modernmt: [MMT4] Doesn't start with country code

In the latest release, MMT4 doesn’t start when having a country code, while it was starting previously:

java.lang.IllegalArgumentException: en_GB
	at eu.modernmt.lang.Language.fromString(Language.java:96)
	at eu.modernmt.decoder.neural.ModelConfig.getAvailableModels(ModelConfig.java:68)
	at eu.modernmt.decoder.neural.NeuralDecoder.<init>(NeuralDecoder.java:70)
	at eu.modernmt.engine.Engine.load(Engine.java:105)
	at eu.modernmt.cluster.ClusterNode.start(ClusterNode.java:280)
	at eu.modernmt.facade.ModernMT.start(ModernMT.java:54)
	at eu.modernmt.cli.ClusterNodeMain.main(ClusterNodeMain.java:182)

About this issue

  • Original URL
  • State: closed
  • Created 5 years ago
  • Comments: 17 (7 by maintainers)

Most upvoted comments

@davidecaroselli I finally got MMT4 working, but my initial impressions are the same. I’m not sure whether loss values of MMT3 and MMT4 can be compared directly, but MMT3 would converge at 1 with my English-Dutch dataset, while MMT4 converges at 4 (!). As said, I’m not sure how the values compare, but an intermediate run in a live production environment confirms that the results from MMT4 are far, far worse, to the extent that translating manually is quicker. MMT3 would result in an actual production boost.

Results from MMT4 are generated faster, but I’d rather sacrifice speed for quality.

This is based on a 30 million word corpus.

Ok, found this one: runtime/<model>/logs/embedded-cassandra.log Now waiting for a new blocking situation…

@EtienneAb3d

The right syntax is en-GB