With the rapid development of computer vision, Vision Transformers (ViTs) offer the tantalizing prospect of unified information processing across visual and textual domains. But due to the lack of inherent inductive biases in ViTs, they require enormous amount of data for training. To make their applications practical, we introduce an innovative ensemble-based distillation approach distilling inductive bias from complementary lightweight teacher models. Prior systems relied solely on convolution-based teaching. However, this method incorporates an ensemble of light teachers with different architectural tendencies, such as convolution and involution, to instruct the student transformer jointly. Because of these unique inductive biases, instructors can accumulate a wide range of knowledge, even from readily identifiable stored datasets, which leads to enhanced student performance. Our proposed framework also involves precomputing and storing logits in advance, essentially the unnormalized predictions of the model. This optimization can accelerate the distillation process by eliminating the need for repeated forward passes during knowledge distillation, significantly reducing the computational burden and enhancing efficiency.
Vision Transformers (ViTs) have emerged as state-of-the-art models for various vision tasks recently. However, their heavy computation costs remain daunting for resource-limited devices. Consequently, researchers have dedicated themselves to compressing redundant information in ViTs for acceleration. However, they generally sparsely drop redundant image tokens by token pruning or brutally remove channels by channel pruning, leading to a sub-optimal balance between model performance and inference speed. They are also disadvantageous in transferring compressed models to downstream vision tasks that require the spatial structure of images, such as semantic segmentation. To tackle these issues, we propose a joint compression method for ViTs that offers both high accuracy and fast inference speed, while also maintaining favorable transferability to downstream tasks (CAIT). Specifically, we introduce an asymmetric token merging (ATME) strategy to effectively integrate neighboring tokens. It can successfully compress redundant token information while preserving the spatial structure of images. We further employ a consistent dynamic channel pruning (CDCP) strategy to dynamically prune unimportant channels in ViTs. Thanks to CDCP, insignificant channels in multi-head self-attention modules of ViTs can be pruned uniformly, greatly enhancing the model compression. Extensive experiments on benchmark datasets demonstrate that our proposed method can achieve state-of-the-art performance across various ViTs. For example, our pruned DeiT-Tiny and DeiT-Small achieve speedups of 1.7$\times$ and 1.9$\times$, respectively, without accuracy drops on ImageNet. On the ADE20k segmentation dataset, our method can enjoy up to 1.31$\times$ speedups with comparable mIoU. Our code will be publicly available.
Recent advances with self-supervised learning have allowed speech recognition systems to achieve state-of-the-art (SOTA) word error rates (WER) while requiring only a fraction of the labeled training data needed by its predecessors. Notwithstanding, while such models achieve SOTA performance in matched train/test conditions, their performance degrades substantially when tested in unseen conditions. To overcome this problem, strategies such as data augmentation and/or domain shift training have been explored. Available models, however, are still too large to be considered for edge speech applications on resource-constrained devices, thus model compression tools are needed. In this paper, we explore the effects that train/test mismatch conditions have on speech recognition accuracy based on compressed self-supervised speech models. In particular, we report on the effects that parameter quantization and model pruning have on speech recognition accuracy based on the so-called robust wav2vec 2.0 model under noisy, reverberant, and noise-plus-reverberation conditions.
Keyword spotting (KWS) refers to the task of identifying a set of predefined words in audio streams. With the advances seen recently with deep neural networks, it has become a popular technology to activate and control small devices, such as voice assistants. Relying on such models for edge devices, however, can be challenging due to hardware constraints. Moreover, as adversarial attacks have increased against voice-based technologies, developing solutions robust to such attacks has become crucial. In this work, we propose VIC-KD, a robust distillation recipe for model compression and adversarial robustness. Using self-supervised speech representations, we show that imposing geometric priors to the latent representations of both Teacher and Student models leads to more robust target models. Experiments on the Google Speech Commands datasets show that the proposed methodology improves upon current state-of-the-art robust distillation methods, such as ARD and RSLAD, by 12% and 8% in robust accuracy, respectively.
The ability to dynamically adjust the computational load of neural models during inference is crucial for on-device processing scenarios characterised by limited and time-varying computational resources. A promising solution is presented by early-exit architectures, in which additional exit branches are appended to intermediate layers of the encoder. In self-attention models for automatic speech recognition (ASR), early-exit architectures enable the development of dynamic models capable of adapting their size and architecture to varying levels of computational resources and ASR performance demands. Previous research on early-exiting ASR models has relied on pre-trained self-supervised models, fine-tuned with an early-exit loss. In this paper, we undertake an experimental comparison between fine-tuning pre-trained backbones and training models from scratch with the early-exiting objective. Experiments conducted on public datasets reveal that early-exit models trained from scratch not only preserve performance when using fewer encoder layers but also exhibit enhanced task accuracy compared to single-exit or pre-trained models. Furthermore, we explore an exit selection strategy grounded in posterior probabilities as an alternative to the conventional frame-based entropy approach. Results provide insights into the training dynamics of early-exit architectures for ASR models, particularly the efficacy of training strategies and exit selection methods.
