Transformer-based pre-trained language models can achieve superior performance on most NLP tasks due to large parameter capacity, but also lead to huge computation cost. Fortunately, we find by empirical study that, most inputs only activate a tiny ratio of neurons during inference. Hence, we explore to accelerate large-model inference by conditional computation based on the sparse activation phenomenon. We propose to transform a large model into its mixture-of-experts (MoE) version with equal model size, namely MoEfication. Model MoEfication consists of two steps: (1) splitting the parameters of feed-forward neural networks (FFNs) into multiple parts as experts, and (2) building expert routers to decide which experts will be used for each input. To further improve the performance of MoEfied models, we can also fine-tune the models on downstream tasks, namely parameter calibration. Experimental results show that the MoEfied models can significantly reduce computation cost, e.g., only activating 20% FFN parameters of a 700-million-parameter model without performance degradation on several downstream tasks including text classification and reading comprehension.
相關內容
In the past years, significant improvements in the field of neural architecture search(NAS) have been made. However, it is still challenging to search for efficient networks due to the gap between the searched constraint and real inference time exists. To search for a high-performance network with low inference time, several previous works set a computational complexity constraint for the search algorithm. However, many factors affect the speed of inference(e.g., FLOPs, MACs). The correlation between a single indicator and the latency is not strong. Currently, some re-parameterization(Rep) techniques are proposed to convert multi-branch to single-path architecture which is inference-friendly. Nevertheless, multi-branch architectures are still human-defined and inefficient. In this work, we propose a new search space that is suitable for structural re-parameterization techniques. RepNAS, a one-stage NAS approach, is present to efficiently search the optimal diverse branch block(ODBB) for each layer under the branch number constraint. Our experimental results show the searched ODBB can easily surpass the manual diverse branch block(DBB) with efficient training.
This paper explores a better codebook for BERT pre-training of vision transformers. The recent work BEiT successfully transfers BERT pre-training from NLP to the vision field. It directly adopts one simple discrete VAE as the visual tokenizer, but has not considered the semantic level of the resulting visual tokens. By contrast, the discrete tokens in NLP field are naturally highly semantic. This difference motivates us to learn a perceptual codebook. And we surprisingly find one simple yet effective idea: enforcing perceptual similarity during the dVAE training. We demonstrate that the visual tokens generated by the proposed perceptual codebook do exhibit better semantic meanings, and subsequently help pre-training achieve superior transfer performance in various downstream tasks. For example, we achieve 84.5 Top-1 accuracy on ImageNet-1K with ViT-B backbone, outperforming the competitive method BEiT by +1.3 with the same pre-training epochs. It can also improve the performance of object detection and segmentation tasks on COCO val by +1.3 box AP and +1.0 mask AP, semantic segmentation on ADE20k by +1.0 mIoU, The code and models will be available at \url{//github.com/microsoft/PeCo}.
This paper presents an efficient multi-scale vision Transformer, called ResT, that capably served as a general-purpose backbone for image recognition. Unlike existing Transformer methods, which employ standard Transformer blocks to tackle raw images with a fixed resolution, our ResT have several advantages: (1) A memory-efficient multi-head self-attention is built, which compresses the memory by a simple depth-wise convolution, and projects the interaction across the attention-heads dimension while keeping the diversity ability of multi-heads; (2) Position encoding is constructed as spatial attention, which is more flexible and can tackle with input images of arbitrary size without interpolation or fine-tune; (3) Instead of the straightforward tokenization at the beginning of each stage, we design the patch embedding as a stack of overlapping convolution operation with stride on the 2D-reshaped token map. We comprehensively validate ResT on image classification and downstream tasks. Experimental results show that the proposed ResT can outperform the recently state-of-the-art backbones by a large margin, demonstrating the potential of ResT as strong backbones. The code and models will be made publicly available at //github.com/wofmanaf/ResT.
Deep Learning has revolutionized the fields of computer vision, natural language understanding, speech recognition, information retrieval and more. However, with the progressive improvements in deep learning models, their number of parameters, latency, resources required to train, etc. have all have increased significantly. Consequently, it has become important to pay attention to these footprint metrics of a model as well, not just its quality. We present and motivate the problem of efficiency in deep learning, followed by a thorough survey of the five core areas of model efficiency (spanning modeling techniques, infrastructure, and hardware) and the seminal work there. We also present an experiment-based guide along with code, for practitioners to optimize their model training and deployment. We believe this is the first comprehensive survey in the efficient deep learning space that covers the landscape of model efficiency from modeling techniques to hardware support. Our hope is that this survey would provide the reader with the mental model and the necessary understanding of the field to apply generic efficiency techniques to immediately get significant improvements, and also equip them with ideas for further research and experimentation to achieve additional gains.
