com.johnsnowlabs.ml.ai.BartClassification.scala Maven / Gradle / Ivy
The newest version!
/*
* Copyright 2017-2022 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.ml.ai
import ai.onnxruntime.OnnxTensor
import com.johnsnowlabs.ml.ai.util.PrepareEmbeddings
import com.johnsnowlabs.ml.onnx.{OnnxSession, OnnxWrapper}
import com.johnsnowlabs.ml.openvino.OpenvinoWrapper
import com.johnsnowlabs.ml.tensorflow.sign.{ModelSignatureConstants, ModelSignatureManager}
import com.johnsnowlabs.ml.tensorflow.{TensorResources, TensorflowWrapper}
import com.johnsnowlabs.ml.util.{ONNX, Openvino, TensorFlow}
import com.johnsnowlabs.nlp.annotators.common._
import com.johnsnowlabs.nlp.annotators.tokenizer.bpe.BpeTokenizer
import com.johnsnowlabs.nlp.annotators.tokenizer.wordpiece.{BasicTokenizer, WordpieceEncoder}
import com.johnsnowlabs.nlp.{ActivationFunction, Annotation}
import org.intel.openvino.Tensor
import org.slf4j.{Logger, LoggerFactory}
import org.tensorflow.ndarray.buffer.IntDataBuffer
import scala.collection.JavaConverters._
/** @param tensorflowWrapper
* Bert Model wrapper with TensorFlow Wrapper
* @param sentenceStartTokenId
* Id of sentence start Token
* @param sentenceEndTokenId
* Id of sentence end Token.
* @param configProtoBytes
* Configuration for TensorFlow session
* @param tags
* labels which model was trained with in order
* @param signatures
* TF v2 signatures in Spark NLP
*/
private[johnsnowlabs] class BartClassification(
val tensorflowWrapper: Option[TensorflowWrapper],
val onnxWrapper: Option[OnnxWrapper],
val openvinoWrapper: Option[OpenvinoWrapper],
val sentenceStartTokenId: Int,
val sentenceEndTokenId: Int,
val sentencePadTokenId: Int,
configProtoBytes: Option[Array[Byte]] = None,
tags: Map[String, Int],
signatures: Option[Map[String, String]] = None,
merges: Map[(String, String), Int],
vocabulary: Map[String, Int],
threshold: Float = 0.5f)
extends Serializable
with XXXForClassification {
val _tfBartSignatures: Map[String, String] =
signatures.getOrElse(ModelSignatureManager.apply())
val detectedEngine: String =
if (tensorflowWrapper.isDefined) TensorFlow.name
else if (onnxWrapper.isDefined) ONNX.name
else if (openvinoWrapper.isDefined) Openvino.name
else TensorFlow.name
private val onnxSessionOptions: Map[String, String] = new OnnxSession().getSessionOptions
protected val sigmoidThreshold: Float = threshold
protected val logger: Logger = LoggerFactory.getLogger("BartClassification")
def tokenizeWithAlignment(
sentences: Seq[TokenizedSentence],
maxSeqLength: Int,
caseSensitive: Boolean): Seq[WordpieceTokenizedSentence] = {
val bpeTokenizer = BpeTokenizer.forModel("bart", merges, vocabulary)
sentences.map { tokenIndex =>
// filter empty and only whitespace tokens
val bertTokens =
tokenIndex.indexedTokens.filter(x => x.token.nonEmpty && !x.token.equals(" ")).map {
token =>
val content = if (caseSensitive) token.token else token.token.toLowerCase()
val sentenceBegin = token.begin
val sentenceEnd = token.end
val sentenceIndex = tokenIndex.sentenceIndex
val result =
bpeTokenizer.tokenize(Sentence(content, sentenceBegin, sentenceEnd, sentenceIndex))
if (result.nonEmpty) result.head else IndexedToken("")
}
val wordpieceTokens =
bertTokens.flatMap(token => bpeTokenizer.encode(token)).take(maxSeqLength)
WordpieceTokenizedSentence(wordpieceTokens)
}
}
def tokenizeSeqString(
candidateLabels: Seq[String],
maxSeqLength: Int,
caseSensitive: Boolean): Seq[WordpieceTokenizedSentence] = {
val basicTokenizer = new BasicTokenizer(caseSensitive)
val encoder = new WordpieceEncoder(vocabulary, unkToken = "")
