Vision-Language Transformers (VLTs) have achieved remarkable success, yet their high computational costs remain challenging due to numerous input tokens and large model parameters. Existing VLT compression methods primarily rely on single-modality-based token pruning or coarse-grained weight pruning techniques. However, these methods face significant obstacles, such as ignoring the critical alignment of different modalities and lacking layer-wise dynamic token pruning flexibility, exhibiting inevitable performance degradation due to coarsegrained weight pruning, and struggling with the simultaneous compression of both input tokens and model parameters. To address those limitations, we propose MADTP++, a novel approach that integrates custom-made token and weight pruning processes into a unified framework, achieving superior compression in both parameter counts and computational costs. Specifically, for the token pruning process, we introduce the Multi-modality Alignment Guidance (MAG) module and the Dynamic Token Pruning (DTP) module to align semantic features across different modalities and guide the dynamic elimination of redundant tokens based on different input instances. For the weight pruning process, we propose a Hardware-aware Weight Pruning (HWP) module that leverages the Sparse Tensor Cores across diverse hardware setups to enable fine-grained parameter pruning within VLTs. To further unify token and weight pruning, we also propose a Cooperative Optimization Training Strategy that automatically allocates GFLOPs and parameter reductions per branch before pruning and employs Knowledge Distillation Constraints to facilitate joint optimization of both pruning dimensions. Extensive experiments conducted on various VLT models and datasets demonstrate that MADTP++ can significantly reduce model parameters and computational costs while maintaining competitive performance.

Vision-language model (VLM) is one of the most important models for multi-modal tasks. Real industrial applications often meet the challenge of adapting VLMs to different scenarios, such as varying hardware platforms or performance requirements. Traditional methods involve training or fine-tuning to adapt multiple unique VLMs or using model compression techniques to create multiple compact models. These approaches are complex and resource-intensive. This paper introduces a novel paradigm called Once-Tuning-Multiple-Variants (OTMV). OTMV requires only a single tuning process to inject dynamic weight expansion capacity into the original VLM structure. This tuned VLM can then be expanded into multiple variants tailored for different scenarios in inference. The tuning mechanism of OTMV is inspired by the mathematical series expansion theorem, which helps to reduce the parameter size and memory requirements while maintaining accuracy for VLM. Experiment results show that OTMV-tuned models achieve comparable accuracy to baseline VLMs across various visual-language tasks. The experiments also demonstrate the dynamic expansion capability of OTMV-tuned VLMs, outperforming traditional model compression and adaptation techniques in terms of accuracy and efficiency.

Model quantization reduces the bit-width of weights and activations, improving memory efficiency and inference speed in diffusion models. However, achieving 4-bit quantization remains challenging. Existing methods, primarily based on integer quantization and post-training quantization fine-tuning, struggle with inconsistent performance. Inspired by the success of floating-point (FP) quantization in large language models, we explore low-bit FP quantization for diffusion models and identify key challenges: the failure of signed FP quantization to handle asymmetric activation distributions, the insufficient consideration of temporal complexity in the denoising process during fine-tuning, and the misalignment between fine-tuning loss and quantization error. To address these challenges, we propose the mixup-sign floating-point quantization (MSFP) framework, first introducing unsigned FP quantization in model quantization, along with timestep-aware LoRA (TALoRA) and denoising-factor loss alignment (DFA), which ensure precise and stable fine-tuning. Extensive experiments show that we are the first to achieve superior performance in 4-bit FP quantization for diffusion models, outperforming existing PTQ fine-tuning methods in 4-bit INT quantization.

The rapid improvement of deep learning models with the integration of the physical world has dramatically improved embodied AI capabilities. Meanwhile, the powerful embodied AI models and their scales place an increasing burden on deployment efficiency. The efficiency issue is more apparent on embodied AI platforms than on data centers because they have more limited computational resources and memory bandwidth. Meanwhile, most embodied AI scenarios, like autonomous driving and robotics, are more sensitive to fast responses. Theoretically, the traditional model compression techniques can help embodied AI models with more efficient computation, lower memory and energy consumption, and reduced latency. Because the embodied AI models are expected to interact with the physical world, the corresponding compressed models are also expected to resist natural corruption caused by real-world events such as noise, blur, weather conditions, and even adversarial corruption. This paper explores the novel paradigm to boost the efficiency of the embodied AI models and the robust compression boundary. The efficacy of our method has been proven to find the optimal balance between accuracy, efficiency, and robustness in real-world conditions.

