The Temporal Fusion Transformer TFT model is a state-of-the-art architecture for interpretable, multi-horizon time-series prediction. The model was first developed and implemented by Google with the collaboration with the University of Oxford. This implementation differs from the reference implementation by addressing the issue of missing data, which is common in production datasets, by either masking their values in attention matrices or embedding them as a special value in the latent space. This model enables the prediction of confidence intervals for future values of time series for multiple future timesteps.

This model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results 1.45x faster than training without Tensor Cores while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.

Model architecture

The TFT model is a hybrid architecture joining LSTM encoding of time series and interpretability of transformer attention layers. Prediction is based on three types of variables: static (constant for a given time series), known (known in advance for whole history and future), observed (known only for historical data). All these variables come in two flavors: categorical, and continuous. In addition to historical data, we feed the model with historical values of time series. All variables are embedded in high-dimensional space by learning an embedding vector. Categorical variables embeddings are learned in the classical sense of embedding discrete values. The model learns a single vector for each continuous variable, which is then scaled by this variable's value for further processing. The next step is to filter variables through the Variable Selection Network (VSN), which assigns weights to the inputs in accordance with their relevance to the prediction. Static variables are used as a context for variable selection of other variables and as an initial state of LSTM encoders. After encoding, variables are passed to multi-head attention layers (decoder), which produce the final prediction. Whole architecture is interwoven with residual connections with gating mechanisms that allow the architecture to adapt to various problems by skipping some parts of it. For the sake of explainability, heads of self-attention layers share value matrices. This allows interpreting self-attention as an ensemble of models predicting different temporal patterns over the same feature set. The other feature that helps us understand the model is VSN activations, which tells us how relevant the given feature is to the prediction. image source: https://arxiv.org/abs/1912.09363

Default configuration

The specific configuration of the TFT model depends on the dataset used. Not only is the volume of the model subject to change but so are the data sampling and preprocessing strategies. During preprocessing, data is normalized per feature. For a part of the datasets, we apply scaling per-time-series, which takes into account shifts in distribution between entities (i.e., a factory consumes more electricity than an average house). The model is trained with the quantile loss: For quantiles in [0.1, 0.5, 0.9]. The default configurations are tuned for distributed training on DGX-1-32G with mixed precision. We use dynamic loss scaling. Specific values are provided in the table below.

Feature support matrix

The following features are supported by this model:

Features

Automatic Mixed Precision provides an easy way to leverage Tensor Cores' performance. It allows the execution of parts of a network in lower precision. Refer to Mixed precision training for more information.

PyTorch DistributedDataParallel - a module wrapper that enables easy multiprocess distributed data-parallel training.

Mixed precision training

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

Enabling mixed precision

Mixed precision is enabled in PyTorch by using the Automatic Mixed Precision torch.cuda.amp module, which 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 applied automatically by the GradScaler class. All the necessary steps to implement AMP are verbosely described here.

To enable mixed precision for TFT, simply add the --use_amp option to the training script.

Enabling TF32

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.

Glossary

Multi horizon prediction
Process of estimating values of a time series for multiple future time steps.

Quantiles
Cut points dividing the range of a probability distribution intervals with equal probabilities.

Time series
Series of data points indexed and equally spaced in time.

Transformer
The paper Attention Is All You Need introduces a novel architecture called Transformer that uses an attention mechanism and transforms one sequence into another.