Only the author mentions half the bleu-4 score

[CallMeRabbit]

The model was trained on flicker8k, but the results achieved only half the BLEU-4 score mentioned by the authors (about 0.14-0.15). I have not modified any parameters in train.py. May I ask why such a result occurs?

[yiliSept]

I also have the problem

[Kevinskt]

I also have the problem

[togawasakikom]

Yes, I also encountered the same problem. The model did not perform well on the validation machine. I tried using other convolutional networks to extract image features and pre-trained word embedding vectors, but the results were similar and did not improve.
log0.txt