Large Scale Language Modeling: Converging on 40GB of Text in Four Hours

By Raul Puri, Robert Kirby, Nikolai Yakovenko, Bryan Catanzaro. 2018

31 citations recorded by the source corpus.

Recent work has shown how to train Convolutional Neural Networks (CNNs) rapidly on large image datasets [1], then transfer the knowledge gained from these models to a variety of tasks [2]. Following [3], in this work, we demonstrate similar scalability and transfer for Recurrent Neural Networks (RNNs) for Natural Language tasks. By utilizing mixed precision arithmetic and a 32k batch size distributed across 128 NVIDIA Tesla V100 GPUs, we are able to train a character-level 4096-dimension multiplicative LSTM (mLSTM) [4] for unsupervised text reconstruction over 3 epochs of the 40 GB Amazon Reviews dataset [5] in four hours. This runtime compares favorably with previous work taking one month to train the same size and configuration for one epoch over the same dataset [3]. Converging large batch RNN models can be challenging. Recent work has suggested scaling the learning rate as a function of batch size, but we find that simply scaling the learning rate as a function of batch size leads either to significantly worse convergence or immediate divergence for this problem. We provide a learning rate schedule that allows our model to converge with a 32k batch size. Since our model converges over the Amazon Reviews dataset in hours, and our compute requirement of 128 Tesla V100 GPUs, while substantial, is commercially available, this work opens up large scale unsupervised NLP training to most commercial applications and deep learning researchers 11Our code is publicly available: https://github.com/NVIDIA/sentiment-discovery, A model can be trained over most public or private text datasets overnight.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Large Language Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Large Language Models

Citations in the corpus

An Empirical Model of Large-Batch Training (2018, methodology): “…sizes of 8 thousand [GDG + 17], 16 thousand [SKYL17], 32 thousand [YGG17, YZH + 17, ASF17], and even 64 thousand [JSH + 18] examples have been effectively employed to train ImageNet, and batch sizes of thousands have been effective for language models and generative models [OEGA18, PKYC18, BDS18].”
Scaling Neural Machine Translation (2018, background): “Recent work by Puri et al. (2018) considers large-scale distributed training of language models (LM) achieving 109x scaling with 128 GPUs.”
Beyond human-level accuracy: computational challenges in deep learning (2019, methodology): “Researchers have shown how to train image and language applications quickly using data-parallel scaling to reduce per-accelerator activation memory requirements [14, 26, 40].”
Demystifying Parallel and Distributed Deep Learning (2018, methodology): “The use of large-batch methods and reduced floating-point precision has sparked a “race” toward accelerating the training process of specific model/dataset combinations in image processing [84, 123, 176, 191, 256] and language modeling [192, 203, 257].”
On the Computational Inefficiency of Large Batch Sizes for Stochastic Gradient Descent (2018, methodology): “Although [24] showed that training on the Amazon Reviews dataset [22] can be done within 4 hours, they tune hyper-parameters heavily.”
Compressing Gradient Optimizers via Count-Sketches (2019, background): “For example, (Puri et al., 2018) showed that doubling the dimensionality of a multiplicative LSTM (Krause et al.”
PruneTrain: Gradual Structured Pruning from Scratch for Faster Neural Network Training (2019, result): “ rate. For other deep learning applications that have a different relation, an appropriate learning rate adjustment rule can be adopted instead (e.g., the square root scaling rule for language models [42]). We evaluate dynamic mini-batch adjustment by training ResNet50 on both CIFAR and ImageNet dat”
A Unified Architecture for Accelerating Distributed DNN Training in Heterogeneous GPU/CPU Clusters (2020, methodology): “All-reduce has been optimized for many years, and most state-of-the-art training speed records are achieved using all-reduce, including classical CNN-based ImageNet tasks [27, 36, 49, 73], RNN-based language modeling tasks [56], and the pre-training of BERT [26, 74].”