Why the program throw not enough memory when writting model? about opennmt HOT 6 CLOSED

This a known limitation with LuaJIT which has a low memory limit (1-2GB).

The quick workaround for this issue is to use the default Lua interpreter which can be easily achieved when installing Torch:

TORCH_LUA_VERSION=LUA52 ./install.sh

We are looking for ways to mitigate this issue during the preprocessing.

from opennmt.

You may also be able to workaround by replacing th with lua on the command above.

from opennmt.

(BTW I would really like to find a workaround here. This is an annoying issue.)

from opennmt.

There should be some improvements with the above commit. torch.save is now told to not keep track of serialized objects as it has a memory cost. See this documentation for reference.

from opennmt.

@maplewizard : is ac65838 fixing the problem for you?

from opennmt.

Fixed by #29.

from opennmt.