Large-scale pre-trained sequence-to-sequence models like BART and T5 achieve state-of-the-art performance on many generative NLP tasks. However, such models pose a great challenge in resource-constrained scenarios owing to their large memory requirements and high latency. To alleviate this issue, we propose to jointly distill and quantize the model, where knowledge is transferred from the full-precision teacher model to the quantized and distilled low-precision student model. Empirical analyses show that, despite the challenging nature of generative tasks, we were able to achieve a 16.5x model footprint compression ratio with little performance drop relative to the full-precision counterparts on multiple summarization and QA datasets. We further pushed the limit of compression ratio to 27.7x and presented the performance-efficiency trade-off for generative tasks using pre-trained models. To the best of our knowledge, this is the first work aiming to effectively distill and quantize sequence-to-sequence pre-trained models for language generation tasks.
The recent focus on the efficiency of deep neural networks (DNNs) has led to significant work on model compression approaches, of which weight pruning is one of the most popular. At the same time, there is rapidly-growing computational support for efficiently executing the unstructured-sparse models obtained via pruning. Yet, most existing pruning methods minimize just the number of remaining weights, i.e. the size of the model, rather than optimizing for inference time. We address this gap by introducing SPDY, a new compression method which automatically determines layer-wise sparsity targets achieving a desired inference speedup on a given system, while minimizing accuracy loss. SPDY is composed of two new techniques: the first is an efficient dynamic programming algorithm for solving the speedup-constrained layer-wise compression problem assuming a set of given layer-wise sensitivity scores; the second is a local search procedure for determining accurate layer-wise sensitivity scores. Experiments across popular vision and language models show that SPDY guarantees speedups while recovering higher accuracy relative to existing strategies, both for one-shot and gradual pruning scenarios, and is compatible with most existing pruning approaches. We also extend our approach to the recently-proposed task of pruning with very little data, where we achieve the best known accuracy recovery when pruning to the GPU-supported 2:4 sparsity pattern.
Neural networks are more expressive when they have multiple layers. In turn, conventional training methods are only successful if the depth does not lead to numerical issues such as exploding or vanishing gradients, which occur less frequently when the layers are sufficiently wide. However, increasing width to attain greater depth entails the use of heavier computational resources and leads to overparameterized models. These subsequent issues have been partially addressed by model compression methods such as quantization and pruning, some of which relying on normalization-based regularization of the loss function to make the effect of most parameters negligible. In this work, we propose instead to use regularization for preventing neurons from dying or becoming linear, a technique which we denote as jumpstart regularization. In comparison to conventional training, we obtain neural networks that are thinner, deeper, and - most importantly - more parameter-efficient.
In modern neural networks like Transformers, linear layers require significant memory to store activations during backward pass. This study proposes a memory reduction approach to perform backpropagation through linear layers. Since the gradients of linear layers are computed by matrix multiplications, we consider methods for randomized matrix multiplications and demonstrate that they require less memory with a moderate decrease of the test accuracy. Also, we investigate the variance of the gradient estimate induced by the randomized matrix multiplication. We compare this variance with the variance coming from gradient estimation based on the batch of samples. We demonstrate the benefits of the proposed method on the fine-tuning of the pre-trained RoBERTa model on GLUE tasks.
Model compression methods can reduce model complexity on the premise of maintaining acceptable performance, and thus promote the application of deep neural networks under resource constrained environments. Despite their great success, the selection of suitable compression methods and design of details of the compression scheme are difficult, requiring lots of domain knowledge as support, which is not friendly to non-expert users. To make more users easily access to the model compression scheme that best meet their needs, in this paper, we propose AutoMC, an effective automatic tool for model compression. AutoMC builds the domain knowledge on model compression to deeply understand the characteristics and advantages of each compression method under different settings. In addition, it presents a progressive search strategy to efficiently explore pareto optimal compression scheme according to the learned prior knowledge combined with the historical evaluation information. Extensive experimental results show that AutoMC can provide satisfying compression schemes within short time, demonstrating the effectiveness of AutoMC.
