This repository provides a script and recipe to train the GNMT v2 model to achieve state of the art accuracy, and is tested and maintained by NVIDIA.

• Model overview
  • Model architecture
  • Default configuration
  • Feature support matrix
    • Features
  • Mixed precision training
    • Enabling mixed precision
    • Enabling TF32
• Setup
  • Requirements
• Quick Start Guide
• Advanced
  • Scripts and sample code
  • Parameters
  • Command-line options
  • Getting the data
    • Dataset guidelines
  • Training process
  • Inference process
• Performance
  • Benchmarking
    • Training performance benchmark
    • Inference performance benchmark
  • Results
    • Training accuracy results
      • Training accuracy: NVIDIA DGX A100 (8x A100 40GB)
      • Training accuracy: NVIDIA DGX-1 (8x V100 16GB)
      • Training accuracy: NVIDIA DGX-2H (16x V100 32GB)
      • Training stability test
    • Training throughput results
      • Training throughput: NVIDIA DGX A100 (8x A100 40GB)
      • Training throughput: NVIDIA DGX-1 (8x V100 16GB)
      • Training throughput: NVIDIA DGX-2H (16x V100 32GB)
    • Inference accuracy results
      • Inference accuracy: NVIDIA A100 40GB
      • Inference accuracy: NVIDIA Tesla V100 16GB
      • Inference accuracy: NVIDIA T4
    • Inference throughput results
      • Inference throughput: NVIDIA A100 40GB
      • Inference throughput: NVIDIA T4
    • Inference latency results
      • Inference latency: NVIDIA A100 40GB
      • Inference latency: NVIDIA T4
• Release notes
  • Changelog
  • Known issues

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 time step.

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

The following features were implemented in this 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 standard PyTorch nn.LSTM layer
  • dropout is applied on input to all LSTM layers, probability of dropout is set to 0.2
  • hidden state of LSTM layers is initialized with zeros
  • weights and bias of LSTM layers is initialized with uniform(-0.1,0.1) distribution
• decoder:
  • 4-layer unidirectional LSTM with hidden size 1024 and fully-connected classifier
  • with residual connections starting from 3rd layer
  • uses standard PyTorch nn.LSTM layer
  • dropout is applied on input to all LSTM layers, probability of dropout is set to 0.2
  • hidden state of LSTM layers is initialized with zeros
  • weights and bias of LSTM layers is initialized with uniform(-0.1,0.1) distribution
  • weights and bias of fully-connected classifier is initialized with uniform(-0.1,0.1) distribution
• 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
  • linear transform of keys and queries is initialized with uniform(-0.1, 0.1), normalization scalar is initialized with 1.0/sqrt(1024), normalization bias is initialized with zero
• inference:
  • beam search with default beam size of 5
  • with coverage penalty and length normalization, coverage penalty factor is set to 0.1, length normalization factor is set to 0.6 and length normalization constant is set to 5.0
  • de-tokenized BLEU computed by SacreBLEU
  • motivation for choosing SacreBLEU

When comparing the BLEU score, there are various tokenization approaches and BLEU calculation methodologies; therefore, ensure you 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 train.py training script.

The following features are supported by this model.

Apex AMP - a tool that enables Tensor Core-accelerated training. Refer to the Enabling mixed precision section for more details.

Apex DistributedDataParallel - a module wrapper that enables easy multiprocess distributed data parallel training, similar to torch.nn.parallel.DistributedDataParallel. DistributedDataParallel is optimized for use with NCCL. It achieves high performance by overlapping communication with computation during backward() and bucketing smaller gradient transfers to reduce the total number of transfers required.

Mixed precision is the combined use of different numerical precisions in a computational method. Mixed precision training offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of Tensor Cores in Volta, and following with both the Turing and Ampere architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using mixed precision training previously required two steps:

1. Porting the model to use the FP16 data type where appropriate.
2. Manually adding loss scaling to preserve small gradient values.

The ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in CUDA 8 in the NVIDIA Deep Learning SDK.

For information about:

• How to train using mixed precision, see the Mixed Precision Training paper and Training With Mixed Precision documentation.
• Techniques used for mixed precision training, see the Mixed-Precision Training of Deep Neural Networks blog.
• APEX tools for mixed precision training, see the NVIDIA Apex: Tools for Easy Mixed-Precision Training in PyTorch .

Mixed precision is enabled in PyTorch by using the Automatic Mixed Precision (AMP), library from APEX that casts variables to half-precision upon retrieval, while storing variables in single-precision format. Furthermore, to preserve small gradient magnitudes in backpropagation, a loss scaling step must be included when applying gradients. In PyTorch, loss scaling can be easily applied by using scale_loss() method provided by AMP. The scaling value to be used can be dynamic or fixed.