Taylor-Series-Expansion (TSE) is a mathematics theorem. It proves that the expansion of the first few finite Taylor Series is a good approximation of a nonlinear function in most cases. Inspired by the TSE theorem, a brand-new TSE-based vision transformer is designed. TSE-based vision transformer uses the shared first-order TSE transformer block's weight (in analogy with the Taylor-Series first-order term), its finite multiple multiplications (in analogy with the Taylor-Series expanded high-order terms), and the corresponding learnable TSE coefficients to approximate the naive vision transformer. In this manner, the TSE-based vision model reduces the memory burden but keeps a similar accuracy as the naive counterpart. Derived from adding the Taylor skip mechanism in training, the TSE-based vision transformer has good dynamic expansion capability. Experiment results show TSE-based models can boost actual deployment latency by 1.30-1.36× on A100 GPU and 1.34-1.45× on AGX Orin with negligible accuracy degradation on ImageNet classification, COCO detection, and ADE20K segmentation benchmarking tasks. Moreover, TSE-based optimization is orthogonal to model compression. Combining with the state-of-the-art vision transformer compression method, it can boost actual deployment performance by 1.70-1.87× and 3.29-3.61× of latency and throughput on A100 GPU, and 1.67-1.74× and 2.76-2.94× improvement of latency and throughput on AGX Orin.

Improving the efficiency of Neural Architecture Search (NAS) is a challenging but significant task that has received much attention. Previous studies mainly adopt the Differentiable Architecture Search (DARTS) and improve its search strategies or modules to enhance search efficiency. Recently, some methods have started considering data reduction for speedup, but they are not tightly coupled with the architecture search process and cannot capture the training dynamics of DARTS well, resulting in sub-optimal performances. To this end, this work pioneers an exploration into the critical role of dataset characteristics in the bi-level optimization of DARTS, and then proposes a novel Bi-level Data Pruning (BDP) paradigm that targets the weights and architecture levels of DARTS to enhance efficiency from a data perspective. Specifically, we introduce a progressive bi-level data pruning strategy that utilizes supernet prediction dynamics as the metric to gradually prune unsuitable samples for DARTS during the search. An effective automatic class balance constraint is also integrated into BDP, to suppress potential class imbalances resulting from data-efficient algorithms. Comprehensive evaluations on the NAS-Bench-201 search space, DARTS search space, and MobileNet-like search space validate that BDP reduces search costs by over 50% while achieving superior performance when applied to the baseline DARTS. Besides, we demonstrate that BDP can harmoniously integrate with advanced DARTS variants, like P-DARTS, PC-DARTS, EG-NAS, and$\beta $-DARTS, offering an approximately$2\times $speedup with minimal performance compromise.

Recent research understands the residual networks from a new perspective of the implicit ensemble model. From this view, previous methods such as stochastic depth and stimulative training have further improved the performance of the residual network by sampling and training of its subnets. However, they both use the same supervision for all subnets of different capacities and neglect the valuable knowledge generated by subnets during training. In this manuscript, we mitigate the significant knowledge distillation gap caused by using the same kind of supervision and advocate leveraging the subnets to provide diverse knowledge. Based on this motivation, we propose a group knowledge based training framework for boosting the performance of residual networks. Specifically, we implicitly divide all subnets into hierarchical groups by subnet-in-subnet sampling, aggregate the knowledge of different subnets in each group during training, and exploit upper-level group knowledge to supervise lower-level subnet group. Meanwhile, we also develop a subnet sampling strategy that naturally samples larger subnets, which are found to be more helpful than smaller subnets in boosting performance for hierarchical groups. Compared with typical subnet training and other methods, our method achieves the best efficiency and performance trade-offs on multiple datasets and network structures. The code is at https://github.com/tsj-001/AAAI24-GKT.

