[translation] tf-nightly-gpu==1.9.0.dev20180419 is no longer available about training

[nvcforster]

See also #84

[lucienwang1009]

OK, didn't see it before. Feel free to close this issue.

[nvcforster]

Hi Lucien (@lucienwang1009 ), I wanted to link the two issues. Your finding makes it more urgent to update the TF version specified in the Dockerfile.

[bitfort]

This is unfortunate. Thank for for letting us know. If this is blocking you, feel free to link a PR here and I'll take a look at it. Otherwise we will try to address this in the coming days.

[lucienwang1009]

I agree with @nvcforster that it ought to use a stable version rather than an alterable nightly one. Besides, I prefer providing variables in Dockerfile to assign versions of TF, CUDA, and CUDNN etc.
By the definition, a reference result is the quality or the performance of a benchmark running in a system with a framework. So the Dockerfile, as a definition of the system environment, should provide a way to specify these software versions.

Anyway, I really want you guys to consider and address this issue ASAP, at least stick to an available stable version.
BTW: in translation benchmark, tf-nightly-gpu is installed again in docker container, which is obviously redundant.

[David-Levinthal]

could somebody at least post a functioning edit for the dockerfile line?
RUN pip3 install --upgrade numpy scipy sklearn tf-nightly-gpu==1.9.0.dev20180419

[bitfort]

I'll take a look at this in the next day or two... sorry for the issue here.

In the mean time, one option is taking a look at the source of this fork:
https://github.com/tensorflow/models/tree/master/official/transformer

The "official" version will be very similar (since we forked it with only minor modification) and better maintained, it may give insight into these kinds of problems.

[bitfort]

PTAL #119

[johntran-nv]

Closing because GNMT is deprecated from the benchmark suite.