In recent years, neural signed distance function (SDF) has become one of the most effective representation methods for 3D models. By learning continuous SDFs in 3D space, neural networks can predict the distance from a given query space point to its closest object surface,whose positive and negative signs denote inside and outside of the object, respectively. Training a specific network for each 3D model, which individually embeds its shape, can realize compressed representation of objects by storing fewer network (and possibly latent) parameters. Consequently, reconstruction through network inference and surface recovery can be achieved. In this paper, we propose an SDF prediction network using explicit key spheres as input. Key spheres are extracted from the internal space of objects, whose centers either have relatively larger SDF values (sphere radii), or are located at essential positions. By inputting the spatial information of multiple spheres which imply different local shapes, the proposed method can significantly improve the reconstruction accuracy with a negligible storage cost. Compared to previous works, our method achieves the high-fidelity and high-compression 3D object coding and reconstruction. Experiments conducted on three datasets verify the superior performance of our method.
As the training of giant dense models hits the boundary on the availability and capability of the hardware resources today, Mixture-of-Experts (MoE) models become one of the most promising model architectures due to their significant training cost reduction compared to a quality-equivalent dense model. Its training cost saving is demonstrated from encoder-decoder models (prior works) to a 5x saving for auto-aggressive language models (this work along with parallel explorations). However, due to the much larger model size and unique architecture, how to provide fast MoE model inference remains challenging and unsolved, limiting its practical usage. To tackle this, we present DeepSpeed-MoE, an end-to-end MoE training and inference solution as part of the DeepSpeed library, including novel MoE architecture designs and model compression techniques that reduce MoE model size by up to 3.7x, and a highly optimized inference system that provides 7.3x better latency and cost compared to existing MoE inference solutions. DeepSpeed-MoE offers an unprecedented scale and efficiency to serve massive MoE models with up to 4.5x faster and 9x cheaper inference compared to quality-equivalent dense models. We hope our innovations and systems help open a promising path to new directions in the large model landscape, a shift from dense to sparse MoE models, where training and deploying higher-quality models with fewer resources becomes more widely possible.
With the continuous development of neural networks for computer vision tasks, more and more network architectures have achieved outstanding success. As one of the most advanced neural network architectures, DenseNet shortcuts all feature maps to solve the model depth problem. Although this network architecture has excellent accuracy with low parameters, it requires an excessive inference time. To solve this problem, HarDNet reduces the connections between the feature maps, making the remaining connections resemble harmonic waves. However, this compression method may result in a decrease in the model accuracy and an increase in the parameters and model size. This network architecture may reduce the memory access time, but its overall performance can still be improved. Therefore, we propose a new network architecture, ThreshNet, using a threshold mechanism to further optimize the connection method. Different numbers of connections for different convolution layers are discarded to accelerate the inference of the network. The proposed network has been evaluated with image classification using CIFAR 10 and SVHN datasets under platforms of NVIDIA RTX 3050 and Raspberry Pi 4. The experimental results show that, compared with HarDNet68, GhostNet, MobileNetV2, ShuffleNet, and EfficientNet, the inference time of the proposed ThreshNet79 is 5%, 9%, 10%, 18%, and 20% faster, respectively. The number of parameters of ThreshNet95 is 55% less than that of HarDNet85. The new model compression and model acceleration methods can speed up the inference time, enabling network models to operate on mobile devices.
Quantized neural networks typically require smaller memory footprints and lower computation complexity, which is crucial for efficient deployment. However, quantization inevitably leads to a distribution divergence from the original network, which generally degrades the performance. To tackle this issue, massive efforts have been made, but most existing approaches lack statistical considerations and depend on several manual configurations. In this paper, we present an adaptive-mapping quantization method to learn an optimal latent sub-distribution that is inherent within models and smoothly approximated with a concrete Gaussian Mixture (GM). In particular, the network weights are projected in compliance with the GM-approximated sub-distribution. This sub-distribution evolves along with the weight update in a co-tuning schema guided by the direct task-objective optimization. Sufficient experiments on image classification and object detection over various modern architectures demonstrate the effectiveness, generalization property, and transferability of the proposed method. Besides, an efficient deployment flow for the mobile CPU is developed, achieving up to 7.46$\times$ inference acceleration on an octa-core ARM CPU. Our codes have been publicly released at \url{https://github.com/RunpeiDong/DGMS}.