Vision-Language Transformers (VLTs) have shown great success recently, but are meanwhile accompanied by heavy computation costs, where a major reason can be attributed to the large number of visual and language tokens. Existing token pruning research for compressing VLTs mainly follows a single-modality-based scheme yet ignores the critical role of aligning different modalities for guiding the token pruning process, causing the important tokens for one modality to be falsely pruned in another modality branch. Meanwhile, existing VLT pruning works also lack the flexibility to dynamically compress each layer based on different input samples. To this end, we propose a novel framework named Multimodal Alignment-Guided Dynamic Token Pruning (MADTP) for accelerating various VLTs. Specifically, we first introduce a well-designed Multi-modality Alignment Guidance (MAG) module that can align features of the same semantic concept from different modalities, to ensure the pruned tokens are less important for all modalities. We further design a novel Dynamic Token Pruning (DTP) module, which can adaptively adjust the token compression ratio in each layer based on different input instances. Extensive experiments on various benchmarks demonstrate that MADTP significantly reduces the computational complexity of kinds of multimodal models while preserving competitive performance. Notably, when applied to the BLIP model in the NLVR2 dataset, MADTP can reduce the GFLOPs by 80% with less than 4% performance degradation. The code is available at https://github.com/double125IMADTP.

Recent Vision Transformer Compression (VTC) works mainly follow a two-stage scheme, where the importance score of each model unit is first evaluated or preset in each submodule, followed by the sparsity score evaluation ac-cording to the target sparsity constraint. Such a separate evaluation process induces the gap between importance and sparsity score distributions, thus causing high search costs for VTC. In this work, for the first time, we investigate how to integrate the evaluations of importance and sparsity scores into a single stage, searching the optimal subnets in an effi-cient manner. Specifically, we present OFB, a cost-efficient approach that simultaneously evaluates both importance and sparsity scores, termed Once for Both (OFB), for VTC. First, a bi-mask scheme is developed by entangling the importance score and the differentiable sparsity score to jointly deter-mine the pruning potential (prunability) of each unit. Such a bi-mask search strategy is further used together with a proposed adaptive one-hot loss to realize the progressive-and-efficient search for the most important subnet. Finally, Progressive Masked Image Modeling (PMIM) is proposed to regularize the feature space to be more representative during the search process, which may be degraded by the dimension reduction. Extensive experiments demonstrate that OFB can achieve superior compression performance over state-of-the-art searching-based and pruning-based methods under various Vision Transformer architectures, meanwhile pro-moting search efficiency significantly, e.g., costing one GPU search day for the compression of DeiT-S on ImageNet-1K.

Dehazing real-world hazy images is challenging due to the complexity of natural haze, varying haze conditions, details preservation, and the risk of overexposure. Existing methods excel in synthetic hazy scenarios but struggle in the real world because they don’t use all available features. Classical dehazing techniques primarily focus on low-level dehazing enhancements, whereas deep learning-based methods extract more intricate weather-related features. However, both of these approaches exhibit limitations in effectively addressing the real-world dehazing. To address these challenges, we introduce an innovative approach that combines the strengths of both modalities to dehaze and enhance the visibility of real-world hazy scenes. Firstly, we extract both low-level and deep features and then employ a pre-trained vector quantization GAN to create well-detailed data patches. A decoder, with a normalized module, effectively utilizes these high-quality features. Additionally, we introduce a controllable operation to improve feature matching. To further enhance dehazing and generalizability, the decoder’s output undergoes a sequence of gamma-correction operations and generates a series of multi-exposure images that are combined to create a haze-free and higher-quality image. Our method effectively reduces haziness, enhances sharpness, preserves natural colors, and minimizes artifacts in challenging real-world scenarios. The approach surpasses five SOTA methods in both qualitative and quantitative evaluations across three key metrics, utilizing three synthetic and two real-world hazy datasets. Notably, it achieves a substantial improvement in real-world datasets over the second-best method, with 0.5702 and 0.129 in FADE metrics for the RTTS and Fattal datasets, respectively.

Recently, differentiable mask pruning methods optimize the continuous relaxation architecture (soft network) as the proxy of the pruned discrete network (hard network) for superior sub-architecture search. However, due to the agnostic impact of the discretization process, the hard network struggles with the equivalent representational capacity as the soft network, namely discretization gap, which severely spoils the pruning performance. In this paper, we first investigate the discretization gap and propose a novel structural differentiable mask pruning framework named S2HPruner to bridge the discretization gap in a one-stage manner. In the training procedure, SH2Pruner forwards both the soft network and its corresponding hard network, then distills the hard network under the supervision of the soft network. To optimize the mask and prevent performance degradation, we propose a decoupled bidirectional knowledge distillation. It blocks the weight updating from the hard to the soft network while maintaining the gradient corresponding to the mask. Compared with existing pruning arts, S2HPruner achieves surpassing pruning performance without fine-tuning on comprehensive benchmarks, including CIFAR-100, Tiny ImageNet, and ImageNet with a variety of network architectures. Besides, investigation and analysis experiments explain the effectiveness of S2HPruner. Codes will be released soon.

