|
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
// See the LICENSE file in the project root for more information.
using System;
using System.Collections.Generic;
using System.Globalization;
using System.Linq;
using System.Runtime.CompilerServices;
using System.Text;
using Microsoft.ML;
using Microsoft.ML.Data;
using Microsoft.ML.Data.IO;
using Microsoft.ML.Internal.Utilities;
using Microsoft.ML.Runtime;
using Microsoft.ML.Tokenizers;
using Microsoft.ML.TorchSharp.Extensions;
using Microsoft.ML.TorchSharp.NasBert;
using Microsoft.ML.TorchSharp.NasBert.Models;
using TorchSharp;
using static Microsoft.ML.TorchSharp.NasBert.NasBertTrainer;
using static TorchSharp.torch;
[assembly: LoadableClass(typeof(NerTransformer), null, typeof(SignatureLoadModel),
NerTransformer.UserName, NerTransformer.LoaderSignature)]
[assembly: LoadableClass(typeof(IRowMapper), typeof(NerTransformer), null, typeof(SignatureLoadRowMapper),
NerTransformer.UserName, NerTransformer.LoaderSignature)]
namespace Microsoft.ML.TorchSharp.NasBert
{
using TargetType = VBuffer<long>;
/// <summary>
/// The <see cref="IEstimator{TTransformer}"/> for training a Deep Neural Network(DNN) to classify text.
/// </summary>
/// <remarks>
/// <format type="text/markdown"><).
///
/// ### Input and Output Columns
/// The input label column data must be a Vector of [string](xref:Microsoft.ML.Data.TextDataViewType) type and the sentence columns must be of type<xref:Microsoft.ML.Data.TextDataViewType>.
///
/// This trainer outputs the following columns:
///
/// | Output Column Name | Column Type | Description|
/// | -- | -- | -- |
/// | `PredictedLabel` | Vector of [key](xref:Microsoft.ML.Data.KeyDataViewType) type | The predicted label's index. If its value is i, the actual label would be the i-th category in the key-valued input label type. |
/// | | |
/// | -- | -- |
/// | Machine learning task | Multiclass classification |
/// | Is normalization required? | No |
/// | Is caching required? | No |
/// | Required NuGet in addition to Microsoft.ML | Microsoft.ML.TorchSharp and libtorch-cpu or libtorch-cuda-11.3 or any of the OS specific variants. |
/// | Exportable to ONNX | No |
///
/// ### Training Algorithm Details
/// Trains a Deep Neural Network(DNN) by leveraging an existing pre-trained NAS-BERT roBERTa model for the purpose of named entity recognition.
/// ]]>
/// </format>
/// </remarks>
///
public class NerTrainer : NasBertTrainer<VBuffer<uint>, TargetType>
{
private const char StartChar = (char)(' ' + 256);
public class NerOptions : NasBertOptions
{
public NerOptions()
{
LearningRate = new List<double>() { 2e-4 };
EncoderOutputDim = 384;
EmbeddingDim = 128;
Arches = new int[] { 15, 16, 14, 0, 0, 0, 15, 16, 14, 0, 0, 0, 17, 14, 15, 0, 0, 0, 17, 14, 15, 0, 0, 0 };
TaskType = BertTaskType.NamedEntityRecognition;
}
}
internal NerTrainer(IHostEnvironment env, NerOptions options) : base(env, options)
{
}
internal NerTrainer(IHostEnvironment env,
string labelColumnName = DefaultColumnNames.Label,
string predictionColumnName = DefaultColumnNames.PredictedLabel,
string sentence1ColumnName = "Sentence1",
int batchSize = 32,
int maxEpochs = 10,
IDataView validationSet = null,
BertArchitecture architecture = BertArchitecture.Roberta) :
this(env, new NerOptions
{
PredictionColumnName = predictionColumnName,
ScoreColumnName = default,
