com.johnsnowlabs.nlp.embeddings.MPNetEmbeddings.scala Maven / Gradle / Ivy
The newest version!
/*
* Copyright 2017-2023 John Snow Labs
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package com.johnsnowlabs.nlp.embeddings
import com.johnsnowlabs.ml.ai.MPNet
import com.johnsnowlabs.ml.onnx.{OnnxWrapper, ReadOnnxModel, WriteOnnxModel}
import com.johnsnowlabs.ml.tensorflow._
import com.johnsnowlabs.ml.util.LoadExternalModel.{
loadTextAsset,
modelSanityCheck,
notSupportedEngineError
}
import com.johnsnowlabs.ml.util.{ONNX, TensorFlow}
import com.johnsnowlabs.nlp._
import com.johnsnowlabs.nlp.annotators.common._
import com.johnsnowlabs.nlp.annotators.tokenizer.wordpiece.{BasicTokenizer, WordpieceEncoder}
import com.johnsnowlabs.nlp.serialization.MapFeature
import com.johnsnowlabs.storage.HasStorageRef
import org.apache.spark.broadcast.Broadcast
import org.apache.spark.ml.param._
import org.apache.spark.ml.util.Identifiable
import org.apache.spark.sql.{DataFrame, SparkSession}
import org.slf4j.{Logger, LoggerFactory}
/** Sentence embeddings using MPNet.
*
* The MPNet model was proposed in MPNet: Masked and Permuted Pre-training for Language
* Understanding by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. MPNet adopts a novel
* pre-training method, named masked and permuted language modeling, to inherit the advantages of
* masked language modeling and permuted language modeling for natural language understanding.
*
* Note that this annotator is only supported for Spark Versions 3.4 and up.
*
* Pretrained models can be loaded with `pretrained` of the companion object:
* {{{
* val embeddings = MPNetEmbeddings.pretrained()
* .setInputCols("document")
* .setOutputCol("mpnet_embeddings")
* }}}
* The default model is `"all_mpnet_base_v2"`, if no name is provided.
*
* For available pretrained models please see the
* [[https://sparknlp.org/models?q=MPNet Models Hub]].
*
* For extended examples of usage, see
* [[https://github.com/JohnSnowLabs/spark-nlp/blob/master/src/test/scala/com/johnsnowlabs/nlp/embeddings/MPNetEmbeddingsTestSpec.scala MPNetEmbeddingsTestSpec]].
*
* '''Sources''' :
*
* [[https://arxiv.org/abs/2004.09297 MPNet: Masked and Permuted Pre-training for Language Understanding]]
*
* [[https://github.com/microsoft/MPNet MPNet Github Repository]]
*
* ''' Paper abstract '''
*
* ''BERT adopts masked language modeling (MLM) for pre-training and is one of the most
* successful pre-training models. Since BERT neglects dependency among predicted tokens, XLNet
* introduces permuted language modeling (PLM) for pre-training to address this problem. However,
* XLNet does not leverage the full position information of a sentence and thus suffers from
* position discrepancy between pre-training and fine-tuning. In this paper, we propose MPNet, a
* novel pre-training method that inherits the advantages of BERT and XLNet and avoids their
* limitations. MPNet leverages the dependency among predicted tokens through permuted language
* modeling (vs. MLM in BERT), and takes auxiliary position information as input to make the
* model see a full sentence and thus reducing the position discrepancy (vs. PLM in XLNet). We
* pre-train MPNet on a large-scale dataset (over 160GB text corpora) and fine-tune on a variety
* of down-streaming tasks (GLUE, SQuAD, etc). Experimental results show that MPNet outperforms
* MLM and PLM by a large margin, and achieves better results on these tasks compared with
* previous state-of-the-art pre-trained methods (e.g., BERT, XLNet, RoBERTa) under the same
* model setting.''