The transformer extends its success from the language to the vision domain. Because of the stacked self-attention and cross-attention blocks, the acceleration deployment of vision transformer on GPU hardware is challenging and also rarely studied. This paper thoroughly designs a compression scheme to maximally utilize the GPU-friendly 2:4 fine-grained structured sparsity and quantization. Specially, an original large model with dense weight parameters is first pruned into a sparse one by 2:4 structured pruning, which considers the GPU's acceleration of 2:4 structured sparse pattern with FP16 data type, then the floating-point sparse model is further quantized into a fixed-point one by sparse-distillation-aware quantization aware training, which considers GPU can provide an extra speedup of 2:4 sparse calculation with integer tensors. A mixed-strategy knowledge distillation is used during the pruning and quantization process. The proposed compression scheme is flexible to support supervised and unsupervised learning styles. Experiment results show GPUSQ-ViT scheme achieves state-of-the-art compression by reducing vision transformer models$\mathbf{6.4}-\mathbf{12.7}\times$on model size and$\mathbf{30.3}-\mathbf{62} \times$on FLOPs with negligible accuracy degradation on ImageNet classification, COCO detection and ADE20K segmentation benchmarking tasks. Moreover, GPUSQ-ViT can boost actual deployment performance by$\mathbf{1.39}-\mathbf{1.79}\times$and$\mathbf{3.22}-\mathbf{3.43}\times$of latency and throughput on A100 GPU, and$\mathbf{1.57}-\mathbf{1.69}\times$and$\mathbf{2.11}-\mathbf{2.51}\times$improvement of latency and throughput on AGX Orin.

Adversarial attack is commonly regarded as a huge threat to neural networks because of misleading behavior. This paper presents an opposite perspective: adversarial attacks can be harnessed to improve neural models if amended correctly. Unlike traditional adversarial defense or adversarial training schemes that aim to improve the adversarial robustness, the proposed adversarial amendment (AdvAmd) method aims to improve the original accuracy level of neural models on benign samples. We thoroughly analyze the distribution mismatch between the benign and adversarial samples. This distribution mismatch and the mutual learning mechanism with the same learning ratio applied in prior art defense strategies is the main cause leading the accuracy degradation for benign samples. The proposed AdvAmd is demonstrated to steadily heal the accuracy degradation and even leads to a certain accuracy boost of common neural models on benign classification, object detection, and segmentation tasks. The efficacy of the AdvAmd is contributed by three key components: mediate samples (to reduce the influence of distribution mismatch with a fine-grained amendment), auxiliary batch norm (to solve the mutual learning mechanism and the smoother judgment surface), and AdvAmd loss (to adjust the learning ratios according to different attack vulnerabilities) through quantitative and ablation experiments.

Semi-supervised video object segmentation (Semi-VOS), which requires only annotating the first frame of a video to segment future frames, has received increased attention recently. Among existing Semi-VOS pipelines, the memory-matching-based one is becoming the main research stream, as it can fully utilize the temporal sequence information to obtain high-quality segmentation results. Even though this type of method has achieved promising performance, the overall framework still suffers from heavy computation overhead, mainly caused by the per-frame dense convolution operations between high-resolution feature maps and each kernel filter. Therefore, we propose a sparse baseline of VOS named SpVOS in this work, which develops a novel triple sparse convolution to reduce the computation costs of the overall VOS framework. The designed triple gate, taking full consideration of both spatial and temporal redundancy between adjacent video frames, adaptively makes a triple decision to decide how to apply the sparse convolution on each pixel to control the computation overhead of each layer, while maintaining sufficient discrimination capability to distinguish similar objects and avoid error accumulation. A mixed sparse training strategy, coupled with a designed objective considering the sparsity constraint, is also developed to balance the VOS segmentation performance and computation costs. Experiments are conducted on two mainstream VOS datasets, including DAVIS and Youtube-VOS. Results show that, the proposed SpVOS achieves superior performance over other state-of-the-art sparse methods, and even maintains comparable performance, e.g., an 83.04% (79.29%) overall score on the DAVIS-2017 (Youtube-VOS) validation set, with the typical non-sparse VOS baseline (82.88% for DAVIS-2017 and 80.36% for Youtube-VOS) while saving up to 42% FLOPs, showing its application potential for resource-constrained scenarios.

