microsoft/onnxruntime-extensions
Publicmirrored fromhttps://github.com/microsoft/onnxruntime-extensionsAvailable
operators/tokenizer/wordpiece_tokenizer.hpp
46lines · modecode
| 1 | // Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | // Licensed under the MIT License. |
| 3 | |
| 4 | #pragma once |
| 5 | |
| 6 | #include <unordered_map> |
| 7 | #include <vector> |
| 8 | #include "ocos.h" |
| 9 | #include "ustring.h" |
| 10 | #include "string_utils.h" |
| 11 | #include "string_tensor.h" |
| 12 | |
| 13 | struct KernelWordpieceTokenizer : BaseKernel { |
| 14 | KernelWordpieceTokenizer(OrtApi api, const OrtKernelInfo* info); |
| 15 | void Compute(OrtKernelContext* context); |
| 16 | |
| 17 | private: |
| 18 | int64_t max_input_chars_per_word_; |
| 19 | std::u32string suffix_indicator_; |
| 20 | ustring unk_token_; |
| 21 | std::unordered_map<std::u32string, int32_t> vocab_; |
| 22 | }; |
| 23 | |
| 24 | struct CustomOpWordpieceTokenizer : Ort::CustomOpBase<CustomOpWordpieceTokenizer, KernelWordpieceTokenizer> { |
| 25 | void* CreateKernel(OrtApi api, const OrtKernelInfo* info) const; |
| 26 | const char* GetName() const; |
| 27 | size_t GetInputTypeCount() const; |
| 28 | ONNXTensorElementDataType GetInputType(size_t index) const; |
| 29 | size_t GetOutputTypeCount() const; |
| 30 | ONNXTensorElementDataType GetOutputType(size_t index) const; |
| 31 | }; |
| 32 | |
| 33 | void KernelWordpieceTokenizer_Split(const std::u32string& suffix_indicator, |
| 34 | const std::u32string& text, |
| 35 | std::vector<std::u32string>& words); |
| 36 | |
| 37 | void KernelWordpieceTokenizer_Tokenizer(const std::unordered_map<std::u32string, int32_t>& vocab, |
| 38 | const std::u32string& suffix_indicator, |
| 39 | const ustring& unk_token, |
| 40 | const std::vector<ustring>& texts, |
| 41 | std::vector<ustring>& tokens, |
| 42 | std::vector<int32_t>& indices, |
| 43 | std::vector<int64_t>& rows, |
| 44 | const int64_t* existing_rows = nullptr, |
| 45 | int64_t n_existing_rows = 0, |
| 46 | int64_t max_input_chars_per_word = 200); |
| 47 | |