For an in-depth walk through on AMP, check out sample usage here. APEX is a PyTorch extension that contains utility libraries, such as AMP, which require minimal network code changes to leverage Tensor Cores performance.

The following steps were needed to enable mixed precision training in GNMT:

• Import AMP from APEX (file: seq2seq/train/trainer.py):

from apex import amp

• Initialize AMP and wrap the model and the optimizer (file: seq2seq/train/trainer.py, class: Seq2SeqTrainer):

self.model, self.optimizer = amp.initialize(
    self.model,
    self.optimizer,
    cast_model_outputs=torch.float16,
    keep_batchnorm_fp32=False,
    opt_level='O2')

• Apply scale_loss context manager (file: seq2seq/train/fp_optimizers.py, class: AMPOptimizer):

with amp.scale_loss(loss, optimizer) as scaled_loss:
    scaled_loss.backward()

• Apply gradient clipping on single precision master weights (file: seq2seq/train/fp_optimizers.py, class: AMPOptimizer):

if self.grad_clip != float('inf'):
    clip_grad_norm_(amp.master_params(optimizer), self.grad_clip)

TensorFloat-32 (TF32) is the new math mode in NVIDIA A100 GPUs for handling the matrix math also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require high dynamic range for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x blog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.

The following section lists the requirements in order to start training the GNMT v2 model.

This repository contains Dockerfile which extends the PyTorch NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:

• NVIDIA Docker
• PyTorch 20.06-py3 NGC container
• GPU architecture:
  • NVIDIA Volta
  • NVIDIA Turing
  • NVIDIA Ampere architecture

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,
• Running PyTorch.

For those unable to use the Pytorch NGC container, to set up the required environment or create your own container, see the versioned NVIDIA Container Support Matrix.

To train your model using mixed or TF32 precision with Tensor Cores or using FP32, perform the following steps using the default parameters of the GNMT v2 model on the WMT16 English German dataset. For the specifics concerning training and inference, see the Advanced section.

1. Clone the repository.

git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/PyTorch/Translation/GNMT

2. Build the GNMT v2 Docker container.

bash scripts/docker/build.sh

3. Start an interactive session in the container to run training/inference.

bash scripts/docker/interactive.sh

4. Download and preprocess the 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

5. Start training.

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 gnmt directory (on the host) or to the /workspace/gnmt/gnmt directory (in the container). By default, the train.py script will launch mixed precision training with Tensor Cores. You can change this behavior by setting:

• the --math fp32 flag to launch single precision training (for NVIDIA Volta and NVIDIA Turing architectures) or
• the --math tf32 flag to launch TF32 training with Tensor Cores (for NVIDIA Ampere architecture)

for the train.py training script.

To launch mixed precision training on 1, 4 or 8 GPUs, run:

python3 -m torch.distributed.launch --nproc_per_node=<#GPUs> train.py --seed 2 --train-global-batch-size 1024

To launch mixed precision training on 16 GPUs, run:

python3 -m torch.distributed.launch --nproc_per_node=16 train.py --seed 2 --train-global-batch-size 2048

By default, the training script will launch training with batch size 128 per GPU. If --train-global-batch-size is specified and larger than 128 times the number of GPUs available for the training then the training script will accumulate gradients over consecutive iterations and then perform the weight update. For example, 1 GPU training with --train-global-batch-size 1024 will accumulate gradients over 8 iterations before doing the weight update with accumulated gradients.

6. Start evaluation.

The training process automatically runs evaluation and outputs the BLEU score after each training epoch. Additionally, after the training is done, you can manually run inference on the test dataset with the checkpoint saved during the training.

To launch FP16 inference on the newstest2014.en test set, run:

python3 translate.py \
  --input data/wmt16_de_en/newstest2014.en \
  --reference data/wmt16_de_en/newstest2014.de \
  --output /tmp/output \
  --model gnmt/model_best.pth

The script will load the checkpoint specified by the --model option, then it will launch inference on the file specified by the --input option, and compute BLEU score against the reference translation specified by the --reference option. Outputs will be stored to the location specified by the --output option.

Additionally, one can pass the input text directly from the command-line:

python3 translate.py \
  --input-text "The quick brown fox jumps over the lazy dog" \
  --model gnmt/model_best.pth

Translated output will be printed to the console:

(...)
0: Translated output:
Der schnelle braune Fuchs springt über den faulen Hund

By default, the translate.py script will launch FP16 inference with Tensor Cores. You can change this behavior by setting:

• the --math fp32 flag to launch single precision inference (for NVIDIA Volta and NVIDIA Turing architectures) or
• the --math tf32 flag to launch TF32 inference with Tensor Cores (for NVIDIA Ampere architecture)

for the translate.py inference script.

The following sections provide greater details of the dataset, running training and inference, and the training results.

