如何将分词器从 Huggingface 转换器导出到 CoreML？

Question

我从 Huggingface 转换器加载分词器和 Bert 模型，并将 Bert 模型导出到 CoreML：

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("huawei-noah/TinyBERT_General_4L_312D")

# Load the model
model = AutoModelForTokenClassification.from_pretrained("huawei-noah/TinyBERT_General_4L_312D")

# Example usage
text = "Hugging Face is creating a tool that democratizes AI."
inputs = tokenizer(text, return_tensors="pt")

要求：

pip install transformers torch

如何将分词器从 Huggingface 转换器导出到 CoreML？

Answer 1

这是我使用的 BERT 分词器，效果很好。其中大部分内容来自 Zach Nagengast 和 Julien Chaumond。希望能帮助到你！您所需要的只是分词器词汇的 vocab.txt 文件，可以在这里找到 - https://huggingface.co/google-bert/bert-base-cased/blob/main/vocab.txt

import Foundation

enum TokenizerError: Error {
    case tooLong(String)
}

class BertTokenizer {
    private let basicTokenizer = BasicTokenizer()
    private let wordpieceTokenizer: WordpieceTokenizer
    private let maxLen = 512
    
    private let vocab: [String: Int]
    private let ids_to_tokens: [Int: String]
    
    init() {
        let url = Bundle.main.url(forResource: "vocab", withExtension: "txt")!
        let vocabTxt = try! String(contentsOf: url)
        let tokens = vocabTxt.split(separator: "\n").map { String($0) }
        var vocab: [String: Int] = [:]
        var ids_to_tokens: [Int: String] = [:]
        for (i, token) in tokens.enumerated() {
            vocab[token] = i
            ids_to_tokens[i] = token
        }
        self.vocab = vocab
        self.ids_to_tokens = ids_to_tokens
        self.wordpieceTokenizer = WordpieceTokenizer(vocab: self.vocab)
    }
    
    
    func tokenize(text: String) -> [String] {
        var tokens: [String] = []
        for token in basicTokenizer.tokenize(text: text) {
            for subToken in wordpieceTokenizer.tokenize(word: token) {
                tokens.append(subToken)
            }
        }
        return tokens
    }
    
    private func createAttentionMask(tokenIds: [Int]) -> [Int] {
        return tokenIds.map { $0 != 0 ? 1 : 0 }
    }
    
    private func convertTokensToIds(tokens: [String]) -> [Int] {
        if tokens.count > maxLen {
            let truncatedTokens = Array(tokens.prefix(maxLen))
            return truncatedTokens.map { vocab[$0]! }
        } else {
            return tokens.map { vocab[$0]! }
        }
    }
    
    private func padSequence(_ sequence: [Int], toLength length: Int, paddingValue: Int = 0) -> [Int] {
        if sequence.count >= length {
            return Array(sequence.prefix(length))
        } else {
            return sequence + Array(repeating: paddingValue, count: length - sequence.count)
        }
    }
    
    /// Main entry point
    func tokenizeToIds(text: String, maxLength: Int = 512) -> (tokenIds: [Int], attentionMask: [Int]) {
        let tokens = ["[CLS]"] + tokenize(text: text) + ["[SEP]"]
        var tokenIds = convertTokensToIds(tokens: tokens)
        tokenIds = padSequence(tokenIds, toLength: maxLength)
        let attentionMask = createAttentionMask(tokenIds: tokenIds)
        return (tokenIds, attentionMask)
    }
    
    func tokenToId(token: String) -> Int {
        return vocab[token]!
    }
    
    /// Un-tokenization: get tokens from tokenIds
    func unTokenize(tokens: [Int]) -> [String] {
        return tokens.map { ids_to_tokens[$0]! }
    }
    
    /// Un-tokenization:
    func convertWordpieceToBasicTokenList(_ wordpieceTokenList: [String]) -> String {
        var tokenList: [String] = []
        var individualToken: String = ""
        
        for token in wordpieceTokenList {
            if token.starts(with: "##") {
                individualToken += String(token.suffix(token.count - 2))
            } else {
                if individualToken.count > 0 {
                    tokenList.append(individualToken)
                }
                
                individualToken = token
            }
        }
        
        tokenList.append(individualToken)
        
        return tokenList.joined(separator: " ")
    }
}



class BasicTokenizer {
    let neverSplit = [
        "[UNK]", "[SEP]", "[PAD]", "[CLS]", "[MASK]"
    ]
    
    func tokenize(text: String) -> [String] {
        let splitTokens = text.folding(options: .diacriticInsensitive, locale: nil)
            .components(separatedBy: NSCharacterSet.whitespaces)
        let tokens = splitTokens.flatMap({ (token: String) -> [String] in
            if neverSplit.contains(token) {
                return [token]
            }
            var toks: [String] = []
            var currentTok = ""
            for c in token.lowercased() {
                if c.isLetter || c.isNumber || c == "°" {
                    currentTok += String(c)
                } else if currentTok.count > 0 {
                    toks.append(currentTok)
                    toks.append(String(c))
                    currentTok = ""
                } else {
                    toks.append(String(c))
                }
            }
            if currentTok.count > 0 {
                toks.append(currentTok)
            }
            return toks
        })
        return tokens
    }
}


class WordpieceTokenizer {
    private let unkToken = "[UNK]"
    private let maxInputCharsPerWord = 100
    private let vocab: [String: Int]
    
    init(vocab: [String: Int]) {
        self.vocab = vocab
    }
    
    func substr(_ s: String, _ r: Range<Int>) -> String? {
        let stringCount = s.count
        if stringCount < r.upperBound || stringCount < r.lowerBound {
            return nil
        }
        let startIndex = s.index(s.startIndex, offsetBy: r.lowerBound)
        let endIndex = s.index(startIndex, offsetBy: r.upperBound - r.lowerBound)
        return String(s[startIndex..<endIndex])
    }
    
    func tokenize(word: String) -> [String] {
        if word.count > maxInputCharsPerWord {
            return [unkToken]
        }
        var outputTokens: [String] = []
        var isBad = false
        var start = 0
        var subTokens: [String] = []
        while start < word.count {
            var end = word.count
            var cur_substr: String? = nil
            while start < end {
                var substr = substr(word, start..<end)!
                if start > 0 {
                    substr = "##\(substr)"
                }
                if vocab[substr] != nil {
                    cur_substr = substr
                    break
                }
                end -= 1
            }
            if cur_substr == nil {
                isBad = true
                break
            }
            subTokens.append(cur_substr!)
            start = end
        }
        if isBad {
            outputTokens.append(unkToken)
        } else {
            outputTokens.append(contentsOf: subTokens)
        }
        return outputTokens
    }
}```