You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have tried two different options for training the punctuator model on Ubuntu 16.04.
In one, I have built the lingvo_py_binary 'trainer' (cuda enabled) binary which has its python_version parameter set 'PY2' and has model_imports and trainer_lib in deps. Intuitively, this build allows me to run my training fine using Python2.7 interpreter. Then I remove this build and instead I build the 'ipython_kernel' binary which includes the same deps as trainer (and some more) but the python_version is set 'PY3'. Strangely, even with this build I can run my training fine using the same Python2.7 interpreter. Bazel docs says this flag is supposed to control which python version you build the binary for.
Question:
Should the python_version in the BUILD file be changed to the interpreter used? Am I going to have some issues with the latter ('PY3') build if I run it with python2.7 interpreter despite the training looking okay for now, perhaps certain functions in the background aren't working as they are supposed to or the performance may be affected?
Or is it safe to ignore the value of the 'python_version' in the BUILD file and go ahead no matter what it is?
The text was updated successfully, but these errors were encountered:
Okay that sounds good. The bazel docs relevant section does mention the bug that you may be referring to:
Bug warning: This attribute sets the version for which Bazel builds your target, but due to #4815, the resulting stub script may still invoke the wrong interpreter version at runtime. See this workaround, which involves defining a py_runtime target that points to either Python version as needed, and activating this py_runtime by setting --python_top.
But the referred issue recently has been deprioritized saying it's now fixed (Except for Windows). Not sure this affects any behavior then but since you say lingvo is compatible with both PY2 and PY3 anyway that's reassuring. Thank you
I have tried two different options for training the punctuator model on Ubuntu 16.04.
In one, I have built the lingvo_py_binary 'trainer' (cuda enabled) binary which has its python_version parameter set 'PY2' and has model_imports and trainer_lib in deps. Intuitively, this build allows me to run my training fine using Python2.7 interpreter. Then I remove this build and instead I build the 'ipython_kernel' binary which includes the same deps as trainer (and some more) but the python_version is set 'PY3'. Strangely, even with this build I can run my training fine using the same Python2.7 interpreter. Bazel docs says this flag is supposed to control which python version you build the binary for.
Question:
Should the python_version in the BUILD file be changed to the interpreter used? Am I going to have some issues with the latter ('PY3') build if I run it with python2.7 interpreter despite the training looking okay for now, perhaps certain functions in the background aren't working as they are supposed to or the performance may be affected?
Or is it safe to ignore the value of the 'python_version' in the BUILD file and go ahead no matter what it is?
The text was updated successfully, but these errors were encountered: