows);
}
} // 加载出来的模型数据举例如下
modelRows = {ArrayList@10100} size = 4
0 = {Row@10103} "2097152,{"clusterId":1,"weight":62.0,"vec":{"data":[5.901612903225806,2.7483870967741932,4.393548387096773,1.4338709677419355]}}"
1 = {Row@10104} "0,{"vectorCol":"\"features\"","latitudeCol":null,"longitudeCol":null,"distanceType":"\"EUCLIDEAN\"","k":"3","vectorSize":"4"}"
2 = {Row@10105} "3145728,{"clusterId":2,"weight":50.0,"vec":{"data":[5.005999999999999,3.418,1.4639999999999997,0.24400000000000002]}}"
3 = {Row@10106} "1048576,{"clusterId":0,"weight":38.0,"vec":{"data":[6.85,3.0736842105263156,5.742105263157894,2.0710526315789477]}}"

this.mapper.loadModel(modelRows) 会调用KMeansModelMapper.loadModel,其最后调用到

  • ModelConverterUtils.extractModelMetaAndData 来进行反序列化,把DataSet转换回Tuple。
  • 最终调用到KMeansUtil.KMeansTrainModelData生成用来预测的模型KMeansTrainModelData
/**
* The abstract class for a kind of {@link ModelDataConverter} where the model data can serialize to "Tuple2&jt;Params, Iterable&jt;String&gt;&gt;". Here "Params" is the meta data of the model, and "Iterable&jt;String&gt;" is concrete data of the model.
*/
public abstract class SimpleModelDataConverter<M1, M2> implements ModelDataConverter<M1, M2> {
@Override
public M2 load(List<Row> rows) {
Tuple2<Params, Iterable<String>> metaAndData = ModelConverterUtils.extractModelMetaAndData(rows);
return deserializeModel(metaAndData.f0, metaAndData.f1);
}
} metaAndData = {Tuple2@10267} "(Params {vectorCol="features", latitudeCol=null, longitudeCol=null, distanceType="EUCLIDEAN", k=3, vectorSize=4},com.alibaba.alink.common.model.ModelConverterUtils$StringDataIterable@7e9c1b42)"
f0 = {Params@10252} "Params {vectorCol="features", latitudeCol=null, longitudeCol=null, distanceType="EUCLIDEAN", k=3, vectorSize=4}"
params = {HashMap@10273} size = 6
"vectorCol" -> ""features""
"latitudeCol" -> null
"longitudeCol" -> null
"distanceType" -> ""EUCLIDEAN""
"k" -> "3"
"vectorSize" -> "4"
f1 = {ModelConverterUtils$StringDataIterable@10262}
iterator = {ModelConverterUtils$StringDataIterator@10272}
modelRows = {ArrayList@10043} size = 4
order = {Integer[4]@10388}
curr = "{"clusterId":0,"weight":38.0,"vec":{"data":[6.85,3.0736842105263156,5.742105263157894,2.0710526315789477]}}"
listPos = 2

可以看到getModelRows就是从广播变量中读取数据。

public class BroadcastVariableModelSource implements ModelSource {
public List<Row> getModelRows(RuntimeContext runtimeContext) {
return runtimeContext.getBroadcastVariable(modelVariableName);
}
}

6.3 预测

最后预测是在ModelMapperAdapter的map函数。这实际上是 flink根据用户代码生成的执行计划进行相应处理后自己执行的。

/**
* Adapt a {@link ModelMapper} to run within flink.
* <p>
* This adapter class hold the target {@link ModelMapper} and it's {@link ModelSource}. Upon open(),
* it will load model rows from {@link ModelSource} into {@link ModelMapper}.
*/
public class ModelMapperAdapter extends RichMapFunction<Row, Row> implements Serializable {
@Override
public Row map(Row row) throws Exception {
return this.mapper.map(row);
}
}

