Bert文本对分类器 (BertTextPairClassifier)

Java 类名：com.alibaba.alink.pipeline.classification.BertTextPairClassifier

Python 类名：BertTextPairClassifier

功能介绍

Bert 文本对分类器。

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
labelCol	标签列名	输入表中的标签列名	String	✓
predictionCol	预测结果列名	预测结果列名	String	✓
textCol	文本列	文本列	String	✓
textPairCol	文本对列	文本对列	String	✓
batchSize	数据批大小	数据批大小	Integer			32
bertModelName	BERT模型名字	BERT模型名字： Base-Chinese,Base-Multilingual-Cased,Base-Uncased,Base-Cased	String			“Base-Chinese”
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			null
customConfigJson	自定义参数	对应 https://github.com/alibaba/EasyTransfer/blob/master/easytransfer/app_zoo/app_config.py 中的config_json	String
inferBatchSize	推理数据批大小	推理数据批大小	Integer			256
intraOpParallelism	Op 间并发度	Op 间并发度	Integer			4
learningRate	学习率	学习率	Double			0.001
maxSeqLength	句子截断长度	句子截断长度	Integer			128
modelFilePath	模型的文件路径	模型的文件路径	String			null
numEpochs	epoch 数	epoch 数	Double			0.01
numFineTunedLayers	微调层数	微调层数	Integer			1
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
overwriteSink	是否覆写已有数据	是否覆写已有数据	Boolean			false
predictionDetailCol	预测详细信息列名	预测详细信息列名	String
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
reservedCols	算法保留列名	算法保留列	String[]			null
modelStreamFilePath	模型流的文件路径	模型流的文件路径	String			null
modelStreamScanInterval	扫描模型路径的时间间隔	描模型路径的时间间隔，单位秒	Integer			10
modelStreamStartTime	模型流的起始时间	模型流的起始时间。默认从当前时刻开始读。使用yyyy-mm-dd hh:mm:ss.fffffffff格式，详见Timestamp.valueOf(String s)	String			null

代码示例

** 以下代码仅用于示意，可能需要修改部分代码或者配置环境后才能正常运行！**

Python 代码

url = "http://alink-algo-packages.oss-cn-hangzhou-zmf.aliyuncs.com/data/MRPC/train.tsv"
schemaStr = "f_quality bigint, f_id_1 string, f_id_2 string, f_string_1 string, f_string_2 string"
data = CsvSourceBatchOp() \
    .setFilePath(url) \
    .setSchemaStr(schemaStr) \
    .setFieldDelimiter("\t") \
    .setIgnoreFirstLine(True) \
    .setQuoteChar(None)
data = ShuffleBatchOp().linkFrom(data)

classifier = BertTextPairClassifier() \
    .setTextCol("f_string_1").setTextPairCol("f_string_2").setLabelCol("f_quality") \
    .setNumEpochs(0.1) \
    .setMaxSeqLength(32) \
    .setNumFineTunedLayers(1) \
    .setBertModelName("Base-Uncased") \
    .setPredictionCol("pred") \
    .setPredictionDetailCol("pred_detail")
model = classifier.fit(data)
predict = model.transform(data.firstN(300))
predict.print()

Java 代码

import com.alibaba.alink.operator.batch.BatchOperator;
import com.alibaba.alink.operator.batch.dataproc.ShuffleBatchOp;
import com.alibaba.alink.operator.batch.source.CsvSourceBatchOp;
import com.alibaba.alink.pipeline.classification.BertClassificationModel;
import com.alibaba.alink.pipeline.classification.BertTextClassifier;
import org.junit.Test;

public class BertTextClassifierTest {
	@Test
	public void test() throws Exception {
		String url = "http://alink-test.oss-cn-beijing.aliyuncs.com/jiqi-temp/tf_ut_files/ChnSentiCorp_htl_small.csv";
		String schemaStr = "label bigint, review string";
		BatchOperator <?> data = new CsvSourceBatchOp()
			.setFilePath(url)
			.setSchemaStr(schemaStr)
			.setIgnoreFirstLine(true);
		data = data.where("review is not null");
		data = new ShuffleBatchOp().linkFrom(data);

		BertTextClassifier classifier = new BertTextClassifier()
			.setTextCol("review")
			.setLabelCol("label")
			.setNumEpochs(0.01)
			.setNumFineTunedLayers(1)
			.setMaxSeqLength(128)
			.setBertModelName("Base-Chinese")
			.setPredictionCol("pred")
			.setPredictionDetailCol("pred_detail");
		BertClassificationModel model = classifier.fit(data);
		BatchOperator <?> predict = model.transform(data.firstN(300));
		predict.print();
	}
}

运行结果

f_quality	f_id_1	f_id_2	f_string_1	f_string_2	pred	pred_detail
0	218017	218035	Application Intelligence will be included as p…	The new application intelligence features will…	1	{“0”:0.20335173606872559,“1”:0.7966482639312744}
1	1642169	1642368	The new 25-member Governing Council ’s first m…	Its first decisions were to scrap all holidays…	1	{“0”:0.20335173606872559,“1”:0.7966482639312744}
1	3399091	3399055	Also in Mosul , rebel gunmen on Friday assassi…	Near a mosque in the northern town of Mosul , …	1	{“0”:0.20335173606872559,“1”:0.7966482639312744}
0	2583299	2583319	“ We ‘re still confident that Gephardt will ge… |Whether or not we get to the two-thirds , we ’…|1 |{”0“:0.20335173606872559,”1“:0.7966482639312744}| |1 |1568540 |1568627 |Monday , the CIA said analysts concluded that … |The CIA on Monday said voice and sound analyst…|1 |{”0“:0.20335173606872559,”1“:0.7966482639312744}| |… |… |… |… |… |… |… | |1 |1805639 |1805436 |Printer maker Lexmark International Inc. spurt… |Other gainers included Lexmark , which rose $ …|1 |{”0“:0.23678696155548096,”1“:0.763213038444519} | |1 |2182211 |2182122 |It ended a diplomatic drought between the two … |The contact between the delegations ended a di…|1 |{”0“:0.23678696155548096,”1“:0.763213038444519} | |1 |774666 |774871 |An unclear number of people were killed and mo… |Four people were killed and 50 injured in the …|1 |{”0“:0.23678696155548096,”1“:0.763213038444519} | |0 |2582380 |2582198 |Shaklee spokeswoman Jenifer Thompson said the … |Shaklee spokeswoman Jenifer Thompson referred …|1 |{”0“:0.23678696155548096,”1“:0.763213038444519} | |1 |427232 |427141 |After three months , Atkins dieters had lost a… |Three months into the study , the Atkins group…|1 |{”0“:0.23678696155548096,”1":0.763213038444519}