Large language models(LLMs) containing tens of billions of parameters (or even more) have demonstrated impressive capabilities in various NLP tasks. However, substantial model size poses challenges to training, inference, and deployment so that it is necessary to compress the model. At present, most model compression for LLMs requires manual design of pruning features, which has problems such as complex optimization pipeline and difficulty in retaining the capabilities of certain parts of the model.Therefore, we propose a novel pruning approach: firstly, a training set of a certain number of architecture-accuracy pairs is established, and then a non-neural model is trained as an accuracy predictor. Using the accuracy predictor to further optimize the search space and search, the optimal model can be automatically selected. Experiments show that our proposed approach is effective and efficient. Compared with the baseline, the perplexity(PPL) on Wikitext2 and PTB dropped by 9.48% and 5,76% respectively, and the average accuracy of MMLU increased by 6.28%.
Tiny, causal models are crucial for embedded audio machine learning applications. Model compression can be achieved via distilling knowledge from a large teacher into a smaller student model. In this work, we propose a novel two-step approach for tiny speech enhancement model distillation. In contrast to the standard approach of a weighted mixture of distillation and supervised losses, we firstly pre-train the student using only the knowledge distillation (KD) objective, after which we switch to a fully supervised training regime. We also propose a novel fine-grained similarity-preserving KD loss, which aims to match the student's intra-activation Gram matrices to that of the teacher. Our method demonstrates broad improvements, but particularly shines in adverse conditions including high compression and low signal to noise ratios (SNR), yielding signal to distortion ratio gains of 0.9 dB and 1.1 dB, respectively, at -5 dB input SNR and 63x compression compared to baseline.
Large-scale self-supervised pre-trained speech encoders outperform conventional approaches in speech recognition and translation tasks. Due to the high cost of developing these large models, building new encoders for new tasks and deploying them to on-device applications are infeasible. Prior studies propose model compression methods to address this issue, but those works focus on smaller models and less realistic tasks. Thus, we propose Contrastive Layer-to-layer Distillation (CoLLD), a novel knowledge distillation method to compress pre-trained speech encoders by leveraging masked prediction and contrastive learning to train student models to copy the behavior of a large teacher model. CoLLD outperforms prior methods and closes the gap between small and large models on multilingual speech-to-text translation and recognition benchmarks.
The edge processing of deep neural networks (DNNs) is becoming increasingly important due to its ability to extract valuable information directly at the data source to minimize latency and energy consumption. Frequency-domain model compression, such as with the Walsh-Hadamard transform (WHT), has been identified as an efficient alternative. However, the benefits of frequency-domain processing are often offset by the increased multiply-accumulate (MAC) operations required. This paper proposes a novel approach to an energy-efficient acceleration of frequency-domain neural networks by utilizing analog-domain frequency-based tensor transformations. Our approach offers unique opportunities to enhance computational efficiency, resulting in several high-level advantages, including array micro-architecture with parallelism, ADC/DAC-free analog computations, and increased output sparsity. Our approach achieves more compact cells by eliminating the need for trainable parameters in the transformation matrix. Moreover, our novel array micro-architecture enables adaptive stitching of cells column-wise and row-wise, thereby facilitating perfect parallelism in computations. Additionally, our scheme enables ADC/DAC-free computations by training against highly quantized matrix-vector products, leveraging the parameter-free nature of matrix multiplications. Another crucial aspect of our design is its ability to handle signed-bit processing for frequency-based transformations. This leads to increased output sparsity and reduced digitization workload. On a 16$\times$16 crossbars, for 8-bit input processing, the proposed approach achieves the energy efficiency of 1602 tera operations per second per Watt (TOPS/W) without early termination strategy and 5311 TOPS/W with early termination strategy at VDD = 0.8 V.
As the size of large language models (LLMs) continues to grow, model compression without sacrificing accuracy has become a crucial challenge for deployment. While some quantization methods, such as GPTQ, have made progress in achieving acceptable 4-bit weight-only quantization, attempts at lower-bit quantization often result in severe performance degradation. In this paper, we introduce a technique called norm tweaking, which can be used as a plugin in current PTQ methods to achieve high precision while being cost-efficient. Our approach is inspired by the observation that rectifying the quantized activation distribution to match its float counterpart can readily restore accuracy for LLMs. To achieve this, we carefully design a tweaking strategy that includes calibration data generation and channel-wise distance constraint to update the weights of normalization layers for better generalization. We conduct extensive experiments on various datasets using several open-sourced LLMs. Our method demonstrates significant improvements in both weight-only quantization and joint quantization of weights and activations, surpassing existing PTQ methods. On GLM-130B and OPT-66B, our method even achieves the same level of accuracy at 2-bit quantization as their float ones. Our simple and effective approach makes it more practical for real-world applications.