*
* ==Example==
* {{{
* import spark.implicits._
* import com.johnsnowlabs.nlp.base.DocumentAssembler
* import com.johnsnowlabs.nlp.annotators.Tokenizer
* import com.johnsnowlabs.nlp.embeddings.MPNetEmbeddings
* import com.johnsnowlabs.nlp.EmbeddingsFinisher
* import org.apache.spark.ml.Pipeline
*
* val documentAssembler = new DocumentAssembler()
* .setInputCol("text")
* .setOutputCol("document")
*
* val embeddings = MPNetEmbeddings.pretrained("all_mpnet_base_v2", "en")
* .setInputCols("document")
* .setOutputCol("mpnet_embeddings")
*
* val embeddingsFinisher = new EmbeddingsFinisher()
* .setInputCols("mpnet_embeddings")
* .setOutputCols("finished_embeddings")
* .setOutputAsVector(true)
*
* val pipeline = new Pipeline().setStages(Array(
* documentAssembler,
* embeddings,
* embeddingsFinisher
* ))
*
* val data = Seq("This is an example sentence", "Each sentence is converted").toDF("text")
* val result = pipeline.fit(data).transform(data)
*
* result.selectExpr("explode(finished_embeddings) as result").show(1, 80)
* +--------------------------------------------------------------------------------+
* | result|
* +--------------------------------------------------------------------------------+
* |[[0.022502584, -0.078291744, -0.023030775, -0.0051000593, -0.080340415, 0.039...|
* |[[0.041702367, 0.0010974605, -0.015534201, 0.07092203, -0.0017729357, 0.04661...|
* +--------------------------------------------------------------------------------+
* }}}
*
* @see
* [[https://sparknlp.org/docs/en/annotators Annotators Main Page]] for a list of transformer
* based embeddings
* @param uid
* required uid for storing annotator to disk
* @groupname anno Annotator types
* @groupdesc anno
* Required input and expected output annotator types
* @groupname Ungrouped Members
* @groupname param Parameters
* @groupname setParam Parameter setters
* @groupname getParam Parameter getters
* @groupname Ungrouped Members
* @groupprio param 1
* @groupprio anno 2
* @groupprio Ungrouped 3
* @groupprio setParam 4
* @groupprio getParam 5
* @groupdesc param
* A list of (hyper-)parameter keys this annotator can take. Users can set and get the
* parameter values through setters and getters, respectively.
*/
class MPNetEmbeddings(override val uid: String)
extends AnnotatorModel[MPNetEmbeddings]
with HasBatchedAnnotate[MPNetEmbeddings]
with WriteTensorflowModel
with WriteOnnxModel
with HasEmbeddingsProperties
with HasStorageRef
with HasCaseSensitiveProperties
with HasEngine {
/** Annotator reference id. Used to identify elements in metadata or to refer to this annotator
* type
*/
override val inputAnnotatorTypes: Array[String] =
Array(AnnotatorType.DOCUMENT)
override val outputAnnotatorType: AnnotatorType = AnnotatorType.SENTENCE_EMBEDDINGS
/** ConfigProto from tensorflow, serialized into byte array. Get with
* `config_proto.SerializeToString()`
*
* @group param
*/
val configProtoBytes = new IntArrayParam(
this,
"configProtoBytes",
"ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()")
/** Max sentence length to process (Default: `128`)
*
* @group param
*/
val maxSentenceLength =
new IntParam(this, "maxSentenceLength", "Max sentence length to process")
def sentenceStartTokenId: Int = {
$$(vocabulary)("")
}
/** @group setParam */
def sentenceEndTokenId: Int = {
$$(vocabulary)("")
}
/** Vocabulary used to encode the words to ids with WordPieceEncoder
*
* @group param
*/
val vocabulary: MapFeature[String, Int] = new MapFeature(this, "vocabulary").setProtected()
/** @group setParam */
def setVocabulary(value: Map[String, Int]): this.type = set(vocabulary, value)
/** It contains TF model signatures for the laded saved model
*
* @group param
*/
val signatures =
new MapFeature[String, String](model = this, name = "signatures").setProtected()
private var _model: Option[Broadcast[MPNet]] = None
def this() = this(Identifiable.randomUID("MPNET_EMBEDDINGS"))
/** @group setParam */
def setConfigProtoBytes(bytes: Array[Int]): MPNetEmbeddings.this.type =
set(this.configProtoBytes, bytes)
/** @group setParam */
def setMaxSentenceLength(value: Int): this.type = {
require(
value <= 512,
"MPNet models do not support sequences longer than 512 because of trainable positional embeddings.")
