Added support for Bert models (#137)

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

Author: jkrukowski

Sources/Tokenizers/Decoder.swift

@@ -23,7 +23,7 @@ extension Decoder {
 
 enum DecoderType: String {
     case Sequence
-//    case WordPiece
+    case WordPiece
     case ByteLevel
     case Replace
     case ByteFallback
@@ -47,11 +47,47 @@ struct DecoderFactory {
         case .Fuse        : return FuseDecoder(config: config)
         case .Strip       : return StripDecoder(config: config)
         case .Metaspace   : return MetaspaceDecoder(config: config)
+        case .WordPiece   : return WordPieceDecoder(config: config)
         default           : fatalError("Unsupported Decoder type: \(typeName)")
         }
     }
 }
 
+class WordPieceDecoder: Decoder {
+    let prefix: String
+    let cleanup: Bool
+
+    required public init(config: Config) {
+        guard let prefix = config.prefix?.stringValue else { fatalError("Missing `prefix` configuration for WordPieceDecoder.") }
+        self.prefix = prefix
+        self.cleanup = config.cleanup?.boolValue ?? false
+    }
+
+    func decode(tokens: [String]) -> [String] {
+        var newTokens = [String]()
+        newTokens.reserveCapacity(tokens.count)
+        for (index, token) in tokens.enumerated() {
+            var decodedToken = token
+            if index != 0 {
+                if decodedToken.hasPrefix(prefix) {
+                    decodedToken = String(decodedToken.dropFirst(prefix.count))
+                } else {
+                    decodedToken = " \(decodedToken)"
+                }
+            }
+            if cleanup {
+                decodedToken = cleanUpTokenization(decodedToken)
+            }
+            newTokens.append(decodedToken)
+        }
+        return newTokens
+    }
+
+    private func cleanUpTokenization(_ token: String) -> String {
+        return token.trimmingCharacters(in: .whitespacesAndNewlines)
+    }
+}
+
 class DecoderSequence: Decoder {
     let decoders: [Decoder]
 

Sources/Tokenizers/Normalizer.swift

@@ -31,6 +31,7 @@ enum NormalizerType: String {
     case NFKD
     case NFKC
     case Bert
+    case BertNormalizer
     case Precompiled
     case StripAccents
     case Strip
@@ -51,7 +52,7 @@ struct NormalizerFactory {
         case .NFC: return NFCNormalizer(config: config)
         case .NFKD: return NFKDNormalizer(config: config)
         case .NFKC: return NFKCNormalizer(config: config)
-        case .Bert: return BertNormalizer(config: config)
+        case .Bert, .BertNormalizer: return BertNormalizer(config: config)
         case .Precompiled: return PrecompiledNormalizer(config: config)
         case .StripAccents: return StripAccentsNormalizer(config: config)
         case .Strip: return StripNormalizer(config: config)

Sources/Tokenizers/PreTokenizer.swift

@@ -46,6 +46,7 @@ enum PreTokenizerType: String {
     case Whitespace
     case WhitespaceSplit
     case Metaspace
+    case BertPreTokenizer
     // Several more to be supported
     case Unknown = ""
 }
@@ -63,11 +64,25 @@ struct PreTokenizerFactory {
         case .Split: return SplitPreTokenizer(config: config)
         case .Whitespace, .WhitespaceSplit: return WhitespacePreTokenizer(config: config)
         case .Metaspace: return MetaspacePreTokenizer(config: config)
+        case .BertPreTokenizer: return BertPreTokenizer(config: config)
         default: fatalError("Unsupported PreTokenizer type: \(typeName)")
         }
     }
 }
 
+class BertPreTokenizer: PreTokenizer {
+    let re: String
+
+    required init(config: Config) {
+        // Ref: https://github.com/huggingface/transformers.js/blob/27920d84831e323275b38f0b5186644b7936e1a2/src/tokenizers.js#L1002
+        re = "[^\\s\(Constants.PUNCTUATION_REGEX)]+|[\(Constants.PUNCTUATION_REGEX)]"
+    }
+
+    func preTokenize(text: String, options: PreTokenizerOptions = [.firstSection]) -> [String] {
+        return text.ranges(of: re).map { String(text[$0]) }
+    }
+}
+
 class PreTokenizerSequence: PreTokenizer {
     let preTokenizers: [PreTokenizer]
 
@@ -184,11 +199,10 @@ class ByteLevelPreTokenizer: PreTokenizer {
 }
 
 class PunctuationPreTokenizer: PreTokenizer {
-    let PUNCTUATION_REGEX = #"\p{P}\u0021-\u002F\u003A-\u0040\u005B-\u0060\u007B-\u007E"#
     let re: String
 
     required init(config: Config) {
-        re = "[^\(PUNCTUATION_REGEX)]+|[\(PUNCTUATION_REGEX)]+"
+        re = "[^\(Constants.PUNCTUATION_REGEX)]+|[\(Constants.PUNCTUATION_REGEX)]+"
     }
 
     func preTokenize(text: String, options: PreTokenizerOptions = [.firstSection]) -> [String] {

Sources/Tokenizers/Utils.swift

@@ -77,3 +77,7 @@ struct Utils {
     }
 }
 
+enum Constants {
+    static let PUNCTUATION_REGEX = #"\p{P}\u0021-\u002F\u003A-\u0040\u005B-\u0060\u007B-\u007E"#
+}
+

Tests/PreTokenizerTests/PreTokenizerTests.swift

@@ -171,4 +171,24 @@ class PreTokenizerTests: XCTestCase {
             ["▁Hey", "▁my", "▁friend", "▁", "▁<s>", "▁how", "▁are", "▁you"]
         )
     }
+
+    func testBertPreTokenizer() {
+        let preTokenizer1 = BertPreTokenizer(config: Config([:]))
+        XCTAssertEqual(
+            preTokenizer1.preTokenize(text: "Hey friend!"),
+            ["Hey", "friend", "!"]
+        )
+        XCTAssertEqual(
+            preTokenizer1.preTokenize(text: "Hey friend!     How are you?!?"),
+            ["Hey", "friend", "!", "How", "are", "you", "?", "!", "?"]
+        )
+        XCTAssertEqual(
+            preTokenizer1.preTokenize(text: "   Hey,    friend ,    what's up?  "),
+            ["Hey", ",", "friend", ",", "what", "\'", "s", "up", "?"]
+        )
+        XCTAssertEqual(
+            preTokenizer1.preTokenize(text: "   Hey,    friend ,  0 99  what's up?  "),
+            ["Hey", ",", "friend", ",", "0", "99", "what", "\'", "s", "up", "?"]
+        )
+    }
 }