Sentence1ColumnName = sentence1ColumnName,
Sentence2ColumnName = default,
LabelColumnName = labelColumnName,
BatchSize = batchSize,
MaxEpoch = maxEpochs,
ValidationSet = validationSet,
})
{
}
private protected override TrainerBase CreateTrainer(TorchSharpBaseTrainer<VBuffer<uint>, TargetType> parent, IChannel ch, IDataView input)
{
return new Trainer(parent, ch, input);
}
private protected override TorchSharpBaseTransformer<VBuffer<uint>, TargetType> CreateTransformer(IHost host, Options options, torch.nn.Module model, DataViewSchema.DetachedColumn labelColumn)
{
return new NerTransformer(host, options as NasBertOptions, model as NasBertModel, labelColumn);
}
internal static bool TokenStartsWithSpace(string token) => token is null || (token.Length != 0 && token[0] == StartChar);
private protected class Trainer : NasBertTrainerBase
{
private const string ModelUrlString = "models/pretrained_NasBert_14M_encoder.tsm";
internal static readonly int[] ZeroArray = new int[] { 0 /* InitToken */};
public Trainer(TorchSharpBaseTrainer<VBuffer<uint>, TargetType> parent, IChannel ch, IDataView input) : base(parent, ch, input, ModelUrlString)
{
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private protected override TargetType AddToTargets(VBuffer<uint> target)
{
// keys are 1 based but the model is 0 based
var tl = target.DenseValues().Select(item => (long)item).ToList();
tl.Insert(0, 0);
VBuffer<long> t = new VBuffer<long>(target.Length + 1, tl.ToArray());
return t;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private protected override torch.Tensor CreateTargetsTensor(ref List<TargetType> targets, torch.Device device)
{
var maxLength = 0;
targets.ForEach(x =>
{
if (x.Length > maxLength)
maxLength = x.Length;
}
);
long[,] targetArray = new long[targets.Count(), maxLength];
for (int i = 0; i < targets.Count(); i++)
{
for (int j = 0; j < targets[i].Length; j++)
{
targetArray[i, j] = targets[i].GetValues()[j];
}
for (int j = targets[i].Length; j < maxLength; j++)
{
targetArray[i, j] = 0;
}
}
return torch.tensor(targetArray, device: Device);
}
private protected override torch.Tensor PrepareRowTensor(ref VBuffer<uint> target)
{
ReadOnlyMemory<char> sentenceRom = default;
Sentence1Getter(ref sentenceRom);
var sentence = sentenceRom.ToString();
Tensor t;
IReadOnlyList<EncodedToken> encoding = Tokenizer.EncodeToTokens(sentence, out string normalizedText);
if (target.Length != encoding.Count)
{
var targetIndex = 0;
var targetEditor = VBufferEditor.Create(ref target, encoding.Count);
var newValues = targetEditor.Values;
for (var i = 0; i < encoding.Count; i++)
{
if (NerTrainer.TokenStartsWithSpace(encoding[i].Value))
{
newValues[i] = target.GetItemOrDefault(++targetIndex);
}
else
{
newValues[i] = target.GetItemOrDefault(targetIndex);
}
}
target = targetEditor.Commit();
}
t = torch.tensor((ZeroArray).Concat(Tokenizer.RobertaModel().ConvertIdsToOccurrenceRanks(encoding.Select(t => t.Id).ToArray())).ToList(), device: Device);
if (t.NumberOfElements > 512)
t = t.slice(0, 0, 512, 1);
return t;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private protected override int GetNumCorrect(torch.Tensor predictions, torch.Tensor targets)
{
predictions = predictions ?? throw new ArgumentNullException(nameof(predictions));
return (int)predictions.eq(targets).sum().ToInt64();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private protected override torch.Tensor GetPredictions(torch.Tensor logits)
