1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 // Generate a list of skip grams from an input. 17 // 18 // Options: 19 // ngram_size: num of words for each output item. 20 // max_skip_size: max num of words to skip. 21 // The op generates ngrams when it is 0. 22 // include_all_ngrams: include all ngrams with size up to ngram_size. 23 // 24 // Input: 25 // A string tensor to generate n-grams. 26 // Dim = {1} 27 // 28 // Output: 29 // A list of strings, each of which contains ngram_size words. 30 // Dim = {num_ngram} 31 32 #include <ctype.h> 33 #include <string> 34 #include <vector> 35 36 #include "tensorflow/lite/c/builtin_op_data.h" 37 #include "tensorflow/lite/c/c_api_internal.h" 38 #include "tensorflow/lite/kernels/kernel_util.h" 39 #include "tensorflow/lite/kernels/op_macros.h" 40 #include "tensorflow/lite/string_util.h" 41 42 namespace tflite { 43 namespace ops { 44 namespace builtin { 45 46 namespace { 47 48 TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { 49 TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); 50 TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); 51 52 TF_LITE_ENSURE_EQ(context, GetInput(context, node, 0)->type, kTfLiteString); 53 TF_LITE_ENSURE_EQ(context, GetOutput(context, node, 0)->type, kTfLiteString); 54 return kTfLiteOk; 55 } 56 57 bool ShouldIncludeCurrentNgram(const TfLiteSkipGramParams* params, int size) { 58 if (size <= 0) { 59 return false; 60 } 61 if (params->include_all_ngrams) { 62 return size <= params->ngram_size; 63 } else { 64 return size == params->ngram_size; 65 } 66 } 67 68 bool ShouldStepInRecursion(const TfLiteSkipGramParams* params, 69 const std::vector<int>& stack, int stack_idx, 70 int num_words) { 71 // If current stack size and next word enumeration are within valid range. 72 if (stack_idx < params->ngram_size && stack[stack_idx] + 1 < num_words) { 73 // If this stack is empty, step in for first word enumeration. 74 if (stack_idx == 0) { 75 return true; 76 } 77 // If next word enumeration are within the range of max_skip_size. 78 // NOTE: equivalent to 79 // next_word_idx = stack[stack_idx] + 1 80 // next_word_idx - stack[stack_idx-1] <= max_skip_size + 1 81 if (stack[stack_idx] - stack[stack_idx - 1] <= params->max_skip_size) { 82 return true; 83 } 84 } 85 return false; 86 } 87 88 TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { 89 auto* params = reinterpret_cast<TfLiteSkipGramParams*>(node->builtin_data); 90 91 // Split sentence to words. 92 std::vector<StringRef> words; 93 tflite::StringRef strref = tflite::GetString(GetInput(context, node, 0), 0); 94 int prev_idx = 0; 95 for (int i = 1; i < strref.len; i++) { 96 if (isspace(*(strref.str + i))) { 97 if (i > prev_idx && !isspace(*(strref.str + prev_idx))) { 98 words.push_back({strref.str + prev_idx, i - prev_idx}); 99 } 100 prev_idx = i + 1; 101 } 102 } 103 if (strref.len > prev_idx) { 104 words.push_back({strref.str + prev_idx, strref.len - prev_idx}); 105 } 106 107 // Generate n-grams recursively. 108 tflite::DynamicBuffer buf; 109 if (words.size() < params->ngram_size) { 110 buf.WriteToTensorAsVector(GetOutput(context, node, 0)); 111 return kTfLiteOk; 112 } 113 114 // Stack stores the index of word used to generate ngram. 115 // The size of stack is the size of ngram. 116 std::vector<int> stack(params->ngram_size, 0); 117 // Stack index that indicates which depth the recursion is operating at. 118 int stack_idx = 1; 119 int num_words = words.size(); 120 121 while (stack_idx >= 0) { 122 if (ShouldStepInRecursion(params, stack, stack_idx, num_words)) { 123 // When current depth can fill with a new word 124 // and the new word is within the max range to skip, 125 // fill this word to stack, recurse into next depth. 126 stack[stack_idx]++; 127 stack_idx++; 128 if (stack_idx < params->ngram_size) { 129 stack[stack_idx] = stack[stack_idx - 1]; 130 } 131 } else { 132 if (ShouldIncludeCurrentNgram(params, stack_idx)) { 133 // Add n-gram to tensor buffer when the stack has filled with enough 134 // words to generate the ngram. 135 std::vector<StringRef> gram(stack_idx); 136 for (int i = 0; i < stack_idx; i++) { 137 gram[i] = words[stack[i]]; 138 } 139 buf.AddJoinedString(gram, ' '); 140 } 141 // When current depth cannot fill with a valid new word, 142 // and not in last depth to generate ngram, 143 // step back to previous depth to iterate to next possible word. 144 stack_idx--; 145 } 146 } 147 148 buf.WriteToTensorAsVector(GetOutput(context, node, 0)); 149 return kTfLiteOk; 150 } 151 } // namespace 152 153 TfLiteRegistration* Register_SKIP_GRAM() { 154 static TfLiteRegistration r = {nullptr, nullptr, Prepare, Eval}; 155 return &r; 156 } 157 158 } // namespace builtin 159 } // namespace ops 160 } // namespace tflite 161