microsoft/onnxruntime-extensions
Publicmirrored fromhttps://github.com/microsoft/onnxruntime-extensionsAvailable
operators/tokenizer/basic_tokenizer.cc
132lines · modecode
| 1 | // Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | // Licensed under the MIT License. |
| 3 | |
| 4 | #include "string_utils.h" |
| 5 | #include "basic_tokenizer.hpp" |
| 6 | #include "string_tensor.h" |
| 7 | #include <vector> |
| 8 | #include <locale> |
| 9 | #include <codecvt> |
| 10 | #include <algorithm> |
| 11 | |
| 12 | BasicTokenizer::BasicTokenizer(bool do_lower_case, bool tokenize_chinese_chars, bool strip_accents, bool tokenize_punctuation, bool remove_control_chars): |
| 13 | do_lower_case_(do_lower_case), |
| 14 | strip_accents_(strip_accents), |
| 15 | tokenize_chinese_chars_(tokenize_chinese_chars), |
| 16 | tokenize_punctuation_(tokenize_punctuation), |
| 17 | remove_control_chars_(remove_control_chars) |
| 18 | {} |
| 19 | |
| 20 | std::vector<ustring> BasicTokenizer::Tokenize(ustring text) { |
| 21 | std::vector<ustring> result; |
| 22 | ustring token; |
| 23 | auto push_current_token_and_clear = [&result, &token]() { |
| 24 | if (!token.empty()) { |
| 25 | result.push_back(token); |
| 26 | token.clear(); |
| 27 | } |
| 28 | }; |
| 29 | |
| 30 | auto push_single_char_and_clear = [&result, &token](char32_t c) { |
| 31 | token.push_back(c); |
| 32 | result.push_back(token); |
| 33 | token.clear(); |
| 34 | }; |
| 35 | |
| 36 | // strip accent first |
| 37 | if (strip_accents_) { |
| 38 | for (auto& c : text) { |
| 39 | c = StripAccent(c); |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | if (do_lower_case_) { |
| 44 | for (auto& c : text) { |
| 45 | c = ToLower(c); |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | for (auto c : text) { |
| 50 | if (tokenize_chinese_chars_ && IsCJK(c)) { |
| 51 | push_current_token_and_clear(); |
| 52 | push_single_char_and_clear(c); |
| 53 | continue; |
| 54 | } |
| 55 | |
| 56 | if (strip_accents_ && IsAccent(c)) { |
| 57 | continue; |
| 58 | } |
| 59 | |
| 60 | // 0x2019 unicode is not punctuation in some Linux platform, |
| 61 | // to be consistent, take it as punctuation. |
| 62 | if (tokenize_punctuation_ && IsPunct(c)) { |
| 63 | push_current_token_and_clear(); |
| 64 | push_single_char_and_clear(c); |
| 65 | continue; |
| 66 | } |
| 67 | |
| 68 | // split by space |
| 69 | if (IsSpace(c)) { |
| 70 | push_current_token_and_clear(); |
| 71 | continue; |
| 72 | } |
| 73 | |
| 74 | if (remove_control_chars_ && IsControl(c)) { |
| 75 | continue; |
| 76 | } |
| 77 | |
| 78 | token.push_back(c); |
| 79 | } |
| 80 | |
| 81 | push_current_token_and_clear(); |
| 82 | return result; |
| 83 | } |
| 84 | |
| 85 | KernelBasicTokenizer::KernelBasicTokenizer(const OrtApi& api, const OrtKernelInfo* info) : BaseKernel(api, info) { |
| 86 | bool do_lower_case = TryToGetAttributeWithDefault("do_lower_case", true); |
| 87 | bool tokenize_chinese_chars = TryToGetAttributeWithDefault("tokenize_chinese_chars", true); |
| 88 | bool strip_accents = TryToGetAttributeWithDefault("strip_accents", false); |
| 89 | bool tokenize_punctuation = TryToGetAttributeWithDefault("tokenize_punctuation", false); |
| 90 | bool remove_control_chars = TryToGetAttributeWithDefault("remove_control_chars", true); |
| 91 | |
| 92 | tokenizer_ = std::make_shared<BasicTokenizer>(do_lower_case, tokenize_chinese_chars, strip_accents, tokenize_punctuation, remove_control_chars); |
| 93 | } |
| 94 | |
| 95 | void KernelBasicTokenizer::Compute(OrtKernelContext* context) { |
| 96 | // Setup inputs |
| 97 | const OrtValue* input = ort_.KernelContext_GetInput(context, 0); |
| 98 | std::vector<std::string> input_data; |
| 99 | GetTensorMutableDataString(api_, ort_, context, input, input_data); |
| 100 | |
| 101 | OrtTensorDimensions dimensions(ort_, input); |
| 102 | if (dimensions.size() != 1 && dimensions[0] != 1) { |
| 103 | ORT_CXX_API_THROW("[BasicTokenizer]: only support string scalar.", ORT_INVALID_GRAPH); |
| 104 | } |
| 105 | |
| 106 | OrtValue* output = ort_.KernelContext_GetOutput(context, 0, dimensions.data(), dimensions.size()); |
| 107 | std::vector<ustring> result = tokenizer_->Tokenize(ustring(input_data[0])); |
| 108 | |
| 109 | FillTensorDataString(api_, ort_, context, result, output); |
| 110 | } |
| 111 | |
| 112 | void* CustomOpBasicTokenizer::CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const { |
| 113 | return CreateKernelImpl(api, info); |
| 114 | }; |
| 115 | |
| 116 | const char* CustomOpBasicTokenizer::GetName() const { return "BasicTokenizer"; }; |
| 117 | |
| 118 | size_t CustomOpBasicTokenizer::GetInputTypeCount() const { |
| 119 | return 1; |
| 120 | }; |
| 121 | |
| 122 | ONNXTensorElementDataType CustomOpBasicTokenizer::GetInputType(size_t /*index*/) const { |
| 123 | return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; |
| 124 | }; |
| 125 | |
| 126 | size_t CustomOpBasicTokenizer::GetOutputTypeCount() const { |
| 127 | return 1; |
| 128 | }; |
| 129 | |
| 130 | ONNXTensorElementDataType CustomOpBasicTokenizer::GetOutputType(size_t /*index*/) const { |
| 131 | return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; |
| 132 | }; |
| 133 | |