GNMT v2

The model

The GNMT v2 model is similar to the one discussed in the Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation paper.

The most important difference between the two models is in the attention mechanism. In our model, the output from the first LSTM layer of the decoder goes into the attention module, then the re-weighted context is concatenated with inputs to all subsequent LSTM layers in the decoder at the current timestep.

The same attention mechanism is also implemented in the default GNMT-like models from TensorFlow Neural Machine Translation Tutorial and NVIDIA OpenSeq2Seq Toolkit.

Default configuration of the GNMT v2 model

• general:
  • encoder and decoder are using shared embeddings
  • data-parallel multi-gpu training
  • dynamic loss scaling with backoff for Tensor Cores (mixed precision) training
  • trained with label smoothing loss (smoothing factor 0.1)
• encoder:
  • 4-layer LSTM, hidden size 1024, first layer is bidirectional, the rest are unidirectional
  • with residual connections starting from 3rd layer
  • uses LSTM layer accelerated by cuDNN
• decoder:
  • 4-layer unidirectional LSTM with hidden size 1024 and fully-connected classifier
  • with residual connections starting from 3rd layer
  • uses LSTM layer accelerated by cuDNN
• attention:
  • normalized Bahdanau attention
  • output from first LSTM layer of decoder goes into attention, then re-weighted context is concatenated with the input to all subsequent LSTM layers of the decoder at the current timestep
• inference:
  • beam search with default beam size of 5
  • with coverage penalty and length normalization terms
  • detokenized BLEU computed by SacreBLEU
  • motivation for choosing SacreBLEU

When comparing the BLEU score there are various tokenization approaches and BLEU calculation methodologies, ensure to align similar metrics.

Code from this repository can be used to train a larger, 8-layer GNMT v2 model. Our experiments show that a 4-layer model is significantly faster to train and yields comparable accuracy on the public WMT16 English-German dataset. The number of LSTM layers is controlled by the num_layers parameter in the scripts/train.sh training script.

Setup

Requirements

• PyTorch 18.06-py3 NGC container (or newer)
• SacreBLEU 1.2.10

This repository contains Dockerfile which extends the PyTorch NGC container and encapsulates all dependencies.

For more information about how to get started with NGC containers, see the following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning DGX Documentation: Getting Started Using NVIDIA GPU Cloud, Accessing And Pulling From The NGC container registry and Running PyTorch.

Training using mixed precision with Tensor Cores

Before you can train using mixed precision with Tensor Cores, ensure that you have a NVIDIA Volta based GPU. For information about how to train using mixed precision, see the Mixed Precision Training paper and Training With Mixed Precision documentation.

Another option for adding mixed-precision support is available from NVIDIA’s APEX, A PyTorch Extension, that contains utility libraries, such as AMP, which require minimal network code changes to leverage Tensor Core performance.

Quick start guide

Perform the following steps to run the training using the default parameters of the GNMT v2 model on the WMT16 English-German dataset.

1. Build and launch the GNMT Docker container

bash scripts/docker/build.sh
bash scripts/docker/interactive.sh

2. Download the training dataset

Download and preprocess the WMT16 English-German dataset. Data will be downloaded to the data directory (on the host). The data directory is mounted to the /workspace/gnmt/data location in the Docker container.

bash scripts/wmt16_en_de.sh

3. Run training

By default, the training script will use all available GPUs. The training script saves only one checkpoint with the lowest value of the loss function on the validation dataset. All results and logs are saved to the results directory (on the host) or to the /workspace/gnmt/results directory (in the container). By default, the scripts/train.sh script will launch mixed precision training with Tensor Cores. You can change this behaviour by setting the --math fp32 flag in the scripts/train.sh script.

bash scripts/train.sh

The training script automatically runs the validation and testing after each training epoch. The results from the validation and testing are printed to the standard error (stderr) and saved to log files.

The summary after each training epoch is printed in the following format:

Summary: Epoch: 3	Training Loss: 3.1735	Validation Loss: 3.0511	Test BLEU: 21.89
Performance: Epoch: 3  Training: 300155 Tok/s  Validation: 156066 Tok/s

The training loss is averaged over an entire training epoch, the validation loss is averaged over the validation dataset and the BLEU score is computed by the SacreBLEU package on the test dataset. Performance is reported in total tokens per second. The result is averaged over an entire training epoch and summed over all GPUs participating in the training.

Details

Getting the data

The GNMT v2 model was trained on the WMT16 English-German dataset. Concatenation of the newstest2015 and newstest2016 test sets are used as a validation dataset and the newstest2014 is used as a testing dataset.

This repository contains the scripts/wmt16_en_de.sh download script which will automatically download and preprocess the training, validation and test datasets. By default, data will be downloaded to the data directory.

Our download script is very similar to the wmt16_en_de.sh script from the tensorflow/nmt repository. Our download script contains an extra preprocessing step, which discards all pairs of sentences which can't be decoded by latin-1 encoder.