In the root directory, the most important files are:

• train.py: serves as the entry point to launch the training
• translate.py: serves as the entry point to launch inference
• Dockerfile: container with the basic set of dependencies to run GNMT v2
• requirements.txt: set of extra requirements for running GNMT v2

The seq2seq/model directory contains the implementation of GNMT v2 building blocks:

• attention.py: implementation of normalized Bahdanau attention
• encoder.py: implementation of recurrent encoder
• decoder.py: implementation of recurrent decoder with attention
• seq2seq_base.py: base class for seq2seq models
• gnmt.py: implementation of GNMT v2 model

The seq2seq/train directory encapsulates the necessary tools to execute training:

• trainer.py: implementation of training loop
• smoothing.py: implementation of cross-entropy with label smoothing
• lr_scheduler.py: implementation of exponential learning rate warmup and step decay
• fp_optimizers.py: implementation of optimizers for various floating point precisions

The seq2seq/inference directory contains scripts required to run inference:

• beam_search.py: implementation of beam search with length normalization and length penalty
• translator.py: implementation of auto-regressive inference

The seq2seq/data directory contains implementation of components needed for data loading:

• dataset.py: implementation of text datasets
• sampler.py: implementation of batch samplers with bucketing by sequence length
• tokenizer.py: implementation of tokenizer (maps integer vocabulary indices to text)

Training

The complete list of available parameters for the train.py training script contains:

dataset setup:
  --dataset-dir DATASET_DIR
                        path to the directory with training/test data
                        (default: data/wmt16_de_en)
  --src-lang SRC_LANG   source language (default: en)
  --tgt-lang TGT_LANG   target language (default: de)
  --vocab VOCAB         path to the vocabulary file (relative to DATASET_DIR
                        directory) (default: vocab.bpe.32000)
  -bpe BPE_CODES, --bpe-codes BPE_CODES
                        path to the file with bpe codes (relative to
                        DATASET_DIR directory) (default: bpe.32000)
  --train-src TRAIN_SRC
                        path to the training source data file (relative to
                        DATASET_DIR directory) (default:
                        train.tok.clean.bpe.32000.en)
  --train-tgt TRAIN_TGT
                        path to the training target data file (relative to
                        DATASET_DIR directory) (default:
                        train.tok.clean.bpe.32000.de)
  --val-src VAL_SRC     path to the validation source data file (relative to
                        DATASET_DIR directory) (default:
                        newstest_dev.tok.clean.bpe.32000.en)
  --val-tgt VAL_TGT     path to the validation target data file (relative to
                        DATASET_DIR directory) (default:
                        newstest_dev.tok.clean.bpe.32000.de)
  --test-src TEST_SRC   path to the test source data file (relative to
                        DATASET_DIR directory) (default:
                        newstest2014.tok.bpe.32000.en)
  --test-tgt TEST_TGT   path to the test target data file (relative to
                        DATASET_DIR directory) (default: newstest2014.de)
  --train-max-size TRAIN_MAX_SIZE
                        use at most TRAIN_MAX_SIZE elements from training
                        dataset (useful for benchmarking), by default uses
                        entire dataset (default: None)

results setup:
  --save-dir SAVE_DIR   path to directory with results, it will be
                        automatically created if it does not exist (default:
                        gnmt)
  --print-freq PRINT_FREQ
                        print log every PRINT_FREQ batches (default: 10)

model setup:
  --hidden-size HIDDEN_SIZE
                        hidden size of the model (default: 1024)
  --num-layers NUM_LAYERS
                        number of RNN layers in encoder and in decoder
                        (default: 4)
  --dropout DROPOUT     dropout applied to input of RNN cells (default: 0.2)
  --share-embedding     use shared embeddings for encoder and decoder (use '--
                        no-share-embedding' to disable) (default: True)
  --smoothing SMOOTHING
                        label smoothing, if equal to zero model will use
                        CrossEntropyLoss, if not zero model will be trained
                        with label smoothing loss (default: 0.1)

general setup:
  --math {fp16,fp32,tf32,manual_fp16}
                        precision (default: fp16)
  --seed SEED           master seed for random number generators, if "seed" is
                        undefined then the master seed will be sampled from
                        random.SystemRandom() (default: None)
  --prealloc-mode {off,once,always}
                        controls preallocation (default: always)
  --dllog-file DLLOG_FILE
                        Name of the DLLogger output file (default:
                        train_log.json)
  --eval                run validation and test after every epoch (use '--no-
                        eval' to disable) (default: True)
  --env                 print info about execution env (use '--no-env' to
                        disable) (default: True)
  --cuda                enables cuda (use '--no-cuda' to disable) (default:
                        True)
  --cudnn               enables cudnn (use '--no-cudnn' to disable) (default:
                        True)
  --log-all-ranks       enables logging from all distributed ranks, if
                        disabled then only logs from rank 0 are reported (use
                        '--no-log-all-ranks' to disable) (default: True)