val labelsToSentences = candidateLabels.map { s => Sentence(s, 0, s.length - 1, 0) }
labelsToSentences.map(label => {
val tokens = basicTokenizer.tokenize(label)
val wordpieceTokens = tokens.flatMap(token => encoder.encode(token)).take(maxSeqLength)
WordpieceTokenizedSentence(wordpieceTokens)
})
}
def tokenizeDocument(
docs: Seq[Annotation],
maxSeqLength: Int,
caseSensitive: Boolean): Seq[WordpieceTokenizedSentence] = {
// we need the original form of the token
// let's lowercase if needed right before the encoding
val bpeTokenizer = BpeTokenizer.forModel("Bart", merges, vocabulary)
val sentences = docs.map { s => Sentence(s.result, s.begin, s.end, 0) }
sentences.map { sentence =>
val content = if (caseSensitive) sentence.content else sentence.content.toLowerCase()
val sentenceBegin = sentence.start
val sentenceEnd = sentence.end
val sentenceIndex = sentence.index
// TODO: we should implement dedicated the tokenize and tokenizeSubText methods for full a sentence rather than token by token
val indexedTokens =
bpeTokenizer.tokenize(Sentence(content, sentenceBegin, sentenceEnd, sentenceIndex))
val wordpieceTokens =
indexedTokens.flatMap(token => bpeTokenizer.encode(token)).take(maxSeqLength)
WordpieceTokenizedSentence(wordpieceTokens)
}
}
def tag(batch: Seq[Array[Int]]): Seq[Array[Array[Float]]] = {
val tensors = new TensorResources()
val maxSentenceLength = batch.map(encodedSentence => encodedSentence.length).max
val batchLength = batch.length
val tokenBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
val maskBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
// [nb of encoded sentences , maxSentenceLength]
val shape = Array(batch.length.toLong, maxSentenceLength)
batch.zipWithIndex
.foreach { case (sentence, idx) =>
val offset = idx * maxSentenceLength
tokenBuffers.offset(offset).write(sentence)
maskBuffers
.offset(offset)
.write(sentence.map(x => if (x == sentencePadTokenId) 0 else 1))
}
val runner = tensorflowWrapper.get
.getTFSessionWithSignature(configProtoBytes = configProtoBytes, initAllTables = false)
.runner
val tokenTensors = tensors.createIntBufferTensor(shape, tokenBuffers)
val maskTensors = tensors.createIntBufferTensor(shape, maskBuffers)
runner
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.InputIds.key, "missing_input_id_key"),
tokenTensors)
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.AttentionMask.key, "missing_input_mask_key"),
maskTensors)
.fetch(_tfBartSignatures
.getOrElse(ModelSignatureConstants.LogitsOutput.key, "missing_logits_key"))
val outs = runner.run().asScala
val rawScores = TensorResources.extractFloats(outs.head)
outs.foreach(_.close())
tensors.clearSession(outs)
tensors.clearTensors()
val dim = rawScores.length / (batchLength * maxSentenceLength)
val batchScores: Array[Array[Array[Float]]] = rawScores
.grouped(dim)
.map(scores => calculateSoftmax(scores))
.toArray
.grouped(maxSentenceLength)
.toArray
batchScores
}
def tagSequence(batch: Seq[Array[Int]], activation: String): Array[Array[Float]] = {
val tensors = new TensorResources()
val maxSentenceLength = batch.map(encodedSentence => encodedSentence.length).max
val batchLength = batch.length
val tokenBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
val maskBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
// [nb of encoded sentences , maxSentenceLength]
val shape = Array(batch.length.toLong, maxSentenceLength)
batch.zipWithIndex
.foreach { case (sentence, idx) =>
val offset = idx * maxSentenceLength
tokenBuffers.offset(offset).write(sentence)
maskBuffers
.offset(offset)
.write(sentence.map(x => if (x == sentencePadTokenId) 0 else 1))
}
