Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
Abstract
Deep learning thrives with large neural networks and large datasets. However, larger networks and larger datasets result in longer training times that impede research and development progress. Distributed synchronous SGD offers a potential solution to this problem by dividing SGD minibatches over a pool of parallel workers. Yet to make this scheme efficient, the perworker workload must be large, which implies nontrivial growth in the SGD minibatch size. In this paper, we empirically show that on the ImageNet dataset large minibatches cause optimization dif ficulties, but when these are addressed the trained networks exhibit good generalization. Specifically, we show no loss of accuracy when training with large minibatch sizes up to 8192 images. To achieve this result, we adopt a linear scaling rule for adjusting learning rates as a function of minibatch size and develop a new warmup scheme that overcomes optimization challenges early in training. With these simple techniques, our Caffe2based system trains ResNet 50 with a minibatch size of 8192 on 256 GPUs in one hour, while matching small minibatch accuracy. Using commodity hardware, our implementation achieves ∼90% scaling efficiency when moving from 8 to 256 GPUs. This system enables us to train visual recognition models on internetscale data with high efficiency.
Notes
This paper is focused on parallelizing model training across multiple GPUs. This is a problem as it is currently typically quite difficult to get reasonable speedups when using multiple GPUs to train a model (by difficult, I mean that you get significantly less than linear speedups).
In this paper, by Facebook Research, the authors are able to get almost linear speedups moving from 8 to 256 GPUs (0.9x), which is quite good. Using 256 GPUs, the authors are able to train the ResNet50 model in 1 hour (to give you an idea of how impressive this is, the ImageNet dataset consists of 750 GB of data). AlexNet, which was the breakthrough paper that was a large contributor to Deep Learning's current popularity, took 56 days to train for 90 epochs on two NVIDIA GTX 580s, which is equivalent to 288 GPU hours 1.
The trick that they use to do this is to use a large minibatch size consisting of 8192 images. Using a large minibatch size makes it much easier to fully exploit the GPUs, and hence makes training run faster, however, it makes each gradient update noisier, making the training potentially take longer to converge (or making it so that training might converge to a wrong, i.e. nonoptimal, answer). Additionally, if you are using multiple GPUs, you have to synchronize the weights after each minibatch update, so having smaller minibatches causes the required communication overhead to drastically increase.
To compensate for the noise introduced by the large minibatch size used here, the authors used a "Linear Scaling Rule", where they multiplied the learning rate by \(k\) when they used a minibatch size of \(k\). This allowed the authors to match the accuracy between small and large minibatches.
The authors note that the linear scaling rule is nice theoretically as, after \(k\) iterations of SGD with a learning rate of \(\eta\) and a minibatch of \(n\), the weight vector is
\[w_{t+k} = w_t  \eta \frac{1}{n} \sum \limits_{j < k} \sum \limits_{x \in \mathcal{B}_j} \nabla l(x, w_{t + j})\]When, instead, we take asingle step with a minibatch \(\cup_j \mathcal{B}_j\) of size \(kn\) and learning rate \(\eta'\), the updated weight vector is instead
\[w_{t+1}' = w_t  \eta' \frac{1}{n} \sum \limits_{j < k} \sum \limits_{x \in \mathcal{B}_j} \nabla l(x, w_t),\]which is different. However, if \(\Delta l(x, w_t)\) is close in value to \(\Delta l(x, w_{t+j})\) for \(j < k\), then setting \(\eta' = kn\) makes it so that \(w_{t+1} \approx w_{t+k}\) (I would imagine that you could formalize this with an epsilondelta proof fairly easily). The paper notes that the two updates can only be similar when the linear scaling rule is used; in other words, the linear scaling rule is necessary, but not sufficient.
The authors note that the assumption that the two gradients are similar doesn't hold during the initial training, when the weights are rapidly changing, and that the results hold only for a large, but finite, range of minibatch sizes (which for ImageNet is as large as 8192). The authors use a "warmup" phase to mitigate the problems with divergence during initial training, where the model uses less aggressive learning rates, and then switches to the linear scaling rule after 5 epochs. That didn't work, and instead, they used a gradual warmup that brought the learning rate to increase at a constant rate per iteration so that it reached \(\eta' = k \eta\) after 5 epochs, which worked better.
The authors then go on to discuss results, namely that they were able to train ResNet50 in one hour while still getting state of the art results.
What's novel about this is the size of the parallelization; presumably there's nothing special about using 256 GPUs, and if someone had the resources available (cough Google cough), one could scale this further. Given that GPUs seem to be following Moore's law and doubling in performance every 18 months, this paper seems to be important; if we can train a state of the art model in one hour using 256 GPUs today, then within 3 years, we could train one in 15 minutes. If someone wanted to scale the number of GPUs higher, they could train the model in under 10 minutes.
Conversely, researchers currently tolerate several weeks of training time (Mozilla's implementation of Baidu's DeepSpeech model takes Mozilla roughly 10 days to train on 4 high end GPUs 2); if the amount of model that can be trained in that time drastically increases, all of a sudden researchers are able to consider datasets that are radically larger, and can start approaching even higher performance levels.
Conclusions
Strong paper, effectively lays out how training deep networks can be scaled effectively. This sort of yeoman's work is needed in the field.
Concerns:

Continues the trend of papers that rely on resources only available at a handful of industrial labs. No academic researcher that's not affiliated with a large tech company would be able to muster the 256 GPUs required to reproduce this work.

The amount of proprietary code required for this is a bit insane; you have to have an infrastructure that can support the communication between GPUs required here. Similar to my first point. Reproducibility suffers.