training setup:
  --train-batch-size TRAIN_BATCH_SIZE
                        training batch size per worker (default: 128)
  --train-global-batch-size TRAIN_GLOBAL_BATCH_SIZE
                        global training batch size, this argument does not
                        have to be defined, if it is defined it will be used
                        to automatically compute train_iter_size using the
                        equation: train_iter_size = train_global_batch_size //
                        (train_batch_size * world_size) (default: None)
  --train-iter-size N   training iter size, training loop will accumulate
                        gradients over N iterations and execute optimizer
                        every N steps (default: 1)
  --epochs EPOCHS       max number of training epochs (default: 6)
  --grad-clip GRAD_CLIP
                        enables gradient clipping and sets maximum norm of
                        gradients (default: 5.0)
  --train-max-length TRAIN_MAX_LENGTH
                        maximum sequence length for training (including
                        special BOS and EOS tokens) (default: 50)
  --train-min-length TRAIN_MIN_LENGTH
                        minimum sequence length for training (including
                        special BOS and EOS tokens) (default: 0)
  --train-loader-workers TRAIN_LOADER_WORKERS
                        number of workers for training data loading (default:
                        2)
  --batching {random,sharding,bucketing}
                        select batching algorithm (default: bucketing)
  --shard-size SHARD_SIZE
                        shard size for "sharding" batching algorithm, in
                        multiples of global batch size (default: 80)
  --num-buckets NUM_BUCKETS
                        number of buckets for "bucketing" batching algorithm
                        (default: 5)

optimizer setup:
  --optimizer OPTIMIZER
                        training optimizer (default: Adam)
  --lr LR               learning rate (default: 0.002)
  --optimizer-extra OPTIMIZER_EXTRA
                        extra options for the optimizer (default: {})

mixed precision loss scaling setup:
  --init-scale INIT_SCALE
                        initial loss scale (default: 8192)
  --upscale-interval UPSCALE_INTERVAL
                        loss upscaling interval (default: 128)

learning rate scheduler setup:
  --warmup-steps WARMUP_STEPS
                        number of learning rate warmup iterations (default:
                        200)
  --remain-steps REMAIN_STEPS
                        starting iteration for learning rate decay (default:
                        0.666)
  --decay-interval DECAY_INTERVAL
                        interval between learning rate decay steps (default:
                        None)
  --decay-steps DECAY_STEPS
                        max number of learning rate decay steps (default: 4)
  --decay-factor DECAY_FACTOR
                        learning rate decay factor (default: 0.5)

validation setup:
  --val-batch-size VAL_BATCH_SIZE
                        batch size for validation (default: 64)
  --val-max-length VAL_MAX_LENGTH
                        maximum sequence length for validation (including
                        special BOS and EOS tokens) (default: 125)
  --val-min-length VAL_MIN_LENGTH
                        minimum sequence length for validation (including
                        special BOS and EOS tokens) (default: 0)
  --val-loader-workers VAL_LOADER_WORKERS
                        number of workers for validation data loading
                        (default: 0)

test setup:
  --test-batch-size TEST_BATCH_SIZE
                        batch size for test (default: 128)
  --test-max-length TEST_MAX_LENGTH
                        maximum sequence length for test (including special
                        BOS and EOS tokens) (default: 150)
  --test-min-length TEST_MIN_LENGTH
                        minimum sequence length for test (including special
                        BOS and EOS tokens) (default: 0)
  --beam-size BEAM_SIZE
                        beam size (default: 5)
  --len-norm-factor LEN_NORM_FACTOR
                        length normalization factor (default: 0.6)
  --cov-penalty-factor COV_PENALTY_FACTOR
                        coverage penalty factor (default: 0.1)
  --len-norm-const LEN_NORM_CONST
                        length normalization constant (default: 5.0)
  --intra-epoch-eval N  evaluate within training epoch, this option will
                        enable extra N equally spaced evaluations executed
                        during each training epoch (default: 0)
  --test-loader-workers TEST_LOADER_WORKERS
                        number of workers for test data loading (default: 0)

checkpointing setup:
  --start-epoch START_EPOCH
                        manually set initial epoch counter (default: 0)
  --resume PATH         resumes training from checkpoint from PATH (default:
                        None)
  --save-all            saves checkpoint after every epoch (default: False)
  --save-freq SAVE_FREQ
                        save checkpoint every SAVE_FREQ batches (default:
                        5000)
  --keep-checkpoints KEEP_CHECKPOINTS
                        keep only last KEEP_CHECKPOINTS checkpoints, affects
                        only checkpoints controlled by --save-freq option
                        (default: 0)

benchmark setup:
  --target-perf TARGET_PERF
                        target training performance (in tokens per second)
                        (default: None)
  --target-bleu TARGET_BLEU
                        target accuracy (default: None)