mapper实际调用到KMeansModelMapper,这里就用到了模型数据。

// Find  the closest cluster center for every point.
public class KMeansModelMapper extends ModelMapper {
@Override
public Row map(Row row){
Vector record = KMeansUtil.getKMeansPredictVector(colIdx, row);
......
if(isPredDetail){
double[] probs = KMeansUtil.getProbArrayFromDistanceArray(clusterDistances);
DenseVector vec = new DenseVector(probs.length);
for(int i = 0; i < this.modelData.params.k; i++){
// 这里就用到了模型数据进行预测
vec.set((int)this.modelData.getClusterId(i), probs[i]);
}
res.add(vec.toString());
}
return outputColsHelper.getResultRow(row, Row.of(res.toArray(new Object[0])));
}
} // 模型数据如下
this = {KMeansModelMapper@10822}
modelData = {KMeansPredictModelData@10828}
centroids = {FastDistanceMatrixData@10842}
vectors = {DenseMatrix@10843} "mat[4,3]:\n 5.006,6.85,5.901612903225807\n 3.418,3.0736842105263156,2.7483870967741937\n 1.4639999999999997,5.742105263157894,4.393548387096774\n 0.24400000000000002,2.0710526315789473,1.4338709677419355\n"
label = {DenseMatrix@10844} "mat[1,3]:\n 38.945592000000005,93.63106648199445,63.74191987513008\n"
rows = {Row[3]@10845}
params = {KMeansTrainModelData$ParamSummary@10829}
k = 3
vectorSize = 4
distanceType = {DistanceType@10849} "EUCLIDEAN"
vectorColName = "features"
latitudeColName = null
longtitudeColName = null

0x07 流式预测

我们知道Alink是可以支持批式预测和流式预测。我们看看流式预测是怎么处理的。下面就是KMeans的流式预测。

public class KMeansExampleStream {
AlgoOperator getData(boolean isBatch) {
Row[] array = new Row[] {
Row.of(0, "0 0 0"),
Row.of(1, "0.1,0.1,0.1"),
Row.of(2, "0.2,0.2,0.2"),
Row.of(3, "9 9 9"),
Row.of(4, "9.1 9.1 9.1"),
Row.of(5, "9.2 9.2 9.2")
}; if (isBatch) {
return new MemSourceBatchOp(
Arrays.asList(array), new String[] {"id", "vec"});
} else {
return new MemSourceStreamOp(
Arrays.asList(array), new String[] {"id", "vec"});
}
} public static void main(String[] args) throws Exception {
KMeansExampleStream ks = new KMeansExampleStream();
BatchOperator inOp1 = (BatchOperator)ks.getData(true);
StreamOperator inOp2 = (StreamOperator)ks.getData(false); KMeansTrainBatchOp trainBatch = new KMeansTrainBatchOp().setVectorCol("vec").setK(2);
KMeansPredictBatchOp predictBatch = new KMeansPredictBatchOp().setPredictionCol("pred"); trainBatch.linkFrom(inOp1);
KMeansPredictStreamOp predictStream = new KMeansPredictStreamOp(trainBatch).setPredictionCol("pred");
predictStream.linkFrom(inOp2);
predictStream.print(-1,5);
StreamOperator.execute();
}
}

predictStream.linkFrom是我们这里的要点,其调用到ModelMapStreamOp。ModelMapStreamOp这个类的代码虽然少,但是条理非常清晰,非常适合学习。

  • 首先相关继承关系如下KMeansPredictStreamOp extends ModelMapStreamOp
  • 其次能看出来,流预测所依赖的数据模型依然是一个批处理产生的模型BatchOperator model
  • mapperBuilder是业务模型算子,其构造是通过(modelScheme, dataSchema, params) 得出来的,这恰恰就是机器学习的几个要素。
  • KMeansModelMapper就是具体模型算子 :KMeansModelMapper extends ModelMapper
// Find  the closest cluster center for every point.
public final class KMeansPredictStreamOp extends ModelMapStreamOp <KMeansPredictStreamOp>
implements KMeansPredictParams <KMeansPredictStreamOp> { // @param model trained from kMeansBatchOp
public KMeansPredictStreamOp(BatchOperator model) {
this(model, new Params());
} public KMeansPredictStreamOp(BatchOperator model, Params params) {
super(model, KMeansModelMapper::new, params);
}
}

具体深入代码 ,我们可以看到:

