Files and scripts in this folder show how to train a Google-style transformer model (Vaswani et al, 2017) on WMT-17 (?) English-German data. The problem set has created following the example from the original tensor2tensor repository by Google. It uses 32,000 common BPE units for both languages.

Assuming four GPUs are available (here 0 1 2 3), execute the command below to run the complete example:

./run-me.sh 0 1 2 3

This starts a training run with marian using the following command:

../build/marian \
    --model model/model.npz --type transformer \
    --train-sets data/corpus.bpe.en data/corpus.bpe.de \
    --max-length 100 \
    --vocabs model/vocab.ende.yml model/vocab.ende.yml \
    --mini-batch-fit -w 10000 --maxi-batch 1000 \
    --early-stopping 10 \
    --valid-freq 5000 --save-freq 5000 --disp-freq 500 \
    --valid-metrics cross-entropy perplexity translation \
    --valid-sets data/valid.bpe.en data/valid.bpe.de \
    --valid-script-path ./scripts/validate.sh \
    --valid-translation-output data/valid.bpe.en.output --quiet-translation \
    --valid-mini-batch 64 \
    --beam-size 6 --normalize 0.6 \
    --log model/train.log --valid-log model/valid.log \
    --enc-depth 6 --dec-depth 6 \
    --transformer-heads 8 \
    --transformer-postprocess-emb d \
    --transformer-postprocess dan \
    --transformer-dropout 0.1 --label-smoothing 0.1 \
    --learn-rate 0.0003 --lr-warmup 16000 --lr-decay-inv-sqrt 16000 --lr-report \
    --optimizer-params 0.9 0.98 1e-09 --clip-norm 5 \
    --tied-embeddings-all \
    --devices $GPUS --sync-sgd --seed 1111 \
    --exponential-smoothing

This reproduces a system roughly equivalent to the basic 6-layer transformer described in the original paper.

The training setting includes:

• Fitting mini-batch sizes to 10GB of GPU memory with synchronous SGD (ADAM), which results in large mini-batches. This has been chosen to fit on GPUs with about 12GB of RAM, we leave about 2GB for the remaining training parameters.
• Validation on external data set using cross-entropy, perplexity and BLEU
• 6-layer (or rather block) encoder and 6-layer decoder
• Tied embeddings for source, target and output layer
• Label smoothing
• Learning rate warm-up and cool-down
• Multi-GPU training

You can reduce the workspace size to about -w 6000 if you want to fit on smaller GPUs with 8GB of RAM or increase it further more on GPUs with more RAM. On a Volta GPU with 16GB of RAM you can set -w 14000 when using four GPUs.

The evaluation is performed on WMT test sets from 2014, 2015 and 2016 using sacreBLEU, which provides hassle-free computation of shareable, comparable, and reproducible BLEU scores. The WMT-2013 test set is used as a validation set.

See the basic training example (marian/examples/training-basics/) for more details.