Inference

The complete list of available parameters for the translate.py inference script contains:

data setup:
  -o OUTPUT, --output OUTPUT
                        full path to the output file if not specified, then
                        the output will be printed (default: None)
  -r REFERENCE, --reference REFERENCE
                        full path to the file with reference translations (for
                        sacrebleu, raw text) (default: None)
  -m MODEL, --model MODEL
                        full path to the model checkpoint file (default: None)
  --synthetic           use synthetic dataset (default: False)
  --synthetic-batches SYNTHETIC_BATCHES
                        number of synthetic batches to generate (default: 64)
  --synthetic-vocab SYNTHETIC_VOCAB
                        size of synthetic vocabulary (default: 32320)
  --synthetic-len SYNTHETIC_LEN
                        sequence length of synthetic samples (default: 50)
  -i INPUT, --input INPUT
                        full path to the input file (raw text) (default: None)
  -t INPUT_TEXT [INPUT_TEXT ...], --input-text INPUT_TEXT [INPUT_TEXT ...]
                        raw input text (default: None)
  --sort                sorts dataset by sequence length (use '--no-sort' to
                        disable) (default: False)

inference setup:
  --batch-size BATCH_SIZE [BATCH_SIZE ...]
                        batch size per GPU (default: [128])
  --beam-size BEAM_SIZE [BEAM_SIZE ...]
                        beam size (default: [5])
  --max-seq-len MAX_SEQ_LEN
                        maximum generated sequence length (default: 80)
  --len-norm-factor LEN_NORM_FACTOR
                        length normalization factor (default: 0.6)
  --cov-penalty-factor COV_PENALTY_FACTOR
                        coverage penalty factor (default: 0.1)
  --len-norm-const LEN_NORM_CONST
                        length normalization constant (default: 5.0)

general setup:
  --math {fp16,fp32,tf32} [{fp16,fp32,tf32} ...]
                        precision (default: ['fp16'])
  --env                 print info about execution env (use '--no-env' to
                        disable) (default: False)
  --bleu                compares with reference translation and computes BLEU
                        (use '--no-bleu' to disable) (default: True)
  --cuda                enables cuda (use '--no-cuda' to disable) (default:
                        True)
  --cudnn               enables cudnn (use '--no-cudnn' to disable) (default:
                        True)
  --batch-first         uses (batch, seq, feature) data format for RNNs
                        (default: True)
  --seq-first           uses (seq, batch, feature) data format for RNNs
                        (default: True)
  --save-dir SAVE_DIR   path to directory with results, it will be
                        automatically created if it does not exist (default:
                        gnmt)
  --dllog-file DLLOG_FILE
                        Name of the DLLogger output file (default:
                        eval_log.json)
  --print-freq PRINT_FREQ, -p PRINT_FREQ
                        print log every PRINT_FREQ batches (default: 1)

benchmark setup:
  --target-perf TARGET_PERF
                        target inference performance (in tokens per second)
                        (default: None)
  --target-bleu TARGET_BLEU
                        target accuracy (default: None)
  --repeat REPEAT [REPEAT ...]
                        loops over the dataset REPEAT times, flag accepts
                        multiple arguments, one for each specified batch size
                        (default: [1])
  --warmup WARMUP       warmup iterations for performance counters (default:
                        0)
  --percentiles PERCENTILES [PERCENTILES ...]
                        Percentiles for confidence intervals for
                        throughput/latency benchmarks (default: (90, 95, 99))
  --tables              print accuracy, throughput and latency results in
                        tables (use '--no-tables' to disable) (default: False)

To see the full list of available options and their descriptions, use the -h or --help command line option. For example, for training:

