train_en2de.sh

#!/usr/bin/env bash
python3 -m sockeye.train --source data/en2cs/train.en.tok \
        --target data/en2cs/train.BPE.de.tok \
        --source-graphs data/en2cs/train.en.deps \
        --validation-source data/en2cs/val.en.tok \
        --validation-target data/en2cs/val.BPE.de.tok \
        --val-source-graphs data/en2cs/val.en.deps \
        --edge-vocab data/en2cs/edge_vocab.json \
        --batch-size 24 \
        --batch-type sentence \
        --word-min-count 3:3 \
        --num-embed 360:360 \
        --embed-dropout .5:.5 \
        --max-seq-len 199:199 \
        --encoder gcn \
        --gcn-activation relu \
        --gcn-num-hidden 360 \
        --gcn-pos-embed 360 \
        --decoder rnn \
        --num-layers 2:2 \
        --rnn-num-hidden 360 \
        --rnn-decoder-hidden-dropout 0.2 \
        --checkpoint-frequency 1500 \
        --max-num-checkpoint-not-improved 28 \
        --initial-learning-rate 0.0003 \
        --learning-rate-reduce-factor 0.7 \
        --learning-rate-reduce-num-not-improved 7 \
        --gcn-num-layers 2 \
        --weight-init-xavier-factor-type in \
        --weight-init-scale 2.34 \
        --decode-and-evaluate 300 \
        --output en2de_model \
        --overwrite-output \
        --device-ids 0 \
        --gcn-dropout 0.1 \
        --gcn-adj-norm \
        --rnn-attention-type coverage