Search code examples
machine-translation

Sockeye WMT German to English news translation stopping criterium

I'm doing the tutorial at https://awslabs.github.io/sockeye/tutorials/wmt.html.

The following command gives the output below.

python3 -m sockeye.train -d train_data \
                        -vs newstest2016.tc.BPE.de \
                        -vt newstest2016.tc.BPE.en \
                        --max-seq-len 60 \
                        --decode-and-evaluate 500 \
                        --use-cpu \
                        -o wmt_model

sockeye.utils.SockeyeError: Please specify at least one stopping criteria: --max-samples --max-updates --max-checkpoints --max-num-epochs --max-num-checkpoint-not-improved

What is a reasonable stopping criterium for this task?

Solution

  • The typical stopping criterion is early stopping, i.e., train the model until the validation loss improves. In Sockey, it is --max-num-checkpoint-not-improved. See Sockey documentation for more details.