python3 train.py --help

usage: train.py [-h] [--dataset-dir DATASET_DIR] [--src-lang SRC_LANG]
                [--tgt-lang TGT_LANG] [--vocab VOCAB] [-bpe BPE_CODES]
                [--train-src TRAIN_SRC] [--train-tgt TRAIN_TGT]
                [--val-src VAL_SRC] [--val-tgt VAL_TGT] [--test-src TEST_SRC]
                [--test-tgt TEST_TGT] [--save-dir SAVE_DIR]
                [--print-freq PRINT_FREQ] [--hidden-size HIDDEN_SIZE]
                [--num-layers NUM_LAYERS] [--dropout DROPOUT]
                [--share-embedding] [--smoothing SMOOTHING]
                [--math {fp16,fp32,tf32,manual_fp16}] [--seed SEED]
                [--prealloc-mode {off,once,always}] [--dllog-file DLLOG_FILE]
                [--eval] [--env] [--cuda] [--cudnn] [--log-all-ranks]
                [--train-max-size TRAIN_MAX_SIZE]
                [--train-batch-size TRAIN_BATCH_SIZE]
                [--train-global-batch-size TRAIN_GLOBAL_BATCH_SIZE]
                [--train-iter-size N] [--epochs EPOCHS]
                [--grad-clip GRAD_CLIP] [--train-max-length TRAIN_MAX_LENGTH]
                [--train-min-length TRAIN_MIN_LENGTH]
                [--train-loader-workers TRAIN_LOADER_WORKERS]
                [--batching {random,sharding,bucketing}]
                [--shard-size SHARD_SIZE] [--num-buckets NUM_BUCKETS]
                [--optimizer OPTIMIZER] [--lr LR]
                [--optimizer-extra OPTIMIZER_EXTRA] [--init-scale INIT_SCALE]
                [--upscale-interval UPSCALE_INTERVAL]
                [--warmup-steps WARMUP_STEPS] [--remain-steps REMAIN_STEPS]
                [--decay-interval DECAY_INTERVAL] [--decay-steps DECAY_STEPS]
                [--decay-factor DECAY_FACTOR]
                [--val-batch-size VAL_BATCH_SIZE]
                [--val-max-length VAL_MAX_LENGTH]
                [--val-min-length VAL_MIN_LENGTH]
                [--val-loader-workers VAL_LOADER_WORKERS]
                [--test-batch-size TEST_BATCH_SIZE]
                [--test-max-length TEST_MAX_LENGTH]
                [--test-min-length TEST_MIN_LENGTH] [--beam-size BEAM_SIZE]
                [--len-norm-factor LEN_NORM_FACTOR]
                [--cov-penalty-factor COV_PENALTY_FACTOR]
                [--len-norm-const LEN_NORM_CONST] [--intra-epoch-eval N]
                [--test-loader-workers TEST_LOADER_WORKERS]
                [--start-epoch START_EPOCH] [--resume PATH] [--save-all]
                [--save-freq SAVE_FREQ] [--keep-checkpoints KEEP_CHECKPOINTS]
                [--target-perf TARGET_PERF] [--target-bleu TARGET_BLEU]
                [--local_rank LOCAL_RANK]

For example, for inference:

python3 translate.py --help

usage: translate.py [-h] [-o OUTPUT] [-r REFERENCE] [-m MODEL] [--synthetic]
                    [--synthetic-batches SYNTHETIC_BATCHES]
                    [--synthetic-vocab SYNTHETIC_VOCAB]
                    [--synthetic-len SYNTHETIC_LEN]
                    [-i INPUT | -t INPUT_TEXT [INPUT_TEXT ...]] [--sort]
                    [--batch-size BATCH_SIZE [BATCH_SIZE ...]]
                    [--beam-size BEAM_SIZE [BEAM_SIZE ...]]
                    [--max-seq-len MAX_SEQ_LEN]
                    [--len-norm-factor LEN_NORM_FACTOR]
                    [--cov-penalty-factor COV_PENALTY_FACTOR]
                    [--len-norm-const LEN_NORM_CONST]
                    [--math {fp16,fp32,tf32} [{fp16,fp32,tf32} ...]] [--env]
                    [--bleu] [--cuda] [--cudnn] [--batch-first | --seq-first]
                    [--save-dir SAVE_DIR] [--dllog-file DLLOG_FILE]
                    [--print-freq PRINT_FREQ] [--target-perf TARGET_PERF]
                    [--target-bleu TARGET_BLEU] [--repeat REPEAT [REPEAT ...]]
                    [--warmup WARMUP]
                    [--percentiles PERCENTILES [PERCENTILES ...]] [--tables]
                    [--local_rank LOCAL_RANK]

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 automatically downloads and preprocesses the training, validation and test datasets. By default, data is 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 (Byte Pair Encodings - BPE). By default, the script builds the shared vocabulary of 32,000 tokens.

In order to test with other datasets, the script needs to be customized accordingly.

The process of downloading and preprocessing the data can be found in the scripts/wmt16_en_de.sh script.

Initially, data is downloaded from www.statmt.org. Then europarl-v7, commoncrawl and news-commentary corpora are concatenated to form the training dataset, similarly newstest2015 and newstest2016 are concatenated to form the validation dataset. Raw data is preprocessed with Moses, first by launching Moses tokenizer (tokenizer breaks up text into individual words), then by launching clean-corpus-n.perl which removes invalid sentences and does initial filtering by sequence length.

Second stage of preprocessing is done by launching the scripts/filter_dataset.py script, which discards all pairs of sentences that can't be decoded by latin-1 encoder.

Third state of preprocessing uses the subword-nmt package. First it builds shared byte pair encoding vocabulary with 32,000 merge operations (command subword-nmt learn-bpe), then it applies generated vocabulary to training, validation and test corpora (command subword-nmt apply-bpe).