Recently, Vision Transformer (ViT) has continuously established new milestones in the computer vision field, while the high computation and memory cost makes its propagation in industrial production difficult. Pruning, a traditional model compression paradigm for hardware efficiency, has been widely applied in various DNN structures. Nevertheless, it stays ambiguous on how to perform exclusive pruning on the ViT structure. Considering three key points: the structural characteristics, the internal data pattern of ViTs, and the related edge device deployment, we leverage the input token sparsity and propose a computation-aware soft pruning framework, which can be set up on vanilla Transformers of both flatten and CNN-type structures, such as Pooling-based ViT (PiT). More concretely, we design a dynamic attention-based multi-head token selector, which is a lightweight module for adaptive instance-wise token selection. We further introduce a soft pruning technique, which integrates the less informative tokens generated by the selector module into a package token that will participate in subsequent calculations rather than being completely discarded. Our framework is bound to the trade-off between accuracy and computation constraints of specific edge devices through our proposed computation-aware training strategy. Experimental results show that our framework significantly reduces the computation cost of ViTs while maintaining comparable performance on image classification. Moreover, our framework can guarantee the identified model to meet resource specifications of mobile devices and FPGA, and even achieve the real-time execution of DeiT-T on mobile platforms. For example, our method reduces the latency of DeiT-T to 26 ms (26%$\sim $41% superior to existing works) on the mobile device with 0.25%$\sim $4% higher top-1 accuracy on ImageNet.
Previous knowledge distillation (KD) methods mostly focus on compressing network architectures, which is not thorough enough in deployment as some costs like transmission bandwidth and imaging equipment are related to the image size. Therefore, we propose Pixel Distillation that extends knowledge distillation into the input level while simultaneously breaking architecture constraints. Such a scheme can achieve flexible cost control for deployment, as it allows the system to adjust both network architecture and image quality according to the overall requirement of resources. Specifically, we first propose an input spatial representation distillation (ISRD) mechanism to transfer spatial knowledge from large images to student's input module, which can facilitate stable knowledge transfer between CNN and ViT. Then, a Teacher-Assistant-Student (TAS) framework is further established to disentangle pixel distillation into the model compression stage and input compression stage, which significantly reduces the overall complexity of pixel distillation and the difficulty of distilling intermediate knowledge. Finally, we adapt pixel distillation to object detection via an aligned feature for preservation (AFP) strategy for TAS, which aligns output dimensions of detectors at each stage by manipulating features and anchors of the assistant. Comprehensive experiments on image classification and object detection demonstrate the effectiveness of our method. Code is available at https://github.com/gyguo/PixelDistillation.
Pre-trained Language Models (PLMs) have achieved great success in various Natural Language Processing (NLP) tasks under the pre-training and fine-tuning paradigm. With large quantities of parameters, PLMs are computation-intensive and resource-hungry. Hence, model pruning has been introduced to compress large-scale PLMs. However, most prior approaches only consider task-specific knowledge towards downstream tasks, but ignore the essential task-agnostic knowledge during pruning, which may cause catastrophic forgetting problem and lead to poor generalization ability. To maintain both task-agnostic and task-specific knowledge in our pruned model, we propose ContrAstive Pruning (CAP) under the paradigm of pre-training and fine-tuning. It is designed as a general framework, compatible with both structured and unstructured pruning. Unified in contrastive learning, CAP enables the pruned model to learn from the pre-trained model for task-agnostic knowledge, and fine-tuned model for task-specific knowledge. Besides, to better retain the performance of the pruned model, the snapshots (i.e., the intermediate models at each pruning iteration) also serve as effective supervisions for pruning. Our extensive experiments show that adopting CAP consistently yields significant improvements, especially in extremely high sparsity scenarios. With only 3% model parameters reserved (i.e., 97% sparsity), CAP successfully achieves 99.2% and 96.3% of the original BERT performance in QQP and MNLI tasks. In addition, our probing experiments demonstrate that the model pruned by CAP tends to achieve better generalization ability.
