BLEU scores too high in inference progess

[Shijie97]

Hi seanie!
Really appreciate for your code share, it does help me deepen my understanding of this paper.
But I have some trouble reproducing your code in the inference stage.
After inference stage, I moved the two files "generated.txt" and "golden.txt" to the "qgevalcap\result" folder, while there has been a "src.txt" in "qgevalcap\result" folder.
I have a few silly questions that may need to be answered.

1. what on earth is the file "src.txt" for? Does it participate in the calculation of BLEU? Dost it exsit at the beginning?
2. how is "golden.txt" generated? If this file is invariant, why not saving it as invariant file just like "scr.txt"?
3. My result for BLEU is too high, it being shown below:
   Bleu_1: 0.65958
   Bleu_2: 0.43423
   Bleu_3: 0.30656
   Bleu_4: 0.22511
   ROUGE_L: 0.49419
   CIDEr: 0.56620

I wonder the reason, if you know what the problem is, may you please point it out? Thank you!

[seanie12]

1. It is dummy file for indexing generated sentences
2. golden.txt is the ground truth questions.
3. Did you try my pretrained model?

[leesurelye]

same question, i trained the model at the beginging. the scores is :
Bleu_1: 0.51006
Bleu_2: 0.36335
Bleu_3: 0.27556
Bleu_4: 0.21658
METEOR: 0.22948
ROUGE_L: 0.52004
is there something wrong? the scores are too high