The default training configuration can be launched by running the train.py 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 then performed after each training epoch. Results are stored in the gnmt directory.

The training script launches data-parallel training with batch size 128 per GPU on all available GPUs. We have tested reliance on up to 16 GPUs on a single node. 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 gnmt directory.

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

0: Summary: Epoch: 3	Training Loss: 3.1336	Validation Loss: 2.9587	Test BLEU: 23.18
0: Performance: Epoch: 3	Training: 418772 Tok/s	Validation: 1445331 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 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.

By default, the train.py script will launch mixed precision training with Tensor Cores. You can change this behavior by setting:

• the --math fp32 flag to launch single precision training (for NVIDIA Volta and NVIDIA Turing architectures) or
• the --math tf32 flag to launch TF32 training with Tensor Cores (for NVIDIA Ampere architecture)

for the train.py training script.

To view all available options for training, run python3 train.py --help.

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.

The performance measurements in this document were conducted at the time of publication and may not reflect the performance achieved from NVIDIA’s latest software release. For the most up-to-date performance measurements, go to NVIDIA Data Center Deep Learning Product Performance.

The following section shows how to run benchmarks measuring the model performance in training and inference modes.

Training is launched on batches of text data, different batches have different sequence lengths (number of tokens in the longest sequence). Sequence length and batch efficiency (ratio of non-pad tokens to total number of tokens) affect performance of the training, therefore it's recommended to run the training on a large chunk of training dataset to get a stable and reliable average training performance. Ideally at least one full epoch of training should be launched to get a good estimate of training performance.

The following commands will launch one epoch of training:

To launch mixed precision training on 1, 4 or 8 GPUs, run:

python3 -m torch.distributed.launch --nproc_per_node=<#GPUs> train.py --seed 2 --train-global-batch-size 1024 --epochs 1 --math fp16

To launch mixed precision training on 16 GPUs, run:

python3 -m torch.distributed.launch --nproc_per_node=16 train.py --seed 2 --train-global-batch-size 2048 --epochs 1 --math fp16

Change --math fp16 to --math fp32 to launch single precision training (for NVIDIA Volta and NVIDIA Turing architectures) or to --math tf32 to launch TF32 training with Tensor Cores (for NVIDIA Ampere architecture).

After the training is completed, the train.py script prints a summary to standard output. Performance results are printed in the following format:

(...)
0: Performance: Epoch: 0	Training: 418926 Tok/s	Validation: 1430828 Tok/s
(...)

Training: 418926 Tok/s represents training throughput averaged over an entire training epoch and summed over all GPUs participating in the training.

The inference performance and accuracy benchmarks require a checkpoint from a fully trained model.

Command to launch the inference accuracy benchmark on NVIDIA Volta or on NVIDIA Turing architectures:

python3 translate.py \
  --model gnmt/model_best.pth \
  --input data/wmt16_de_en/newstest2014.en \
  --reference data/wmt16_de_en/newstest2014.de \
  --output /tmp/output \
  --math fp16 fp32 \
  --batch-size 128 \
  --beam-size 1 2 5 \
  --tables

Command to launch the inference accuracy benchmark on NVIDIA Ampere architecture:

python3 translate.py \
  --model gnmt/model_best.pth \
  --input data/wmt16_de_en/newstest2014.en \
  --reference data/wmt16_de_en/newstest2014.de \
  --output /tmp/output \
  --math fp16 tf32 \
  --batch-size 128 \
  --beam-size 1 2 5 \
  --tables

Command to launch the inference throughput and latency benchmarks on NVIDIA Volta or NVIDIA Turing architectures:

python3 translate.py \
  --model gnmt/model_best.pth \
  --input data/wmt16_de_en/newstest2014.en \
  --reference data/wmt16_de_en/newstest2014.de \
  --output /tmp/output \
  --math fp16 fp32 \
  --batch-size 1 2 4 8 32 128 512 \
  --repeat 1 1 1 1 2 8 16 \
  --beam-size 1 2 5 \
  --warmup 5 \
  --tables

Command to launch the inference throughput and latency benchmarks on NVIDIA Ampere architecture:

python3 translate.py \
  --model gnmt/model_best.pth \
  --input data/wmt16_de_en/newstest2014.en \
  --reference data/wmt16_de_en/newstest2014.de \
  --output /tmp/output \
  --math fp16 tf32 \
  --batch-size 1 2 4 8 32 128 512 \
  --repeat 1 1 1 1 2 8 16 \
  --beam-size 1 2 5 \
  --warmup 5 \
  --tables

The following sections provide details on how we achieved our performance and accuracy in training and inference.

Our results were obtained by running the train.py script with the default batch size = 128 per GPU in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs.

