Training-free video understanding methods leverage the strong image comprehension capabilities of pre-trained vision language models (VLMs) by treating videos as a sequences of static frames, thus obviating the need for costly video-specific training. However, this paradigm often suffers from severe visual redundancy and high computational overhead, especially when processing long videos. Crucially, existing keyframe selection strategies, especially those based on CLIP similarity, are prone to biases and may inadvertently overlook critical frames, resulting in suboptimal video comprehension. To address these significant challenges, we propose KTV, a novel two-stage framework for efficient and effective training-free video understanding. In the first stage, KTV performs question-agnostic keyframe selection by clustering frame-level visual features, yielding a compact, diverse, and representative subset of frames that mitigates temporal redundancy. In the second stage, KTV applies key visual token selection, pruning redundant or less informative tokens from each selected keyframe based on token importance and redundancy, which significantly reduces the number of tokens fed into the LLM. Extensive experiments on the Multiple-Choice VideoQA task demonstrate that KTV outperforms state-of-the-art training-free baselines while using significantly fewer visual tokens, e.g., only 504 tokens for a 60 min video with 10800 frames, achieving 44.8% accuracy on the MLVU-Test benchmark. In particular, KTV also exceeds several training-based approaches on certain benchmarks.

Despite the efficiency of prompt learning in transferring vision-language models (VLMs) to downstream tasks, existing methods mainly learn the prompts in a coarse-grained manner where the learned prompt vectors are shared across all categories. Consequently, the tailored prompts often fail to discern class-specific visual concepts, thereby hindering the transferred performance for classes that share similar or complex visual attributes. Recent advances mitigate this challenge by leveraging external knowledge from Large Language Models (LLMs) to furnish class descriptions, yet incurring notable inference costs. In this paper, we introduce TextRefiner, a plug-and-play method to refine the text prompts of existing methods by leveraging the internal knowledge of VLMs. Particularly, TextRefiner builds a novel local cache module to encapsulate fine-grained visual concepts derived from local tokens within the image branch. By aggregating and aligning the cached visual descriptions with the original output of the text branch, TextRefiner can efficiently refine and enrich the learned prompts from existing methods without relying on any external expertise. For example, it improves the performance of CoOp from 71.66% to 76.96% on 11 benchmarks, surpassing CoCoOp which introduced instance-wise feature for text prompts. Equipped with TextRefiner, PromptKD achieves state-of-the-art performance while keep inference efficient.

Despite the efficacy of network sparsity in alleviating the deployment strain of Large Language Models (LLMs), it endures significant performance degradation. Applying Low-Rank Adaptation (LoRA) to fine-tune the sparse LLMs offers an intuitive approach to counter this predicament, while it holds shortcomings include: 1) The inability to integrate LoRA weights into sparse LLMs post-training, and 2) Insufficient performance recovery at high sparsity ratios. In this paper, we introduces dynamic $\textbf{Lo}$w-rank $\textbf{S}$parse $\textbf{A}$daptation $\textbf{(LoSA)}$, a novel method that seamlessly integrates low-rank adaptation into LLM sparsity within a unified framework, thereby enhancing the performance of sparse LLMs without increasing the inference latency. In particular, LoSA dynamically sparsifies the LoRA outcomes based on the corresponding sparse weights during fine-tuning, thus guaranteeing that the LoRA module can be integrated into the sparse LLMs post-training. Besides, to achieve the optimal sparse model architecture, LoSA leverages Representation Mutual Information (RMI) as an indicator to determine the importance of layers, thereby dynamically determining the optimal layer-wise sparsity rates during fine-tuning. Predicated on this, LoSA adjusts the rank of the LoRA module based on the variability in layer-wise reconstruction errors, allocating an appropriate fine-tuning for each layer to reduce the output discrepancies between dense and sparse LLMs. Extensive experiments tell that LoSA can efficiently boost the efficacy of sparse LLMs within a few hours, without introducing any additional inferential burden. For example, LoSA reduced the perplexity of sparse LLaMA-2-7B by $\textbf{68.73}$$\downarrow$ and increased zero-shot accuracy by $\textbf{16.32}$%$\uparrow$, achieving a $\textbf{2.60$\times$}$ speedup on CPU and $\textbf{2.23$\times$}$ speedup on GPU, requiring only $\textbf{45 minutes}$ of fine-tuning on $\textbf{a single}$ NVIDIA A100 80GB GPU. Code is available at https://github.com/wzhuang-xmu/LoSA.