  • 首先 ,把DataSet的数据一次性都取出来,因为都取出来容易造成内存问题,所以 DataSet.collect 注释中有警告:Convenience method to get the elements of a DataSet as a List. As DataSet can contain a lot of data, this method should be used with caution.
  • 其次,通过如下代码this.mapperBuilder.apply(modelSchema, in.getSchema(), this.getParams());构建业务模型KMeansModelMapper。
  • 然后,new ModelMapperAdapter(mapper, modelSource)会建立一个 RichFunction 作为运行适配层。
  • 最后,输入的流数据源 in 会通过in.getDataStream().map((new ModelMapperAdapter(mapper, modelSource));来完成预测。
  • 实际上,这时候只是生成stream graph,具体计算是后续flink会根据graph再进行处理。
public class ModelMapStreamOp<T extends ModelMapStreamOp <T>> extends StreamOperator<T> {

	private final BatchOperator model;
// (modelScheme, dataSchema, params) -> ModelMapper
private final TriFunction<TableSchema, TableSchema, Params, ModelMapper> mapperBuilder; public ModelMapStreamOp(BatchOperator model,
TriFunction<TableSchema, TableSchema, Params, ModelMapper> mapperBuilder,
Params params) {
super(params);
this.model = model;
this.mapperBuilder = mapperBuilder;
} @Override
public T linkFrom(StreamOperator<?>... inputs) {
StreamOperator<?> in = checkAndGetFirst(inputs);
TableSchema modelSchema = this.model.getSchema(); try {
// 把模型数据全都取出来
DataBridge modelDataBridge = DirectReader.collect(model);
DataBridgeModelSource modelSource = new DataBridgeModelSource(modelDataBridge);
ModelMapper mapper = this.mapperBuilder.apply(modelSchema, in.getSchema(), this.getParams());
// 生成runtime适配层和预测算子。把预测结果返回。
// 实际上,这时候只是生成stream graph,具体计算是后续flink会根据graph再进行处理。
DataStream <Row> resultRows = in.getDataStream().map(new ModelMapperAdapter(mapper, modelSource));
TableSchema resultSchema = mapper.getOutputSchema();
this.setOutput(resultRows, resultSchema); return (T) this;
} catch (Exception ex) {
throw new RuntimeException(ex);
}
}
}

0x08 总结

现在我们已经梳理了Alink模型的来龙去脉,让我们再次拿出模型文件内容来验证。

  • 第一行是元数据信息,其中包含schema, 算法类名称,元参数。Alink可以通过这些信息生成流水线的transformer。
  • 后续行是算法类所需要的模型数据。每一行对应一个算法类。Alink会取出这些数据来设置到transformer中。
  • 后续行的模型数据是具体算法相关。
  • 第一行特殊之处在于其index是 -1。后续数据行的index从0开始,如果某一个transformer没有数据,则没有对应行,跳过index。

这样Alink就可以根据模型文件生成流水线模型。

-1,"{""schema"":["""",""model_id BIGINT,model_info VARCHAR""],""param"":[""{\""outputCol\"":\""\\\""features\\\""\"",\""selectedCols\"":\""[\\\""sepal_length\\\"",\\\""sepal_width\\\"",\\\""petal_length\\\"",\\\""petal_width\\\""]\""}"",""{\""vectorCol\"":\""\\\""features\\\""\"",\""maxIter\"":\""100\"",\""reservedCols\"":\""[\\\""category\\\""]\"",\""k\"":\""3\"",\""predictionCol\"":\""\\\""prediction_result\\\""\"",\""predictionDetailCol\"":\""\\\""prediction_detail\\\""\""}""],""clazz"":[""com.alibaba.alink.pipeline.dataproc.vector.VectorAssembler"",""com.alibaba.alink.pipeline.clustering.KMeansModel""]}"

1,"0^{""vectorCol"":""\""features\"""",""latitudeCol"":null,""longitudeCol"":null,""distanceType"":""\""EUCLIDEAN\"""",""k"":""3"",""vectorSize"":""4""}"
1,"1048576^{""clusterId"":0,""weight"":39.0,""vec"":{""data"":[6.8538461538461535,3.0769230769230766,5.7153846153846155,2.0538461538461545]}}"
1,"2097152^{""clusterId"":1,""weight"":61.0,""vec"":{""data"":[5.883606557377049,2.740983606557377,4.388524590163936,1.4344262295081969]}}"
1,"3145728^{""clusterId"":2,""weight"":50.0,""vec"":{""data"":[5.006,3.418,1.4640000000000002,0.24400000000000005]}}"