Recent studies on mobile network design have demonstrated the remarkable effectiveness of channel attention (e.g., the Squeeze-and-Excitation attention) for lifting model performance, but they generally neglect the positional information, which is important for generating spatially selective attention maps. In this paper, we propose a novel attention mechanism for mobile networks by embedding positional information into channel attention, which we call "coordinate attention". Unlike channel attention that transforms a feature tensor to a single feature vector via 2D global pooling, the coordinate attention factorizes channel attention into two 1D feature encoding processes that aggregate features along the two spatial directions, respectively. In this way, long-range dependencies can be captured along one spatial direction and meanwhile precise positional information can be preserved along the other spatial direction. The resulting feature maps are then encoded separately into a pair of direction-aware and position-sensitive attention maps that can be complementarily applied to the input feature map to augment the representations of the objects of interest. Our coordinate attention is simple and can be flexibly plugged into classic mobile networks, such as MobileNetV2, MobileNeXt, and EfficientNet with nearly no computational overhead. Extensive experiments demonstrate that our coordinate attention is not only beneficial to ImageNet classification but more interestingly, behaves better in down-stream tasks, such as object detection and semantic segmentation. Code is available at //github.com/Andrew-Qibin/CoordAttention.
Many real-world applications require the prediction of long sequence time-series, such as electricity consumption planning. Long sequence time-series forecasting (LSTF) demands a high prediction capacity of the model, which is the ability to capture precise long-range dependency coupling between output and input efficiently. Recent studies have shown the potential of Transformer to increase the prediction capacity. However, there are several severe issues with Transformer that prevent it from being directly applicable to LSTF, such as quadratic time complexity, high memory usage, and inherent limitation of the encoder-decoder architecture. To address these issues, we design an efficient transformer-based model for LSTF, named Informer, with three distinctive characteristics: (i) a $ProbSparse$ Self-attention mechanism, which achieves $O(L \log L)$ in time complexity and memory usage, and has comparable performance on sequences' dependency alignment. (ii) the self-attention distilling highlights dominating attention by halving cascading layer input, and efficiently handles extreme long input sequences. (iii) the generative style decoder, while conceptually simple, predicts the long time-series sequences at one forward operation rather than a step-by-step way, which drastically improves the inference speed of long-sequence predictions. Extensive experiments on four large-scale datasets demonstrate that Informer significantly outperforms existing methods and provides a new solution to the LSTF problem.
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
Language model based pre-trained models such as BERT have provided significant gains across different NLP tasks. In this paper, we study different types of pre-trained transformer based models such as auto-regressive models (GPT-2), auto-encoder models (BERT), and seq2seq models (BART) for conditional data augmentation. We show that prepending the class labels to text sequences provides a simple yet effective way to condition the pre-trained models for data augmentation. On three classification benchmarks, pre-trained Seq2Seq model outperforms other models. Further, we explore how different pre-trained model based data augmentation differs in-terms of data diversity, and how well such methods preserve the class-label information.
Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet. To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.4% top-1 / 97.1% top-5 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters. Source code is at //github.com/tensorflow/tpu/tree/master/models/official/efficientnet.
This work focuses on combining nonparametric topic models with Auto-Encoding Variational Bayes (AEVB). Specifically, we first propose iTM-VAE, where the topics are treated as trainable parameters and the document-specific topic proportions are obtained by a stick-breaking construction. The inference of iTM-VAE is modeled by neural networks such that it can be computed in a simple feed-forward manner. We also describe how to introduce a hyper-prior into iTM-VAE so as to model the uncertainty of the prior parameter. Actually, the hyper-prior technique is quite general and we show that it can be applied to other AEVB based models to alleviate the {\it collapse-to-prior} problem elegantly. Moreover, we also propose HiTM-VAE, where the document-specific topic distributions are generated in a hierarchical manner. HiTM-VAE is even more flexible and can generate topic distributions with better variability. Experimental results on 20News and Reuters RCV1-V2 datasets show that the proposed models outperform the state-of-the-art baselines significantly. The advantages of the hyper-prior technique and the hierarchical model construction are also confirmed by experiments.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191