Not getting expected results

[vflux]

I'm trying to run the code on the CNN/Dailymail dataset following the instructions in the readme but the loss doesn't seem to be decreasing and when I try to decode I get output like the following. This is after over 1000000 training steps.

, , , a , , peterson , , .
, , the , .
in .
, .
.
in , , in , .
a , .
to , .

Do you have any idea where the problem could be?

[duzani]

I also tried to pretrain the generator model. But even with 80000 train steps, the loss doesn't seem to be decreasing and I got poor output. I'd like to ask @iwangjian if you succeed to get the appropriate result, like the one in the original paper, with this code.

Thanks!

[iwangjian]

I‘ve fixed some bugs and updated the README, thanks for your comments!

[giriallada]

I'm getting results like the the the . the the , the . . the . , the the the the he the the of the the '' the the to the the` `` How do I achieve paper like results?How many pretrain steps should I take?

[iwangjian]

You can directly use the provided discriminator_train_data.npz to train the full model with pretrain. The full training may be pretty slow.

[hanyegirl]

有没有试过减少迭代次数，比如8万次，这样会有比较好的效果吗

[geekyspartan]

@iwangjian Hey, I am using the 'discriminator_train_data.npz' file and trained for around 650 iterations, but still my output is like this:

b'.'
b'new new : .'
b': .'
b'new : manchester .'
b': : .'
b'.'
b': v .'
b'.'
b'.'
b'new .'
b': may .'
b': manchester manchester .'
b'new'

Commands that I am running: python3 main.py --mode=train --data_path=./data/train.bin --vocab_path=./data/vocab --log_root=./log --pretrain_dis_data_path=./data/discriminator_train_data.npz --restore_best_model=False

For testing: python3 main.py --mode=decode --data_path=./chunked/test_001.bin --vocab_path=./data/vocab --log_root=./log --single_pass=True

chunked data is generated using the same method that you have mentioned in your repo.

Thanks in advance.

[geekyspartan]

Never mind, I was doing a mistake. If I am using discriminator_train_data.npz, then the train command will be this:

python3 main.py --mode=pretrain --data_path=./data/train.bin --vocab_path=./data/vocab --log_root=./log --pretrain_dis_data_path=./data/discriminator_train_data.npz --restore_best_model=False

[Ryze789]

@iwangjian 请问问题如何解决，我是使用了discriminator_train_data.npz，进行full train，得到decode结果也是这样，比较不好 the the the . the the , the . . the . , the the 感谢您的回复