search

ZhenYangIACAS / NMT_GAN

generative adversarial nets for neural machine translation
Apache License 2.0
119 stars 37 forks source link

when to stop pretraining generator #9

Closed jeicy07 closed 6 years ago

jeicy07 commented 6 years ago

Hi, when pretraining the generator after 200 epochs, the accuracy almost reaches 1, and loss is extremely small. However, when pretraining discriminator, the accuracy keeps 0.5 from the beginning to the end. I guess the generator is so powerful that the negative samples generated by it are very close to the positive samples. Then the gan training can't move on under this circumstance. So I wonder when should I stop pretraining generator? (at about 70% accuracy or something else?) Thanks.

ZhenYangIACAS commented 6 years ago

@jeicy07 You mean the accuracy of your generator reaches near 1? How large is your training data size? In our experiments, I train the generator as better as possible. However, the accuracy never reaches around 1.

jeicy07 commented 6 years ago

I use 10w data first to try running whole process. Now I start using 300w data. After 13 epochs, the accuracy keeps at 0.5. Is that normal?

ZhenYangIACAS commented 6 years ago

Haven't you ever trained the NMT model? I think that you should test the BLEU score of your model not just the accuracy. It is normal for me that the NMT model achieves accuracy around 0.7.

jeicy07 commented 6 years ago

It's exactly my first time to train an NMT model. Maybe there's something wrong with my data cuz the BLEU score is only 0.09 on evaluation set. Thanks!

ZhenYangIACAS commented 6 years ago

@jeicy07 That is all right. Check your data first and you will find something wrong.