Vision transformer (ViT) and its variants have swept through visual learning leaderboards and offer state-of-the-art accuracy in tasks such as image classification, object detection, and semantic segmentation by attending to different parts of the visual input and capturing long-range spatial dependencies. However, these models are large and computation-heavy. For instance, the recently proposed ViT-B model has 86M parameters making it impractical for deployment on resource-constrained devices. As a result, their deployment on mobile and edge scenarios is limited. In our work, we aim to take a step toward bringing vision transformers to the edge by utilizing popular model compression techniques such as distillation, pruning, and quantization. Our chosen application environment is an unmanned aerial vehicle (UAV) that is battery-powered and memory-constrained, carrying a single-board computer on the scale of an NVIDIA Jetson Nano with 4GB of RAM. On the other hand, the UAV requires high accuracy close to that of state-of-the-art ViTs to ensure safe object avoidance in autonomous navigation, or correct localization of humans in search-and-rescue. Inference latency should also be minimized given the application requirements. Hence, our target is to enable rapid inference of a vision transformer on an NVIDIA Jetson Nano (4GB) with minimal accuracy loss. This allows us to deploy ViTs on resource-constrained devices, opening up new possibilities in surveillance, environmental monitoring, etc. Our implementation is made available at https://github.com/chensy7/efficient-vit.
Low Rank Decomposition (LRD) is a model compression technique applied to the weight tensors of deep learning models in order to reduce the number of trainable parameters and computational complexity. However, due to high number of new layers added to the architecture after applying LRD, it may not lead to a high training/inference acceleration if the decomposition ranks are not small enough. The issue is that using small ranks increases the risk of significant accuracy drop after decomposition. In this paper, we propose two techniques for accelerating low rank decomposed models without requiring to use small ranks for decomposition. These methods include rank optimization and sequential freezing of decomposed layers. We perform experiments on both convolutional and transformer-based models. Experiments show that these techniques can improve the model throughput up to 60% during training and 37% during inference when combined together while preserving the accuracy close to that of the original models
In an age dominated by resource-intensive foundation models, the ability to efficiently adapt to downstream tasks is crucial. Visual Prompting (VP), drawing inspiration from the prompting techniques employed in Large Language Models (LLMs), has emerged as a pivotal method for transfer learning in the realm of computer vision. As the importance of efficiency continues to rise, research into model compression has become indispensable in alleviating the computational burdens associated with training and deploying over-parameterized neural networks. A primary objective in model compression is to develop sparse and/or quantized models capable of matching or even surpassing the performance of their over-parameterized, full-precision counterparts. Although previous studies have explored the effects of model compression on transfer learning, its impact on visual prompting-based transfer remains unclear. This study aims to bridge this gap, shedding light on the fact that model compression detrimentally impacts the performance of visual prompting-based transfer, particularly evident in scenarios with low data volume. Furthermore, our findings underscore the adverse influence of sparsity on the calibration of downstream visual-prompted models. However, intriguingly, we also illustrate that such negative effects on calibration are not present when models are compressed via quantization. This empirical investigation underscores the need for a nuanced understanding beyond mere accuracy in sparse and quantized settings, thereby paving the way for further exploration in Visual Prompting techniques tailored for sparse and quantized models.
Over the past decade, deep learning models have exhibited considerable advancements, reaching or even exceeding human-level performance in a range of visual perception tasks. This remarkable progress has sparked interest in applying deep networks to real-world applications, such as autonomous vehicles, mobile devices, robotics, and edge computing. However, the challenge remains that state-of-the-art models usually demand significant computational resources, leading to impractical power consumption, latency, or carbon emissions in real-world scenarios. This trade-off between effectiveness and efficiency has catalyzed the emergence of a new research focus: computationally efficient deep learning, which strives to achieve satisfactory performance while minimizing the computational cost during inference. This review offers an extensive analysis of this rapidly evolving field by examining four key areas: 1) the development of static or dynamic light-weighted backbone models for the efficient extraction of discriminative deep representations; 2) the specialized network architectures or algorithms tailored for specific computer vision tasks; 3) the techniques employed for compressing deep learning models; and 4) the strategies for deploying efficient deep networks on hardware platforms. Additionally, we provide a systematic discussion on the critical challenges faced in this domain, such as network architecture design, training schemes, practical efficiency, and more realistic model compression approaches, as well as potential future research directions.
