tobyyouup

There's no pooling layer in the conv seq2seq model reported in the paper https://arxiv.org/abs/1705.03122. The avg_pooling is an extra implementation which is not relevant if you just want to use...

@dylanee2 @anglil I have run the iwslt14 de-en task, the command is shown in the README, and I can get a BLEU score 25 at step 60k with batchsize 32.

Hi anglil, just one tesla K40 GPU. The traning data is IWSLT Germain-English, I concatenate dev2010, tst2010, tst2011 and tst2012 as the test set.

Hi calmevtime， I have test inference code, and the greedy search and beam search, it works fine. Have you encountered any problems?

@loveJasmine Just split the file into two files, with each line aligned.

@jacklone tf.contrib.learn.estimator has the problem that would reload the model every time when calling the functions. Have you ever tried this solutions? https://github.com/marcsto/rl/blob/master/src/fast_predict.py

@loveJasmine Have you load the correct checkpoint? Or can you try on a new folder without loading an existing checkpoint?

@pooyadavoodi I have also found this problem, evaluation is actually doing: feeding all the target tokens to the decoder and comparing the prediction with the target. I thinks this way...

@hscspring I want to know how about your experiment about different writers. Does it work? Do you say the sentence making sense?