Nmt: Is tensorflow-nightly necessary?

[zhangpengGenedock]

Is tensorflow-nightly necessary? or just after a specific tensorflow version like 1.3.0 can work?
why does it has that restrict, and it leads to the slow training, because the unsupport for gpu.

[oahziur]

@zhangpengGenedock The master requires nightly so we can add the latest features to the tutorial. You can check out the tf-1.2 branch if you want to use tensorflow version 1.3.0.

[zhangpengGenedock]

thanks, and is there a different instructions between tf-1.2 branch and master branch?

[oahziur]

I don't think there are much different on instructions now, but they may be different in the future. The instructions in the README file in the branch should contain the up to date instructions for tensorflow version 1.2 and 1.3.

[kmario23]

I don't think the nightly version is necessary because this line checks for whether TF version is >= 1.2.1
https://github.com/tensorflow/nmt/blob/master/nmt/train.py#L33

def check_tensorflow_version():
    if tf.__version__ < "1.2.1":  
        raise EnvironmentError("Tensorflow version must >= 1.2.1")

So, I guess it will work fine for versions >= 1.2.1