Modernmt: Porting models from v2.0.2 to current master

[AndRossi]

In the current master version, and starting from commit 09981f7d35c366d5446e50a65d23c157ad90fabc, our neural models were made a bit different from earlier versions to achieve higher performances and translation quality.

If you have updated MMT to one of these recent commits, therefore, you need to upgrade your older engines in order to use them.

Don't worry: we have made this process entirely automatic :)
All you need to do is to run command from your mmt home:

extras/porting/convert_from_v2.0.2_to_v2.0.3 --engine-path <path to your engine folder>

For instance:

extras/porting/convert_from_v2.0.2_to_v2.0.3 --engine-path engines/default

We will soon publish a new release 2.0.3 with these changes. As soon as that release is published, I will close this issue (and add the upgrading information on the Upgrade Engine page, of course).

If you have problems upgrading your engines, do not hesitate to contact us!

Thank you!

Andrea

[EtienneAb3d]

Can you explain what was improved ?
:)

[davidecaroselli]

@EtienneAb3d a detailed changelog will be included in the release as always. Short version: we have fixed a number of bugs (we will close pending issues in the next days), we implemented a hot/cold strategy for Neural Engine in GPU RAM, we changed the termination condition after a wide range of experiments on different language pairs.

[EtienneAb3d]

Can't wait to get technical details...
:D

[mzeidhassan]

Hi @AndRossi ,

I am a bit confused. I did the following:

git pull origin master

but running the command above throws this error:

`From https://github.com/ModernMT/MMT

• branch master -> FETCH_HEAD
  Updating 52116af..1573f5b
  error: Your local changes to the following files would be overwritten by merge:
  src/core/src/main/java/eu/modernmt/facade/TrainingFacade.java
  Please, commit your changes or stash them before you can merge.
  Aborting`

Any idea?

[davidecaroselli]

Hi @mzeidhassan

you have some changes on your code, that's why git is complaining. For a clean build I always recommend this procedure:
```
git checkout -- .
git clean -fd .

rm -rf build/*
git pull origin master

cd vendor
./compile
cd ../src
mvn clean install

[mzeidhassan]

Perfect, @davidecaroselli ! This fixed the issue and the engine has been ported successfully.

Congratulations! Your 2.0.2 engine was successfully ported to version 2.0.3 :-)

[AndRossi]

Hi everyone,
We have released new MMT version 2.1 (you will find it in our master branch as well as in our releases page) so, as I said, I'm closing this issue now.

An important difference in version 2.1 is that you may also need to update your CUDA driver and PyTorch version too.

(For this reason too, we decided to call the new version 2.1 instead of 2.0.3.
The script porting has been renamed extras/porting/convert_from_v2.0.2_to_v2.1.
Apart from that, the behaviour is exactly the same.)

I have already updated the upgrade page in our Wiki with all the information you should need to perform the whole upgrading process.

Thanks for your support!

Andrea