require(value >= 1, "The maxSentenceLength must be at least 1")
set(maxSentenceLength, value)
this
}
/** @group getParam */
def getMaxSentenceLength: Int = $(maxSentenceLength)
/** @group setParam */
def setSignatures(value: Map[String, String]): this.type = {
if (get(signatures).isEmpty)
set(signatures, value)
this
}
/** @group setParam */
def setModelIfNotSet(
spark: SparkSession,
tensorflowWrapper: Option[TensorflowWrapper],
onnxWrapper: Option[OnnxWrapper]): MPNetEmbeddings = {
if (_model.isEmpty) {
_model = Some(
spark.sparkContext.broadcast(
new MPNet(
tensorflowWrapper,
onnxWrapper,
configProtoBytes = getConfigProtoBytes,
sentenceStartTokenId = sentenceStartTokenId,
sentenceEndTokenId = sentenceEndTokenId,
signatures = getSignatures)))
}
this
}
/** Set Embeddings dimensions for the BERT model Only possible to set this when the first time
* is saved dimension is not changeable, it comes from BERT config file
*
* @group setParam
*/
override def setDimension(value: Int): this.type = {
if (get(dimension).isEmpty)
set(this.dimension, value)
this
}
/** Whether to lowercase tokens or not
*
* @group setParam
*/
override def setCaseSensitive(value: Boolean): this.type = {
if (get(caseSensitive).isEmpty)
set(this.caseSensitive, value)
this
}
setDefault(dimension -> 768, batchSize -> 8, maxSentenceLength -> 128, caseSensitive -> false)
def tokenize(sentences: Seq[Annotation]): Seq[WordpieceTokenizedSentence] = {
val basicTokenizer = new BasicTokenizer($(caseSensitive))
val encoder = new WordpieceEncoder($$(vocabulary))
sentences.map { s =>
val sent = Sentence(
content = s.result,
start = s.begin,
end = s.end,
metadata = Some(s.metadata),
index = s.begin)
val tokens = basicTokenizer.tokenize(sent)
val wordpieceTokens = tokens.flatMap(token => encoder.encode(token))
WordpieceTokenizedSentence(wordpieceTokens)
}
}
/** takes a document and annotations and produces new annotations of this annotator's annotation
* type
*
* @param batchedAnnotations
* Annotations that correspond to inputAnnotationCols generated by previous annotators if any
* @return
* any number of annotations processed for every input annotation. Not necessary one to one
* relationship
*/
override def batchAnnotate(batchedAnnotations: Seq[Array[Annotation]]): Seq[Seq[Annotation]] = {
val allAnnotations = batchedAnnotations
.filter(_.nonEmpty)
.zipWithIndex
.flatMap { case (annotations, i) =>
annotations.filter(_.result.nonEmpty).map(x => (x, i))
}
// Tokenize sentences
val tokenizedSentences = tokenize(allAnnotations.map(_._1))
val processedAnnotations = if (allAnnotations.nonEmpty) {
this.getModelIfNotSet.predict(
sentences = allAnnotations.map(_._1),
tokenizedSentences = tokenizedSentences,
batchSize = $(batchSize),
maxSentenceLength = $(maxSentenceLength))
} else {
Seq()
}
// Group resulting annotations by rows. If there are not sentences in a given row, return empty sequence
batchedAnnotations.indices.map(rowIndex => {
val rowAnnotations = processedAnnotations
// zip each annotation with its corresponding row index
.zip(allAnnotations)
// select the sentences belonging to the current row
.filter(_._2._2 == rowIndex)
// leave the annotation only
.map(_._1)
if (rowAnnotations.nonEmpty)
rowAnnotations
else
Seq.empty[Annotation]
})
}
/** @group getParam */
def getModelIfNotSet: MPNet = _model.get.value
override def onWrite(path: String, spark: SparkSession): Unit = {
super.onWrite(path, spark)
val suffix = "_mpnet"
getEngine match {
case TensorFlow.name =>
writeTensorflowModelV2(
path,
spark,
