The BLEU score of translation model seems abnormal. The model doesn't seem to train effectively. #63
Description
I trained translation model with the version 'GPUs=1' of pipedream. The loss during training is all right. After 10 epochs, the loss function drops normally until it converges. But when I run 'compute_bleu_scores.py' to calculate the BLUE score, I get the score of '0.01'. The model saved in each epochs gets very low score such as '0.01', '0.04'. When I output the translated sentences directly, the translation result is obviously incorrect. The translation results can be seen in the picture. There are many ',' and '.'. I want to know what's wrong, my reason or the source code. Thank you very much!!