The use of large transformer-based models such as BERT, GPT, and T5 has led to significant advancements in natural language processing. However, these models are computationally expensive, necessitating model compression techniques that reduce their size and complexity while maintaining accuracy. This project investigates and applies knowledge distillation for BERT model compression, specifically focusing on the TinyBERT student model. We explore various techniques to improve knowledge distillation, including experimentation with loss functions, transformer layer mapping methods, and tuning the weights of attention and representation loss and evaluate our proposed techniques on a selection of downstream tasks from the GLUE benchmark. The goal of this work is to improve the efficiency and effectiveness of knowledge distillation, enabling the development of more efficient and accurate models for a range of natural language processing tasks.
Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization (\textbf{OmniQuant}) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \url{https://github.com/OpenGVLab/OmniQuant}.
Vision-Language Pre-training (VLP) shows remarkable progress with the assistance of extremely heavy parameters, which challenges deployment in real applications. Knowledge distillation is well recognized as the essential procedure in model compression. However, existing knowledge distillation techniques lack an in-depth investigation and analysis of VLP, and practical guidelines for VLP-oriented distillation are still not yet explored. In this paper, we present DLIP, a simple yet efficient Distilling Language-Image Pre-training framework, through which we investigate how to distill a light VLP model. Specifically, we dissect the model distillation from multiple dimensions, such as the architecture characteristics of different modules and the information transfer of different modalities. We conduct comprehensive experiments and provide insights on distilling a light but performant VLP model. Experimental results reveal that DLIP can achieve a state-of-the-art accuracy/efficiency trade-off across diverse cross-modal tasks, e.g., image-text retrieval, image captioning and visual question answering. For example, DLIP compresses BLIP by 1.9x, from 213M to 108M parameters, while achieving comparable or better performance. Furthermore, DLIP succeeds in retaining more than 95% of the performance with 22.4% parameters and 24.8% FLOPs compared to the teacher model and accelerates inference speed by 2.7x.
Text-based Person Search (TBPS) aims to retrieve the person images using natural language descriptions. Recently, Contrastive Language Image Pretraining (CLIP), a universal large cross-modal vision-language pre-training model, has remarkably performed over various cross-modal downstream tasks due to its powerful cross-modal semantic learning capacity. TPBS, as a fine-grained cross-modal retrieval task, is also facing the rise of research on the CLIP-based TBPS. In order to explore the potential of the visual-language pre-training model for downstream TBPS tasks, this paper makes the first attempt to conduct a comprehensive empirical study of CLIP for TBPS and thus contribute a straightforward, incremental, yet strong TBPS-CLIP baseline to the TBPS community. We revisit critical design considerations under CLIP, including data augmentation and loss function. The model, with the aforementioned designs and practical training tricks, can attain satisfactory performance without any sophisticated modules. Also, we conduct the probing experiments of TBPS-CLIP in model generalization and model compression, demonstrating the effectiveness of TBPS-CLIP from various aspects. This work is expected to provide empirical insights and highlight future CLIP-based TBPS research.
Efficiency has been a critical problem in UAV tracking due to limitations in computation resources, battery capacity, and unmanned aerial vehicle maximum load. Although discriminative correlation filters (DCF)-based trackers prevail in this field for their favorable efficiency, some recently proposed lightweight deep learning (DL)-based trackers using model compression demonstrated quite remarkable CPU efficiency as well as precision. Unfortunately, the model compression methods utilized by these works, though simple, are still unable to achieve satisfying tracking precision with higher compression rates. This paper aims to exploit disentangled representation learning with mutual information maximization (DR-MIM) to further improve DL-based trackers' precision and efficiency for UAV tracking. The proposed disentangled representation separates the feature into an identity-related and an identity-unrelated features. Only the latter is used, which enhances the effectiveness of the feature representation for subsequent classification and regression tasks. Extensive experiments on four UAV benchmarks, including UAV123@10fps, DTB70, UAVDT and VisDrone2018, show that our DR-MIM tracker significantly outperforms state-of-the-art UAV tracking methods.
Multi-view 3D detection based on BEV (bird-eye-view) has recently achieved significant improvements. However, the huge memory consumption of state-of-the-art models makes it hard to deploy them on vehicles, and the non-trivial latency will affect the real-time perception of streaming applications. Despite the wide application of quantization to lighten models, we show in our paper that directly applying quantization in BEV tasks will 1) make the training unstable, and 2) lead to intolerable performance degradation. To solve these issues, our method QD-BEV enables a novel view-guided distillation (VGD) objective, which can stabilize the quantization-aware training (QAT) while enhancing the model performance by leveraging both image features and BEV features. Our experiments show that QD-BEV achieves similar or even better accuracy than previous methods with significant efficiency gains. On the nuScenes datasets, the 4-bit weight and 6-bit activation quantized QD-BEV-Tiny model achieves 37.2% NDS with only 15.8 MB model size, outperforming BevFormer-Tiny by 1.8% with an 8x model compression. On the Small and Base variants, QD-BEV models also perform superbly and achieve 47.9% NDS (28.2 MB) and 50.9% NDS (32.9 MB), respectively.