sockeye-train-transformer.sh

#!/bin/bash
# Weijia

### Parameters
# gpus=1,2,4

# Training
gpu_ids=$(echo $gpus | sed "s/,/ /g")

if [ ! -d $model_dir ]; then
	echo "Start training..."
    CUDA_VISIBLE_DEVICES=0,1 python3 -m sockeye.train \
        -s $train_src \
        -t $train_tgt \
        -vs $dev_src \
        -vt $dev_tgt \
        -o $model_dir \
        $model_args \
        --num-words 50000:50000 \
        --weight-tying \
        --weight-tying-type trg_softmax \
        --label-smoothing 0.1 \
        --encoder transformer \
        --decoder transformer \
        --num-layers 6 \
        --transformer-attention-heads 8 \
        --transformer-model-size 512 \
        --num-embed 512 \
        --transformer-feed-forward-num-hidden 2048 \
        --transformer-preprocess n \
        --transformer-postprocess dr \
        --gradient-clipping-type none \
        --transformer-dropout-attention 0.1 \
        --transformer-dropout-act 0.1 \
        --transformer-dropout-prepost 0.1 \
        --max-seq-len $src_max_len:$tgt_max_len \
        --batch-type word \
        --batch-size 2048 \
        --min-num-epochs 3 \
        --initial-learning-rate .0002 \
        --learning-rate-reduce-factor .7 \
        --learning-rate-reduce-num-not-improved 4 \
        --checkpoint-frequency 1000 \
        --keep-last-params 30 \
        --max-num-checkpoint-not-improved 20 \
        --decode-and-evaluate 1000 \
        --seed $seed \
        --disable-device-locking \
        --device-ids $gpu_ids 
fi;
#