Model compression can significantly reduce the sizes of deep neural network (DNN) models, and thus facilitates the dissemination of sophisticated, sizable DNN models, especially for their deployment on mobile or embedded devices. However, the prediction results of compressed models may deviate from those of their original models. To help developers thoroughly understand the impact of model compression, it is essential to test these models to find those deviated behaviors before dissemination. However, this is a non-trivial task because the architectures and gradients of compressed models are usually not available. To this end, we propose DFLARE, a novel, search-based, black-box testing technique to automatically find triggering inputs that result in deviated behaviors in image classification tasks. DFLARE iteratively applies a series of mutation operations to a given seed image, until a triggering input is found. For better efficacy and efficiency, DFLARE models the search problem as Markov Chains and leverages the Metropolis-Hasting algorithm to guide the selection of mutation operators in each iteration. Further, DFLARE utilizes a novel fitness function to prioritize the mutated inputs that either cause large differences between two models' outputs, or trigger previously unobserved models' probability vectors. We evaluated DFLARE on 21 compressed models for image classification tasks with three datasets. The results show that DFLARE outperforms the baseline in terms of efficacy and efficiency. We also demonstrated that the triggering inputs found by DFLARE can be used to repair up to 48.48% deviated behaviors in image classification tasks and further decrease the effectiveness of DFLARE on the repaired models.
Recently, vision transformer (ViT) and its variants have achieved promising performances in various computer vision tasks. Yet the high computational costs and training data requirements of ViTs limit their application in resource-constrained settings. Model compression is an effective method to speed up deep learning models, but the research of compressing ViTs has been less explored. Many previous works concentrate on reducing the number of tokens. However, this line of attack breaks down the spatial structure of ViTs and is hard to be generalized into downstream tasks. In this paper, we design a unified framework for structural pruning of both ViTs and its variants, namely UP-ViTs. Our method focuses on pruning all ViTs components while maintaining the consistency of the model structure. Abundant experimental results show that our method can achieve high accuracy on compressed ViTs and variants, e.g., UP-DeiT-T achieves 75.79% accuracy on ImageNet, which outperforms the vanilla DeiT-T by 3.59% with the same computational cost. UP-PVTv2-B0 improves the accuracy of PVTv2-B0 by 4.83% for ImageNet classification. Meanwhile, UP-ViTs maintains the consistency of the token representation and gains consistent improvements on object detection tasks.
Network quantization is a dominant paradigm of model compression. However, the abrupt changes in quantized weights during training often lead to severe loss fluctuations and result in a sharp loss landscape, making the gradients unstable and thus degrading the performance. Recently, Sharpness-Aware Minimization (SAM) has been proposed to smooth the loss landscape and improve the generalization performance of the models. Nevertheless, directly applying SAM to the quantized models can lead to perturbation mismatch or diminishment issues, resulting in suboptimal performance. In this paper, we propose a novel method, dubbed Sharpness-Aware Quantization (SAQ), to explore the effect of SAM in model compression, particularly quantization for the first time. Specifically, we first provide a unified view of quantization and SAM by treating them as introducing quantization noises and adversarial perturbations to the model weights, respectively. According to whether the noise and perturbation terms depend on each other, SAQ can be formulated into three cases, which are analyzed and compared comprehensively. Furthermore, by introducing an efficient training strategy, SAQ only incurs a little additional training overhead compared with the default optimizer (e.g., SGD or AdamW). Extensive experiments on both convolutional neural networks and Transformers across various datasets (i.e., ImageNet, CIFAR-10/100, Oxford Flowers-102, Oxford-IIIT Pets) show that SAQ improves the generalization performance of the quantized models, yielding the SOTA results in uniform quantization. For example, on ImageNet, SAQ outperforms AdamW by 1.2% on the Top-1 accuracy for 4-bit ViT-B/16. Our 4-bit ResNet-50 surpasses the previous SOTA method by 0.9% on the Top-1 accuracy.
Recognizing less salient features is the key for model compression. However, it has not been investigated in the revolutionary attention mechanisms. In this work, we propose a novel normalization-based attention module (NAM), which suppresses less salient weights. It applies a weight sparsity penalty to the attention modules, thus, making them more computational efficient while retaining similar performance. A comparison with three other attention mechanisms on both Resnet and Mobilenet indicates that our method results in higher accuracy. Code for this paper can be publicly accessed at https://github.com/Christian-lyc/NAM.