{
logits = logits ?? throw new ArgumentNullException(nameof(logits));
var (_, indexes) = logits.max(-1, false);
return indexes.@int();
}
private protected override int GetRowCountAndSetLabelCount(IDataView input)
{
VBuffer<ReadOnlyMemory<char>> keys = default;
input.Schema[Parent.Option.LabelColumnName].GetKeyValues(ref keys);
var labelCol = input.GetColumn<VBuffer<uint>>(Parent.Option.LabelColumnName);
var rowCount = 0;
foreach (var label in labelCol)
{
rowCount++;
}
Parent.Option.NumberOfClasses = keys.Length + 1;
return rowCount;
}
private protected override torch.Tensor GetTargets(torch.Tensor labels)
{
return labels.view(-1);
}
}
}
public sealed class NerTransformer : NasBertTransformer<VBuffer<uint>, TargetType>
{
internal const string LoadName = "NERTrainer";
internal const string UserName = "NER Trainer";
internal const string ShortName = "NER";
internal const string Summary = "NER with NAS-BERT";
internal const string LoaderSignature = "NER";
private static readonly FuncStaticMethodInfo1<object, Delegate> _decodeInitMethodInfo
= new FuncStaticMethodInfo1<object, Delegate>(DecodeInit<int>);
private static VersionInfo GetVersionInfo()
{
return new VersionInfo(
modelSignature: "NER-BERT",
verWrittenCur: 0x00010001, // Initial
verReadableCur: 0x00010001,
verWeCanReadBack: 0x00010001,
loaderSignature: LoaderSignature,
loaderAssemblyName: typeof(TextClassificationTransformer).Assembly.FullName);
}
internal NerTransformer(IHostEnvironment env, NasBertOptions options, NasBertModel model, DataViewSchema.DetachedColumn labelColumn) : base(env, options, model, labelColumn)
{
}
private protected override IRowMapper GetRowMapper(TorchSharpBaseTransformer<VBuffer<uint>, TargetType> parent, DataViewSchema schema)
{
return new Mapper(parent, schema);
}
private protected override void SaveModel(ModelSaveContext ctx)
{
// *** Binary format ***
// BaseModel
// int: id of label column name
// int: id of the score column name
// int: id of output column name
// int: number of classes
// BinaryStream: TS Model
// int: id of sentence 1 column name
// int: id of sentence 2 column name
// LabelValues
SaveBaseModel(ctx, GetVersionInfo());
var labelColType = LabelColumn.Annotations.Schema[AnnotationUtils.Kinds.KeyValues].Type as VectorDataViewType;
Microsoft.ML.Internal.Utilities.Utils.MarshalActionInvoke(SaveLabelValues<int>, labelColType.ItemType.RawType, ctx);
}
private void SaveLabelValues<T>(ModelSaveContext ctx)
{
ValueGetter<VBuffer<T>> getter = LabelColumn.Annotations.GetGetter<VBuffer<T>>(LabelColumn.Annotations.Schema[AnnotationUtils.Kinds.KeyValues]);
var val = default(VBuffer<T>);
getter(ref val);
BinarySaver saver = new BinarySaver(Host, new BinarySaver.Arguments());
int bytesWritten;
var labelColType = LabelColumn.Annotations.Schema[AnnotationUtils.Kinds.KeyValues].Type as VectorDataViewType;
if (!saver.TryWriteTypeAndValue<VBuffer<T>>(ctx.Writer.BaseStream, labelColType, ref val, out bytesWritten))
throw Host.Except("We do not know how to serialize label names of type '{0}'", labelColType.ItemType);
}
//Factory method for SignatureLoadRowMapper.
private static IRowMapper Create(IHostEnvironment env, ModelLoadContext ctx, DataViewSchema inputSchema)
=> Create(env, ctx).MakeRowMapper(inputSchema);
// Factory method for SignatureLoadModel.