With the development of deep learning, neural networks tend to be deeper and larger to achieve good performance. Trained models are more compute-intensive and memory-intensive, which lead to the big challenges on memory band-width, storage, latency, and throughput. In this paper, we propose the neural network compression method named minimally invasive surgery. Different from traditional model compression and knowledge distillation methods, the proposed method refers to the minimally invasive surgery principle. It learns the principal features from a pair of dense and compressed models in a contrastive manner. It also optimizes the neural networks to meet the specific hardware acceleration requirements. Through qualitative, quantitative, and ablation experiments, the proposed method shows a compelling performance, acceleration, and generalization in various tasks.

We introduce channel permutations as a method to maximize the accuracy of N:M sparse networks. N:M sparsity requires N out of M consecutive elements to be zero and has been shown to maintain accuracy for many models and tasks with a simple prune and fine-tune workflow. By permuting weight matrices along their channel dimension and adjusting the surrounding layers appropriately, we demonstrate accuracy recovery for even small, parameter-efficient networks, without affecting inference run-time. We also present both a quality metric to simplify judging permutations as well as efficient methods to search for high-quality permutations, including two optimizations to escape local minima. Finally, we share an ablation study to show the importance of each part of our search algorithm, experimental results showing correlation between our quality metric and final network accuracy, improved sparse network accuracy using our techniques with insignificant overhead to training time, and the transformation of unstructured to structured sparse workloads. Code to use these techniques when generating a 2:4 sparse network is available at https://github.com/NVIDIA/apex/tree/master/apex/contrib/sparsity.

Deep learning’s success has led to larger and larger models to handle more and more complex tasks; trained models often contain millions of parameters. These large models are compute- and memory-intensive, which makes it a challenge to deploy them with latency, throughput, and storage constraints. Some model compression methods have been successfully applied to image classification and detection or language models, but there has been very little work compressing generative adversarial networks (GANs) performing complex tasks. In this paper, we show that a standard model compression technique, weight pruning and knowledge distillation, cannot be applied to GANs using existing methods. We then develop a self-supervised compression technique which uses the trained discriminator to supervise the training of a compressed generator. We show that this framework has compelling performance to high degrees of sparsity, can be easily applied to new tasks and models, and enables meaningful comparisons between different compression granularities.

Because of the intrinsic complexity in computation, three-dimensional (3D) reconstruction is an essential and challenging topic in computer vision research and applications. The existing methods for 3D reconstruction often produce holes, distortions and obscure parts in the reconstructed 3D models, or can only reconstruct voxelized 3D models for simple isolated objects. So they are not adequate for real usage. From 2014, the Generative Adversarial Network (GAN) is widely used in generating unreal dataset and semi-supervised learning. So the focus of this paper is to achieve high quality 3D reconstruction performance by adopting GAN principle. We propose a novel semi-supervised 3D reconstruction framework, namely SS-3D-GAN, which can iteratively improve any raw 3D reconstruction models by training the GAN models to converge. This new model only takes real-time 2D observation images as the weak supervision, and doesn't rely on prior knowledge of shape models or any referenced observations. Finally, through the qualitative and quantitative experiments & analysis, this new method shows compelling advantages over the current state-of-the-art methods on Tanks & Temples reconstruction benchmark dataset.

In this letter, we proposed a high-resolution satellite image classification method using morphological component analysis of texture and cartoon layers. The construction of the dictionary matrix used in the algorithm is based on independent component analysis. After the decomposition, we obtain the morphological coefficient vectors in both texture and cartoon layers which are termed as the sparse representation of the input high-resolution satellite image. By combining the features from two layers, the total probability of the target image classification is calculated out according to the maximum likelihood mechanism. Quantitative analysis on experiment results and comparisons with classic image classification algorithms have proven that the proposed classification method has better accuracy, efficiency, and performance than most of state-of-the-art classification methods.