In existing multimodal large language models (MLLMs), image resolution plays a significant role for granular visual recognition. However, directly increasing image resolution leads to expensive computational cost for MLLMs. In this paper, we reveal that a combination of low- and high-resolution visual features can efficiently mitigate this shortcoming. Based on this principle, we propose a novel and efficient method for MLLMs, termed Mixture-of-Resolution Adaptation (MRA). In particular, MRA adopts two visual pathways for images of different resolutions, where high-resolution visual information is embedded into the low-resolution pathway via the novel mixture-of-resolution adapters (MR-Adapters). This design also greatly reduces the input sequence length of MLLMs. To validate MRA, we apply it to a recent MLLM called LLaVA, and term the new model LLaVA-HR. We conduct extensive experiments on 17 vision-language (VL) tasks, which show that LLaVA-HR outperforms existing MLLMs on 15 VL tasks, e.g., +5.2\% on TextVQA. More importantly, both training and inference of LLaVA-HR remain efficient with MRA, e.g., 20 training hours and faster inference speed than LLaVA-NeXT. Source codes are released at: https://github.com/luogen1996/LLaVA-HR.

Recent advances in test-time adaptation (TTA) for Vision-Language Models (VLMs) have garnered increasing attention, particularly through the use of multiple augmented views of a single image to boost zero-shot generalization. Unfortunately, existing methods fail to strike a satisfactory balance between performance and efficiency, either due to excessive overhead of tuning text prompts or unstable benefits from handcrafted, training-free visual feature enhancement. In this paper, we present Global-Spatial Bias Learner (GS-Bias), an efficient and effective TTA paradigm that incorporates two learnable biases during TTA, unfolded as the global bias and spatial bias. Particularly, the global bias captures the global semantic features of a test image by learning consistency across augmented views, while spatial bias learns the semantic coherence between regions in the image’s spatial visual representation. It is worth highlighting that these two sets of biases are directly added to the logits outputed by the pretrained VLMs, which circumvent the full backpropagation through VLM that hinders the efficiency of existing TTA methods. This endows GS-Bias with extremely high efficiency while achieving state-of-the-art performance on 15 benchmark datasets. For example, it achieves a 2.23% improvement over TPT in cross-dataset generalization and a 2.72% improvement in domain generalization, while requiring only 6.5% of TPT’s memory usage on ImageNet.

In this paper, we address the challenge of determining the layer-wise sparsity rates of large language models (LLMs) through a theoretical perspective. Specifically, we identify a critical issue of **"reconstruction error explosion"** in existing LLMs sparsification methods. This refers to the cumulative effect of reconstruction errors throughout the sparsification process, where errors from earlier layers propagate and amplify in subsequent layers. As a result, the overall reconstruction error increases significantly, leading to a substantial degradation in model performance. Through theoretical analysis, we derive a simple yet effective approach to layer-wise sparsity allocation that mitigates this issue. Our method uses a monotonically increasing arithmetic progression, reducing the process of determining sparsity rates for multiple layers to the determination of a single common difference hyperparameter. Remarkably, this allows for the optimal layer-wise sparsity rates to be identified with just a few trials. Both our theoretical analysis and experimental results demonstrate that this sparsity allocation scheme is near optimal. Extensive experiments show that our method significantly improves the performance of sparse LLMs across various architectures, outperforming existing layer-wise sparsity methods. Furthermore, it enhances the performance of various compression techniques and is applicable to vision and multimodal models. Notably, our method achieves a reduction of 52.10 in perplexity for the 70% sparse LLaMA2-7B model obtained via Wanda, improves average zero-shot accuracy by 10.50%, and delivers speedups of 2.63$\times$ and 2.23$\times$ on CPU and GPU, respectively. Code is available at https://github.com/wzhuang-xmu/ATP.

N:M sparsity stands as a progressively important tool for DNN compression, achieving practical speedups by stipulating at most N non-zero components within M sequential weights. Unfortunately, most existing works identify the N:M sparse mask through dense backward propagation to update all weights, which incurs exorbitant training costs. In this paper, we introduce BAME, a method that maintains consistent sparsity throughout the N:M sparse training process. BAME perpetually keeps both sparse forward and backward propagation, while iteratively performing weight pruning-and-regrowing within designated weight blocks to tailor the N:M mask. These blocks are selected through a joint assessment based on accumulated mask oscillation frequency and expected loss reduction of mask adaptation, thereby ensuring stable and efficient identification of the optimal N:M mask. Our empirical results substantiate the effectiveness of BAME, illustrating it performs comparably to or better than previous works that fully maintaining dense backward propagation during training. For instance, BAME attains a 72.0% top-1 accuracy while training a 1:16 sparse ResNet-50 on ImageNet, eclipsing SR-STE by 0.5%, despite achieving 2.37 training FLOPs reduction. Code is released at https://github.com/BAME-xmu/BAME

