nlp

data/my_squad_answer.txt

@@ -0,0 +1 @@
+copenhagen telephone exchange

data/my_squad_question.txt

@@ -0,0 +1 @@
+Where did Erlang work?

data/my_squad_source.txt

@@ -0,0 +1 @@
+Erlang worked for the Copenhagen Telephone Exchange and wanted to analyze and optimize its operations

modules/7_nlp/include/model.hpp

@@ -14,4 +14,5 @@ class SQuADModel {
 private:
     Tokenizer tokenizer;
     InferenceEngine::InferRequest req;
+    std::string outputName;
 };

modules/7_nlp/include/tokenizer.hpp

@@ -27,6 +27,6 @@ class Tokenizer {
     std::vector<int> tokensToIndices(const std::vector<std::string>& tokens, int maxNumTokens=128);
 
 private:
-    std::vector<std::string> vocab;
     std::map<std::string, int> vocabMap;
+    std::vector<std::string> vocab;
 };

modules/7_nlp/src/model.cpp

@@ -11,16 +11,23 @@ using namespace InferenceEngine;
 using namespace cv;
 using namespace cv::utils::fs;
 
+Blob::Ptr wrapVecToBlob(std::vector<int> v) {
+    std::vector<size_t> dims = {1, v.size()};
+    return make_shared_blob<int32_t>(TensorDesc(Precision::I32, dims, Layout::NC), (int*)v.data());
+}
+
 SQuADModel::SQuADModel() : tokenizer(join(DATA_FOLDER, "bert-large-uncased-vocab.txt")) {
     Core ie;
 
     // Load deep learning network into memory
     CNNNetwork net = ie.ReadNetwork(join(DATA_FOLDER, "distilbert.xml"),
                                     join(DATA_FOLDER, "distilbert.bin"));
-
+    InputInfo::Ptr inputInfo = net.getInputsInfo()["input.1"];
+   // inputInfo->setLayout(Layout::HW);
+    inputInfo->setPrecision(Precision::I32);
+    outputName = net.getOutputsInfo().begin()->first;
     // Initialize runnable object on CPU device
     ExecutableNetwork execNet = ie.LoadNetwork(net, "CPU");
-
     // Create a single processing thread
     req = execNet.CreateInferRequest();
 }
@@ -39,8 +46,37 @@ std::string SQuADModel::getAnswer(const std::string& question, const std::string
     tokens.push_back("[SEP]");
 
     std::vector<int> indices = tokenizer.tokensToIndices(tokens);
+    Blob::Ptr input = wrapVecToBlob(indices);
+    req.SetBlob("input.1", input);
+    req.Infer(); 
+    float* output1 = req.GetBlob(outputName)->buffer();
+    float* output2 = req.GetBlob("Squeeze_438")->buffer();
+    float max1 = output1[0], max2 = output2[0];
+    int indMax1 = 0, indMax2 = 0;
+    for (int i = 0; i < 128; i++) {
+        if (output1[i] > max1) {
+            max1 = output1[i];
+            indMax1 = i;
+        }
 
-    // TODO: forward indices through the network and return an answer
-
-    return "";
+        if (output2[i] > max2) {
+            max2 = output2[i];
+            indMax2 = i;
+        }
+    }
+
+    std::string result;
+    for (int i = indMax1; i < indMax2 + 1; i++) {
+        std::string word = tokens[i];
+        if (word[0] == '#') {
+            result.pop_back();
+            result += word.substr(2, word.length() - 2);
+            result += ' ';
+        }
+        else {
+            result += word + ' ';
+        }
+    }
+    result.pop_back();
+    return result;
 }

modules/7_nlp/src/tokenizer.cpp

@@ -6,7 +6,26 @@
 #include <opencv2/opencv.hpp>
 
 std::vector<std::string> basicTokenize(const std::string& text) {
-    CV_Error(cv::Error::StsNotImplemented, "basicTokenize");
+    std::vector<std::string> basicTokens;
+    std::string currToken = "";
+    for (auto ch : text) {
+        if (isspace(ch)) {
+            if (!currToken.empty())
+                basicTokens.push_back(currToken);
+            currToken = "";
+        } else if (ispunct(ch)) {
+            if (!currToken.empty())
+                basicTokens.push_back(currToken);
+            currToken = ""; currToken += ch;
+            basicTokens.push_back(currToken);
+            currToken = "";
+        } else {
+            currToken += tolower(ch);
+        }
+    }
+    if (!currToken.empty())
+        basicTokens.push_back(currToken);
+    return basicTokens;
 }
 
 std::vector<std::string> wordTokenize(const std::string& word,