getModelIfNotSet.tensorflowWrapper.get,
suffix,
MPNetEmbeddings.tfFile,
configProtoBytes = getConfigProtoBytes,
savedSignatures = getSignatures)
case ONNX.name =>
writeOnnxModel(
path,
spark,
getModelIfNotSet.onnxWrapper.get,
suffix,
MPNetEmbeddings.onnxFile)
case _ =>
throw new Exception(notSupportedEngineError)
}
}
/** @group getParam */
def getConfigProtoBytes: Option[Array[Byte]] = get(this.configProtoBytes).map(_.map(_.toByte))
/** @group getParam */
def getSignatures: Option[Map[String, String]] = get(this.signatures)
override protected def afterAnnotate(dataset: DataFrame): DataFrame = {
dataset.withColumn(
getOutputCol,
wrapSentenceEmbeddingsMetadata(
dataset.col(getOutputCol),
$(dimension),
Some($(storageRef))))
}
}
trait ReadablePretrainedMPNetModel
extends ParamsAndFeaturesReadable[MPNetEmbeddings]
with HasPretrained[MPNetEmbeddings] {
override val defaultModelName: Some[String] = Some("all_mpnet_base_v2")
/** Java compliant-overrides */
override def pretrained(): MPNetEmbeddings = super.pretrained()
override def pretrained(name: String): MPNetEmbeddings = super.pretrained(name)
override def pretrained(name: String, lang: String): MPNetEmbeddings =
super.pretrained(name, lang)
override def pretrained(name: String, lang: String, remoteLoc: String): MPNetEmbeddings =
super.pretrained(name, lang, remoteLoc)
}
trait ReadMPNetDLModel extends ReadTensorflowModel with ReadOnnxModel {
this: ParamsAndFeaturesReadable[MPNetEmbeddings] =>
override val tfFile: String = "mpnet_tensorflow"
override val onnxFile: String = "mpnet_onnx"
def readModel(instance: MPNetEmbeddings, path: String, spark: SparkSession): Unit = {
instance.getEngine match {
case TensorFlow.name =>
val tfWrapper = readTensorflowModel(path, spark, "_mpnet_tf", initAllTables = false)
instance.setModelIfNotSet(spark, Some(tfWrapper), None)
case ONNX.name =>
val onnxWrapper =
readOnnxModel(path, spark, "_mpnet_onnx", zipped = true, useBundle = false, None)
instance.setModelIfNotSet(spark, None, Some(onnxWrapper))
case _ =>
throw new Exception(notSupportedEngineError)
}
}
addReader(readModel)
def loadSavedModel(modelPath: String, spark: SparkSession): MPNetEmbeddings = {
val (localModelPath, detectedEngine) = modelSanityCheck(modelPath)
val vocabs = loadTextAsset(localModelPath, "vocab.txt").zipWithIndex.toMap
/*Universal parameters for all engines*/
val annotatorModel = new MPNetEmbeddings().setVocabulary(vocabs)
annotatorModel.set(annotatorModel.engine, detectedEngine)
detectedEngine match {
case TensorFlow.name =>
val (wrapper, signatures) = TensorflowWrapper.read(
localModelPath,
zipped = false,
useBundle = true,
tags = Array("serve"),
initAllTables = false)
val _signatures = signatures match {
case Some(s) => s
case None => throw new Exception("Cannot load signature definitions from model!")
}
/** the order of setSignatures is important if we use getSignatures inside
* setModelIfNotSet
*/
annotatorModel
.setSignatures(_signatures)
.setModelIfNotSet(spark, Some(wrapper), None)
case ONNX.name =>
val onnxWrapper =
OnnxWrapper.read(spark, localModelPath, zipped = false, useBundle = true)
annotatorModel
.setModelIfNotSet(spark, None, Some(onnxWrapper))
case _ =>
throw new Exception(notSupportedEngineError)
}
annotatorModel
}
}
/** This is the companion object of [[MPNetEmbeddings]]. Please refer to that class for the
* documentation.
*/
object MPNetEmbeddings extends ReadablePretrainedMPNetModel with ReadMPNetDLModel {
private[MPNetEmbeddings] val logger: Logger =
LoggerFactory.getLogger("MPNetEmbeddings")
}