val session = tensorflowWrapper.get.getTFSessionWithSignature(
configProtoBytes = configProtoBytes,
savedSignatures = signatures,
initAllTables = false)
val runner = session.runner
val tokenTensors = tensors.createIntBufferTensor(shape, tokenBuffers)
val maskTensors = tensors.createIntBufferTensor(shape, maskBuffers)
runner
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.InputIds.key, "missing_input_id_key"),
tokenTensors)
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.AttentionMask.key, "missing_input_mask_key"),
maskTensors)
.fetch(_tfBartSignatures
.getOrElse(ModelSignatureConstants.LogitsOutput.key, "missing_logits_key"))
val outs = runner.run().asScala
val rawScores = TensorResources.extractFloats(outs.head)
outs.foreach(_.close())
tensors.clearSession(outs)
tensors.clearTensors()
val dim = rawScores.length / batchLength
val batchScores: Array[Array[Float]] =
rawScores
.grouped(dim)
.map(scores =>
activation match {
case ActivationFunction.softmax => calculateSoftmax(scores)
case ActivationFunction.sigmoid => calculateSigmoid(scores)
case _ => calculateSoftmax(scores)
})
.toArray
batchScores
}
private def padArrayWithZeros(arr: Array[Int], maxLength: Int): Array[Int] = {
if (arr.length >= maxLength) {
arr
} else {
arr ++ Array.fill(maxLength - arr.length)(sentencePadTokenId)
}
}
def tagZeroShotSequence(
batch: Seq[Array[Int]],
entailmentId: Int,
contradictionId: Int,
activation: String): Array[Array[Float]] = {
val maxSentenceLength = batch.map(encodedSentence => encodedSentence.length).max
val paddedBatch = batch.map(arr => padArrayWithZeros(arr, maxSentenceLength))
val batchLength = paddedBatch.length
val rawScores = detectedEngine match {
case ONNX.name => computeZeroShotLogitsWithONNX(paddedBatch, maxSentenceLength)
case Openvino.name => computeZeroShotLogitsWithOv(paddedBatch, maxSentenceLength)
case TensorFlow.name => computeZeroShotLogitsWithTF(paddedBatch, maxSentenceLength)
}
val dim = rawScores.length / batchLength
rawScores
.grouped(dim)
.toArray
}
def computeZeroShotLogitsWithOv(
batch: Seq[Array[Int]],
maxSentenceLength: Int): Array[Float] = {
val batchLength = batch.length
val shape = Array(batchLength, maxSentenceLength)
val (tokenTensors, maskTensors) =
PrepareEmbeddings.prepareOvLongBatchTensors(batch, maxSentenceLength, batchLength)
val inferRequest = openvinoWrapper.get.getCompiledModel().create_infer_request()
inferRequest.set_tensor("input_ids", tokenTensors)
inferRequest.set_tensor("attention_mask", maskTensors)
inferRequest.infer()
try {
try {
inferRequest
.get_tensor("logits")
.data()
}
} catch {
case e: Exception =>
// Log the exception as a warning
logger.warn("Exception in getRawScoresWithOnnx", e)
// Rethrow the exception to propagate it further
throw e
}
}
def computeZeroShotLogitsWithONNX(
batch: Seq[Array[Int]],
maxSentenceLength: Int): Array[Float] = {
val (runner, env) = onnxWrapper.get.getSession(onnxSessionOptions)
val tokenTensors =
OnnxTensor.createTensor(env, batch.map(x => x.map(x => x.toLong)).toArray)
val maskTensors =
OnnxTensor.createTensor(
env,
batch.map(sentence => sentence.map(x => if (x == sentencePadTokenId) 0L else 1L)).toArray)
val inputs =
Map("input_ids" -> tokenTensors, "attention_mask" -> maskTensors).asJava
try {
val results = runner.run(inputs)
try {
val embeddings = results
.get("logits")
.get()
.asInstanceOf[OnnxTensor]
.getFloatBuffer
.array()
tokenTensors.close()
maskTensors.close()
embeddings
} finally if (results != null) results.close()
}
}
def computeZeroShotLogitsWithTF(
batch: Seq[Array[Int]],
maxSentenceLength: Int): Array[Float] = {
val tensors = new TensorResources()
