Cannot reproduce this experiment about finetune-transformer-lm HOT 4 CLOSED

I get accuracy of ~85% on ROC 2016 test set without having to change any parameters aside from --n_gpu, which I set to 1, just because I wanted to use only 1 GPU (if --n_gpu 2, then the default training batch size is 8x2=16, which caused out of memory issues on one GTX 1080 (8GB of GPU memory). By default n_gpu=4).

from finetune-transformer-lm.

Thank you so much for your reply.
I got 87% after I set n_gpu to 1. I think there may be a problem in implementation of multi-gpu. I test default n_gpu = 4 on a server with 8 Tesla K80 4 times but only get 53%.

from finetune-transformer-lm.

i got 53% in python2.7 with tensorflow 1.2，while 87% with python 3.6 and tensorflow 1.6.maybe you should check the version of tensorflow

from finetune-transformer-lm.

This looks to have been due to an implementation difference between python 2 and python 3. Only python 3 is supported but I've removed the code that was likely causing this issue as discussed in #21 so python 2 should work fine now.

from finetune-transformer-lm.