Knowledge distillation is an effective and stable method for model compression via knowledge transfer. Conventional knowledge distillation (KD) is to transfer knowledge from a large and well pre-trained teacher network to a small student network, which is a one-way process. Recently, deep mutual learning (DML) has been proposed to help student networks learn collaboratively and simultaneously. However, to the best of our knowledge, KD and DML have never been jointly explored in a unified framework to solve the knowledge distillation problem. In this paper, we investigate that the teacher model supports more trustworthy supervision signals in KD, while the student captures more similar behaviors from the teacher in DML. Based on these observations, we first propose to combine KD with DML in a unified framework. Furthermore, we propose a Semi-Online Knowledge Distillation (SOKD) method that effectively improves the performance of the student and the teacher. In this method, we introduce the peer-teaching training fashion in DML in order to alleviate the student's imitation difficulty, and also leverage the supervision signals provided by the well-trained teacher in KD. Besides, we also show our framework can be easily extended to feature-based distillation methods. Extensive experiments on CIFAR-100 and ImageNet datasets demonstrate the proposed method achieves state-of-the-art performance.
Multi-modal cues, including spatial information, facial expression and voiceprint, are introduced to the speech separation and speaker extraction tasks to serve as complementary information to achieve better performance. However, the introduction of these cues brings about an increasing number of parameters and model complexity, which makes it harder to deploy these models on resource-constrained devices. In this paper, we alleviate the aforementioned problem by proposing a Lightweight Multi-modal framework for Speaker Extraction (LiMuSE). We propose to use GC-equipped TCN, which incorporates Group Communication (GC) and Temporal Convolutional Network (TCN) in the Context Codec module, the audio block and the fusion block. The experiments on the MC_GRID dataset demonstrate that LiMuSE achieves on par or better performance with a much smaller number of parameters and less model complexity. We further investigate the impacts of the quantization of LiMuSE. Our code and dataset are provided.
In recent years, large-scale deep models have achieved great success, but the huge computational complexity and massive storage requirements make it a great challenge to deploy them in resource-limited devices. As a model compression and acceleration method, knowledge distillation effectively improves the performance of small models by transferring the dark knowledge from the teacher detector. However, most of the existing distillation-based detection methods mainly imitating features near bounding boxes, which suffer from two limitations. First, they ignore the beneficial features outside the bounding boxes. Second, these methods imitate some features which are mistakenly regarded as the background by the teacher detector. To address the above issues, we propose a novel Feature-Richness Score (FRS) method to choose important features that improve generalized detectability during distilling. The proposed method effectively retrieves the important features outside the bounding boxes and removes the detrimental features within the bounding boxes. Extensive experiments show that our methods achieve excellent performance on both anchor-based and anchor-free detectors. For example, RetinaNet with ResNet-50 achieves 39.7% in mAP on the COCO2017 dataset, which even surpasses the ResNet-101 based teacher detector 38.9% by 0.8%. Our implementation is available at https://github.com/duzhixing/FRS.
Despite their tremendous successes, convolutional neural networks (CNNs) incur high computational/storage costs and are vulnerable to adversarial perturbations. Recent works on robust model compression address these challenges by combining model compression techniques with adversarial training. But these methods are unable to improve throughput (frames-per-second) on real-life hardware while simultaneously preserving robustness to adversarial perturbations. To overcome this problem, we propose the method of Generalized Depthwise-Separable (GDWS) convolution -- an efficient, universal, post-training approximation of a standard 2D convolution. GDWS dramatically improves the throughput of a standard pre-trained network on real-life hardware while preserving its robustness. Lastly, GDWS is scalable to large problem sizes since it operates on pre-trained models and doesn't require any additional training. We establish the optimality of GDWS as a 2D convolution approximator and present exact algorithms for constructing optimal GDWS convolutions under complexity and error constraints. We demonstrate the effectiveness of GDWS via extensive experiments on CIFAR-10, SVHN, and ImageNet datasets. Our code can be found at https://github.com/hsndbk4/GDWS.