Command to launch the training:

python3 -m torch.distributed.launch --nproc_per_node=<#GPUs> train.py --seed 2 --train-global-batch-size 1024 --math fp16

Change --math fp16 to --math tf32 to launch TF32 training with Tensor Cores.

To achieve these same results, follow the Quick Start Guide outlined above.

Our results were obtained by running the train.py script with the default batch size = 128 per GPU in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs.

Command to launch the training:

python3 -m torch.distributed.launch --nproc_per_node=<#GPUs> train.py --seed 2 --train-global-batch-size 1024 --math fp16

Change --math fp16 to --math fp32 to launch single precision training.

To achieve these same results, follow the Quick Start Guide outlined above.

Our results were obtained by running the train.py script with the default batch size = 128 per GPU in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H with 16x V100 32GB GPUs.

To launch mixed precision training on 16 GPUs, run:

python3 -m torch.distributed.launch --nproc_per_node=16 train.py --seed 2 --train-global-batch-size 2048 --math fp16

Change --math fp16 to --math fp32 to launch single precision training.

To achieve these same results, follow the Quick Start Guide outlined above.

The GNMT v2 model was trained for 6 epochs, starting from 32 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-20.06-py3 Docker container on NVIDIA DGX A100 with 8x A100 40GB GPUs. The following table summarizes the results of the stability test.

In the following table, the BLEU scores after each training epoch for different initial random seeds are displayed.

Our results were obtained by running the train.py training script in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Throughput performance numbers (in tokens per second) were averaged over an entire training epoch.

To achieve these same results, follow the Quick Start Guide outlined above.

Our results were obtained by running the train.py training script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Throughput performance numbers (in tokens per second) were averaged over an entire training epoch.

To achieve these same results, follow the Quick Start Guide outlined above.

Our results were obtained by running the train.py training script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H with 16x V100 32GB GPUs. Throughput performance numbers (in tokens per second) were averaged over an entire training epoch.

To achieve these same results, follow the Quick Start Guide outlined above.

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA A100 40GB GPU. Full command to launch the inference accuracy benchmark was provided in the Inference performance benchmark section.

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA Tesla V100 16GB GPU. Full command to launch the inference accuracy benchmark was provided in the Inference performance benchmark section.

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA Tesla T4. Full command to launch the inference accuracy benchmark was provided in the Inference performance benchmark section.

To achieve these same results, follow the Quick Start Guide outlined above.

Tables presented in this section show the average inference throughput (columns Avg (tok/s)) and inference throughput for various confidence intervals (columns N% (ms), where N denotes the confidence interval). Inference throughput is measured in tokens per second. Speedups reported in FP16 subsections are relative to FP32 (for NVIDIA Volta and NVIDIA Turing) and relative to TF32 (for NVIDIA Ampere) numbers for corresponding configuration.

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA A100 40GB. Full command to launch the inference throughput benchmark was provided in the Inference performance benchmark section.

FP16

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA T4. Full command to launch the inference throughput benchmark was provided in the Inference performance benchmark section.

FP16

To achieve these same results, follow the Quick Start Guide outlined above.

Tables presented in this section show the average inference latency (columns Avg (ms)) and inference latency for various confidence intervals (columns N% (ms), where N denotes the confidence interval). Inference latency is measured in milliseconds. Speedups reported in FP16 subsections are relative to FP32 (for NVIDIA Volta and NVIDIA Turing) and relative to TF32 (for NVIDIA Ampere) numbers for corresponding configuration.

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA A100 40GB. Full command to launch the inference latency benchmark was provided in the Inference performance benchmark section.

FP16

Our results were obtained by running the translate.py script in the pytorch-20.06-py3 NGC Docker container with NVIDIA T4. Full command to launch the inference latency benchmark was provided in the Inference performance benchmark section.

FP16

To achieve these same results, follow the Quick Start Guide outlined above.

• July 2020
  • Added support for NVIDIA DGX A100
  • Default container updated to NGC PyTorch 20.06-py3
• June 2019
  • Default container updated to NGC PyTorch 19.05-py3
  • Mixed precision training implemented using APEX AMP
  • Added inference throughput and latency results on NVIDIA T4 and NVIDIA Tesla V100 16GB
  • Added option to run inference on user-provided raw input text from command line
• February 2019
  • Different batching algorithm (bucketing with 5 equal-width buckets)
  • Additional dropouts before first LSTM layer in encoder and in decoder
  • Weight initialization changed to uniform (-0.1,0.1)
  • Switched order of dropout and concatenation with attention in decoder
  • Default container updated to NGC PyTorch 19.01-py3
• December 2018
  • Added exponential warm-up and step learning rate decay
  • Multi-GPU (distributed) inference and validation
  • Default container updated to NGC PyTorch 18.11-py3
  • General performance improvements
• August 2018
  • Initial release

There are no known issues in this release.