private static NerTransformer Create(IHostEnvironment env, ModelLoadContext ctx)
{
Contracts.CheckValue(env, nameof(env));
env.CheckValue(ctx, nameof(ctx));
ctx.CheckAtModel(GetVersionInfo());
// *** Binary format ***
// BaseModel
// int: id of label column name
// int: id of the score column name
// int: id of output column name
// int: number of classes
// BinaryStream: TS Model
// int: id of sentence 1 column name
// int: id of sentence 2 column name
// LabelValues
var options = new NerTrainer.NerOptions()
{
LabelColumnName = ctx.LoadString(),
ScoreColumnName = ctx.LoadStringOrNull(),
PredictionColumnName = ctx.LoadString(),
NumberOfClasses = ctx.Reader.ReadInt32(),
};
var ch = env.Start("Load Model");
var tokenizer = TokenizerExtensions.GetInstance(ch);
EnglishRobertaTokenizer tokenizerModel = tokenizer.RobertaModel();
var model = new NerModel(options, tokenizerModel.PadIndex, tokenizerModel.SymbolsCount, options.NumberOfClasses);
if (!ctx.TryLoadBinaryStream("TSModel", r => model.load(r)))
throw env.ExceptDecode();
options.Sentence1ColumnName = ctx.LoadString();
options.Sentence2ColumnName = ctx.LoadStringOrNull();
options.TaskType = BertTaskType.NamedEntityRecognition;
BinarySaver saver = new BinarySaver(env, new BinarySaver.Arguments());
DataViewType type;
object value;
env.CheckDecode(saver.TryLoadTypeAndValue(ctx.Reader.BaseStream, out type, out value));
var vecType = type as VectorDataViewType;
env.CheckDecode(vecType != null);
env.CheckDecode(value != null);
var labelGetter = Microsoft.ML.Internal.Utilities.Utils.MarshalInvoke(_decodeInitMethodInfo, vecType.ItemType.RawType, value);
var meta = new DataViewSchema.Annotations.Builder();
meta.Add(AnnotationUtils.Kinds.KeyValues, type, labelGetter);
var labelCol = new DataViewSchema.DetachedColumn(options.LabelColumnName, type, meta.ToAnnotations());
return new NerTransformer(env, options, model, labelCol);
}
private static Delegate DecodeInit<T>(object value)
{
VBuffer<T> buffValue = (VBuffer<T>)value;
ValueGetter<VBuffer<T>> buffGetter = (ref VBuffer<T> dst) => buffValue.CopyTo(ref dst);
return buffGetter;
}
private sealed class Mapper : NasBertMapper
{
public Mapper(TorchSharpBaseTransformer<VBuffer<uint>, TargetType> parent, DataViewSchema inputSchema) : base(parent, inputSchema)
{
}
private protected override Delegate CreateGetter(DataViewRow input, int iinfo, TensorCacher outputCacher)
{
var ch = Host.Start("Make Getter");
return MakePredictedLabelGetter(input, ch, outputCacher);
}
private void CondenseOutput(ref VBuffer<UInt32> dst, string sentence, Tokenizer tokenizer, TensorCacher outputCacher)
{
var pre = tokenizer.PreTokenizer.PreTokenize(sentence);
IReadOnlyList<EncodedToken> encoding = tokenizer.EncodeToTokens(sentence, out string normalizedText);
var argmax = (outputCacher as BertTensorCacher).Result.argmax(-1);
var prediction = argmax.ToArray<long>();
var targetIndex = 0;
// Figure out actual count of output tokens
for (var i = 0; i < encoding.Count; i++)
{
if (NerTrainer.TokenStartsWithSpace(encoding[i].Value))
{
targetIndex++;
}
}
var editor = VBufferEditor.Create(ref dst, targetIndex + 1);
var newValues = editor.Values;
targetIndex = 0;
newValues[targetIndex++] = (uint)prediction[0];
for (var i = 1; i < encoding.Count; i++)
{
if (NerTrainer.TokenStartsWithSpace(encoding[i].Value))
{
newValues[targetIndex++] = (uint)prediction[i];
}
}
dst = editor.Commit();
}
private Delegate MakePredictedLabelGetter(DataViewRow input, IChannel ch, TensorCacher outputCacher)
{
ValueGetter<ReadOnlyMemory<char>> getSentence1 = default;
ValueGetter<ReadOnlyMemory<char>> getSentence2 = default;
Tokenizer tokenizer = TokenizerExtensions.GetInstance(ch);
getSentence1 = input.GetGetter<ReadOnlyMemory<char>>(input.Schema[Parent.SentenceColumn.Name]);
ReadOnlyMemory<char> sentence1 = default;
ReadOnlyMemory<char> sentence2 = default;
ValueGetter<VBuffer<UInt32>> classification = (ref VBuffer<UInt32> dst) =>
{
using var disposeScope = torch.NewDisposeScope();
UpdateCacheIfNeeded(input.Position, outputCacher, ref sentence1, ref sentence2, ref getSentence1, ref getSentence2, tokenizer);
var argmax = (outputCacher as BertTensorCacher).Result.argmax(-1);
var prediction = argmax.ToArray<long>();
CondenseOutput(ref dst, sentence1.ToString(), tokenizer, outputCacher);
};
return classification;
}
private protected override Func<int, bool> GetDependenciesCore(Func<int, bool> activeOutput)
{
return col => activeOutput(0) && InputColIndices.Any(i => i == col);
}
}
}
}
|