The scripts/wmt16_en_de.sh script uses the subword-nmt package to segment text into subword units (BPE). By default, the script builds the shared vocabulary of 32,000 tokens.

Running training

The default training configuration can be launched by running the scripts/train.sh training script. By default, the training script saves only one checkpoint with the lowest value of the loss function on the validation dataset, an evaluation is performed after each training epoch. Results are stored in the results/gnmt_wmt16 directory.

The training script launches data-parallel training with batch size 128 per GPU on all available GPUs. After each training epoch, the script runs an evaluation on the validation dataset and outputs a BLEU score on the test dataset (newstest2014). BLEU is computed by the SacreBLEU package. Logs from the training and evaluation are saved to the results directory.

Even though the training script uses all available GPUs, you can change this behavior by setting the CUDA_VISIBLE_DEVICES variable in your environment or by setting the NV_GPU variable at the Docker container launch (see section "GPU isolation").

By default, the scripts/train.sh script will launch mixed precision training with Tensor Cores. You can change this behaviour by setting the --math fp32 flag in the scripts/train.sh script.

Internally, the scripts/train.sh script uses train.py. To view all available options for training, run python3 train.py --help.

Running inference

Inference can be run by launching the translate.py inference script, although, it requires a pre-trained model checkpoint and tokenized input.

The inference script, translate.py, supports batched inference. By default, it launches beam search with beam size of 5, coverage penalty term and length normalization term. Greedy decoding can be enabled by setting the beam size to 1.

To view all available options for inference, run python3 translate.py --help.

Benchmarking scripts

Training performance benchmark

The scripts/benchmark_training.sh benchmarking script runs a few, relatively short training sessions and automatically collects performance numbers. The benchmarking script assumes that the scripts/wmt16_en_de.sh data download script was launched and the datasets are available in the default location (data directory).

Results from the benchmark are stored in the results directory. After the benchmark is done, you can launch the scripts/parse_train_benchmark.sh script to generate a short summary which will contain launch configuration, performance (in tokens per second), and estimated training time needed for one epoch (in seconds).

Inference performance and accuracy benchmark

The scripts/benchmark_inference.sh benchmarking script launches a number of inference runs with different hyperparameters (beam size, batch size, arithmetic type) on sorted and unsorted newstest2014 test dataset. Performance and accuracy results are stored in the results/inference_benchmark directory. BLEU score is computed by the SacreBLEU package.

The scripts/benchmark_inference.sh script assumes that the scripts/wmt16_en_de.sh data download script was launched and the datasets are available in the default location (data directory).

The scripts/benchmark_inference.sh script requires a pre-trained model checkpoint. By default, the script is loading a checkpoint from the results/gnmt_wmt16/model_best.pth location.

Training Accuracy Results

All results were obtained by running the scripts/train.sh script in the pytorch-18.06-py3 Docker container on NVIDIA DGX-1 with 8 V100 16G GPUs.

number of GPUs	mixed precision BLEU	fp32 BLEU	mixed precision training time	fp32 training time
1	22.54	22.25	412 minutes	948 minutes
4	22.45	22.46	118 minutes	264 minutes
8	22.41	22.43	64 minutes	139 minutes

Training Stability Test

The GNMT v2 model was trained for 10 epochs, starting from 96 different initial random seeds. After each training epoch the model was evaluated on the test dataset and the BLEU score was recorded. The training was performed in the pytorch-18.06-py3 Docker container on NVIDIA DGX-1 with 8 V100 16G GPUs. The following table summarizes results of the stability test.

Training Performance Results

All results were obtained by running the scripts/train.sh training script in the pytorch-18.06-py3 Docker container on NVIDIA DGX-1 with 8 V100 16G GPUs. Performance numbers (in tokens per second) were averaged over an entire training epoch.

number of GPUs	mixed precision tokens/s	fp32 tokens/s	mixed precision speedup	mixed precision multi-gpu weak scaling	fp32 multi-gpu weak scaling
1	42337	18581	2.279	1.000	1.000
4	153433	67586	2.270	3.624	3.637
8	300181	132734	2.262	7.090	7.144

Inference Performance Results

All results were obtained by running the scripts/benchmark_inference.sh benchmarking script in the pytorch-18.06-py3 Docker container on NVIDIA DGX-1. Inference was run on a single V100 16G GPU.

batch size	beam size	mixed precision BLEU	fp32 BLEU	mixed precision tokens/s	fp32 tokens/s
512	1	20.63	20.63	62009	31229
512	2	21.55	21.60	32669	16454
512	5	22.34	22.36	21105	8562
512	10	22.34	22.40	12967	4720
128	1	20.62	20.63	27095	19505
128	2	21.56	21.60	13224	9718
128	5	22.38	22.36	10987	6575
128	10	22.35	22.40	8603	4103
32	1	20.62	20.63	9451	8483
32	2	21.56	21.60	4818	4333
32	5	22.34	22.36	4505	3655
32	10	22.37	22.40	4086	2822

Changelog

1. Aug 7, 2018

• Initial release

Known issues

None