about the training efficiency

[yyyf-g]

Hi,you mentioned in your paper that you can run an epoch in 9.7 hours, but when I reproduced it with pytorch, it took me five or six minutes to run a batch. Do you have any tricks to improve the training efficiency?

One more question, when I run your original code with tensorflow, it always reminds me "failed to get article or abstract from example", do you know why? I don't know if it is related to using a lower version of tensorflow, because it will also report warnings that many methods have been deprecated.

thanks very much

[iriscxy]

Hi, for the first question, we didn't use any speed-up trick, and each batch might take about 20 seconds. Maybe you should double-check which function takes up the most time.

For the second question, we use tensorflow 1.6 version. And you should set the json_input_key and json_target_key in corresponding to the keys in the json data. For wikisum datset, the key is document and summary.

[otakusbear]

Hi,you mentioned in your paper that you can run an epoch in 9.7 hours, but when I reproduced it with pytorch, it took me five or six minutes to run a batch. Do you have any tricks to improve the training efficiency?

One more question, when I run your original code with tensorflow, it always reminds me "failed to get article or abstract from example", do you know why? I don't know if it is related to using a lower version of tensorflow, because it will also report warnings that many methods have been deprecated.

thanks very much

Hi, when I try to run this code, The tensorflow1.x legacy code made me mad. And it can‘t work in my CUDA version. So I want to ask could you share your reproduced pytorch code. I really need it and appreciate.
thanks very much

[iriscxy]

Hi, unfortunately, we didn't prepare pytorch version of our code. We used TensorFlow 1.6.0, and I think you can directly run the code based on this version.