Openmt vs openmnt-py similarities and differences

[miguelknals (miguel canals)]

Hi again,

I have a couple of questions related to opennmt-py, excuse if they are obvious.

a) I assume the previous lua opennmt tools wont be migrated (as tokenize.lua and detokenize.lua). I have seen several opennmt-py tools are coming from moses but they are not like the lua versions. I understand that caseing or joiner annotate should be done outside the openmnt-py package. Is this correct?

b) Train uses a loooot of memory If I compare to “old” opennmt. I dont have numbers here by, but for simple models, are using at least 2 times more memory. Is any way to reduce the memory consumption? I was thinking some openmnt models that I was able to run they already were 6MB, this means I will probably wont be able to run them in and end consumer GPU.

c) As rest_translation_server.lua (the api server for the “old” openmt) probably will also become obsolete, I assume that right now the clients we have are the Simple OpenNMT.py REST server that paul has provided (Simple OpenNMT-py REST server) or the CTranslate (https://github.com/OpenNMT/CTranslate). Not sure if CTranslate only was working with the “old” openmnt models. Is this correct?

d) I assume openmt source/target files can be used in opennmt-py. Are the opennmt models and the opennmt-py models compatible? For instance, can I use rest_translation_server.lua with opennmt-py models?

e) One of the things I like about opennmt-py is the ability to create complex models, but to be honest, not sure where to recall some info about how to modify or enhance the standard model (2 x 500) or create a new one. Does anyone knows where to find a suitable model for instance to translate 2 similar languages?

f) Is the openmt-py docker image current? As far as I know, nvidia docker offers close to bare metal performance, so it would be nice to have one current image.

Thanks in advance!
Have a nice day
Miguel

[guillaumekln (Guillaume Klein)]

Hi,

a) I assume the previous lua opennmt tools wont be migrated (as tokenize.lua and detokenize.lua). I have seen several opennmt-py tools are coming from moses but they are not like the lua versions. I understand that caseing or joiner annotate should be done outside the openmnt-py package. Is this correct?

Yes, it should be done outside. Tokenization tools have been migrated to OpenNMT/Tokenizer which also provides a Python wrapper.

b) Train uses a loooot of memory If I compare to “old” opennmt. I dont have numbers here by, but for simple models, are using at least 2 times more memory. Is any way to reduce the memory consumption? I was thinking some openmnt models that I was able to run they already were 6MB, this means I will probably wont be able to run them in and end consumer GPU.

Try -max_generator_batches 1.

c) As rest_translation_server.lua (the api server for the “old” openmt) probably will also become obsolete, I assume that right now the clients we have are the Simple OpenNMT.py REST server that paul has provided (Simple OpenNMT-py REST server) or the CTranslate (https://github.com/OpenNMT/CTranslate). Not sure if CTranslate only was working with the “old” openmnt models. Is this correct?

CTranslate only works for OpenNMT-lua models. So you should use the OpenNMT-py REST server which should be good enough.

d) I assume openmt source/target files can be used in opennmt-py. Are the opennmt models and the opennmt-py models compatible? For instance, can I use rest_translation_server.lua with opennmt-py models?

Text data files are of course compatible but not models.

e) One of the things I like about opennmt-py is the ability to create complex models, but to be honest, not sure where to recall some info about how to modify or enhance the standard model (2 x 500) or create a new one. Does anyone knows where to find a suitable model for instance to translate 2 similar languages?

See for example:

http://opennmt.net/OpenNMT-py/FAQ.html#how-do-i-use-the-transformer-model-do-you-support-multi-gpu

f) Is the openmt-py docker image current? As far as I know, nvidia docker offers close to bare metal performance, so it would be nice to have one current image.

I’m not sure about the version that is pushed to Dockerhub. You should probably rebuild one from the Dockerfile which is pretty simple.