val batchLength = batch.length
val tokenBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
val maskBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
// [nb of encoded sentences , maxSentenceLength]
val shape = Array(batch.length.toLong, maxSentenceLength)
batch.zipWithIndex
.foreach { case (sentence, idx) =>
val offset = idx * maxSentenceLength
tokenBuffers.offset(offset).write(sentence)
maskBuffers.offset(offset).write(sentence.map(x => if (x == 0) 0 else 1))
}
val session = tensorflowWrapper.get.getTFSessionWithSignature(
configProtoBytes = configProtoBytes,
savedSignatures = signatures,
initAllTables = false)
val runner = session.runner
val tokenTensors = tensors.createIntBufferTensor(shape, tokenBuffers)
val maskTensors = tensors.createIntBufferTensor(shape, maskBuffers)
runner
.feed(
_tfBartSignatures.getOrElse(ModelSignatureConstants.InputIds.key, "missing_input_id_key"),
tokenTensors)
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.AttentionMask.key, "missing_input_mask_key"),
maskTensors)
.fetch(_tfBartSignatures
.getOrElse(ModelSignatureConstants.LogitsOutput.key, "missing_logits_key"))
val outs = runner.run().asScala
val rawScores = TensorResources.extractFloats(outs.head)
outs.foreach(_.close())
tensors.clearSession(outs)
tensors.clearTensors()
rawScores
}
def tagSpan(batch: Seq[Array[Int]]): (Array[Array[Float]], Array[Array[Float]]) = {
val tensors = new TensorResources()
val maxSentenceLength = batch.map(encodedSentence => encodedSentence.length).max
val batchLength = batch.length
val tokenBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
val maskBuffers: IntDataBuffer = tensors.createIntBuffer(batchLength * maxSentenceLength)
// [nb of encoded sentences , maxSentenceLength]
val shape = Array(batch.length.toLong, maxSentenceLength)
batch.zipWithIndex
.foreach { case (sentence, idx) =>
val offset = idx * maxSentenceLength
tokenBuffers.offset(offset).write(sentence)
maskBuffers
.offset(offset)
.write(sentence.map(x => if (x == sentencePadTokenId) 0 else 1))
}
val session = tensorflowWrapper.get.getTFSessionWithSignature(
configProtoBytes = configProtoBytes,
savedSignatures = signatures,
initAllTables = false)
val runner = session.runner
val tokenTensors = tensors.createIntBufferTensor(shape, tokenBuffers)
val maskTensors = tensors.createIntBufferTensor(shape, maskBuffers)
runner
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.InputIds.key, "missing_input_id_key"),
tokenTensors)
.feed(
_tfBartSignatures
.getOrElse(ModelSignatureConstants.AttentionMask.key, "missing_input_mask_key"),
maskTensors)
.fetch(_tfBartSignatures
.getOrElse(ModelSignatureConstants.EndLogitsOutput.key, "missing_end_logits_key"))
.fetch(_tfBartSignatures
.getOrElse(ModelSignatureConstants.StartLogitsOutput.key, "missing_start_logits_key"))
val outs = runner.run().asScala
val endLogits = TensorResources.extractFloats(outs.head)
val startLogits = TensorResources.extractFloats(outs.last)
outs.foreach(_.close())
tensors.clearSession(outs)
tensors.clearTensors()
val endDim = endLogits.length / batchLength
val endScores: Array[Array[Float]] =
endLogits.grouped(endDim).map(scores => calculateSoftmax(scores)).toArray
val startDim = startLogits.length / batchLength
val startScores: Array[Array[Float]] =
startLogits.grouped(startDim).map(scores => calculateSoftmax(scores)).toArray
(startScores, endScores)
}
def findIndexedToken(
tokenizedSentences: Seq[TokenizedSentence],
sentence: (WordpieceTokenizedSentence, Int),
tokenPiece: TokenPiece): Option[IndexedToken] = {
tokenizedSentences(sentence._2).indexedTokens.find(p => p.begin == tokenPiece.begin)
}
}