lstmeval: Improve output by ensuring 'Truth:' text is encoded the sam…

…e as OCR output

This ensures that transformations like unicode normalisation are done on
the truth output as well as the OCR output, so that you can compare
the two properly.

Before this a perfect OCR could show different lines for Truth and OCR
if the OCR output included characters that were normalised.

src/training/unicharset/lstmtester.cpp

@@ -91,6 +91,11 @@ std::string LSTMTester::RunEvalSync(int iteration, const double *training_errors
   int error_count = 0;
   while (error_count < total_pages_) {
     const ImageData *trainingdata = test_data_.GetPageBySerial(eval_iteration);
+    std::vector<int> truth_labels;
+    if (!EncodeString(trainingdata->transcription(), &truth_labels)) {
+      continue
+    }
+    std::string truth_text = trainer.DecodeLabels(truth_labels);
     trainer.SetIteration(++eval_iteration);
     NetworkIO fwd_outputs, targets;
     Trainability result = trainer.PrepareForBackward(trainingdata, &fwd_outputs, &targets);
@@ -99,7 +104,7 @@ std::string LSTMTester::RunEvalSync(int iteration, const double *training_errors
       word_error += trainer.NewSingleError(tesseract::ET_WORD_RECERR);
       ++error_count;
       if (verbosity > 1 || (verbosity > 0 && result != PERFECT)) {
-        tprintf("Truth:%s\n", trainingdata->transcription().c_str());
+        tprintf("Truth:%s\n", truth_text.c_str());
         std::vector<int> ocr_labels;
         std::vector<int> xcoords;
         trainer.LabelsFromOutputs(fwd_outputs, &ocr_labels, &xcoords);