Mixture-of-Experts (MoE) architectures enable efficient scaling of large language models but face prohibitive memory demands due to massive parameterization. Existing pruning methods rely on heuristic metrics or impractical enumeration of expert subsets, leading to suboptimal performance or scalability. In this paper, we propose Shapley-MoE, an efficient pruning method for MoE models inspired by cooperative game theory. By quantifying each expert’s contribution via Shapley value, our method identifies important experts without exhaustive combination evaluations. To overcome the NP-hard complexity of exact Shapley computation, we introduce a Monte Carlo sampling strategy for efficient approximation that reduces complexity to quadratic time. However, vanilla Monte Carlo sampling still faces issues of insufficient estimation accuracy and low sampling efficiency. To address these issues, we further propose two novel methods to improve sampling accuracy and efficiency: (1) Early Truncation, which early terminates unstable sampling steps caused by overly small expert subsets, and (2) Router-Guided Importance Sampling, which prioritize sampling important expert subsets using gating activation probabilities. Both theoretical and experimental analyses show that both methods can accelerate Shapley value estimation and improve accuracy. Extensive empirical evaluations demonstrate that our pruned MoE models outperform existing expert pruning methods. Notably, when applied to the Qwen2-57B-A14B model, our method reduces the number of experts by 25% with only a 0.92 increase in perplexity and over 96.4% of the average zero-shot accuracy is maintained.

Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$\mu$s on a single A100 GPU, with end-to-end throughput up to 3$\times$ higher than vanilla decoding.

This paper presents a Spatial Re-parameterization (SpRe) method for the N:M sparsity. SpRe stems from an observation regarding the restricted variety in spatial sparsity of convolution kernels presented in N:M sparsity compared with unstructured sparsity. Particularly, N:M sparsity exhibits a fixed sparsity rate within the spatial domains due to its distinctive pattern that mandates N non-zero components among M successive weights in the input channel dimension of convolution filters. On the contrary, we observe that conventional unstructured sparsity displays a substantial divergence in sparsity across the spatial domains, which we experimentally verify to be very crucial for its robust performance retention compared with N:M sparsity. Therefore, SpRe employs the spatial-sparsity distribution of unstructured sparsity by assigning an extra branch in conjunction with the original N:M branch at training time, which allows the N:M sparse network to sustain a similar distribution of spatial sparsity with unstructured sparsity. During inference, the extra branch can be further re-parameterized into the main N:M branch, without exerting any distortion on the sparse pattern or additional computation costs. SpRe has achieved a commendable feat by matching the performance of N:M sparsity methods with state-of-the-art unstructured sparsity methods across various benchmarks. Our project is available at https://github.com/zyxxmu/SpRE.

Post-training quantization (PTQ) for vision transformers (ViTs) has received increasing attention from both academic and industrial communities due to its minimal data needs and high time efficiency. However, many current methods fail to account for the complex interactions between quantized weights and activations, resulting in significant quantization errors and suboptimal performance. This paper presents ERQ, an innovative two-step PTQ method specifically crafted to reduce quantization errors arising from activation and weight quantization sequentially. The first step, Activation quantization error reduction (Aqer), first applies Reparameterization Initialization aimed at mitigating initial quantization errors in high-variance activations. Then, it further mitigates the errors by formulating a Ridge Regression problem, which updates the weights maintained at full-precision using a closed-form solution. The second step, Weight quantization error reduction (Wqer), first applies Dual Uniform Quantization to handle weights with numerous outliers, which arise from adjustments made during Reparameterization Initialization, thereby reducing initial weight quantization errors. Then, it employs an iterative approach to further tackle the errors. In each iteration, it adopts Rounding Refinement that uses an empirically derived, efficient proxy to refine the rounding directions of quantized weights, complemented by a Ridge Regression solver to reduce the errors. Comprehensive experimental results demonstrate ERQ's superior performance across various ViTs variants and tasks. For example, ERQ surpasses the state-of-the-art GPTQ by a notable 36.81% in accuracy for W3A4 ViT-S.

This paper focuses on addressing the issue of image demoiréing. Unlike the large volume of existing studies that rely on learning from paired real data, we attempt to learn a demoiréing model from unpaired real data, i.e., moiré images associated with irrelevant clean images. The proposed method, referred to as Unpaired Demoiréing(UnDeM), synthesizes pseudo moiré images from unpaired datasets, generating pairs with clean images for training demoiréing models. To achieve this, we divide real moiré images into patches and group them in compliance with their moiré complexity. We introduce a novel moiré generation framework to synthesize moiré images with diverse moiré features, resembling real moiré patches, and details akin to real moiré-free images. Additionally, we introduce an adaptive denoise method to eliminate the low-quality pseudo moiré images that adversely impact the learning of demoiréing models. We conduct extensive experiments on the commonly-used FHDMi and UHDM datasets. Results manifest that our UnDeM performs better than existing methods when using existing demoiréing models such as MBCNN and ESDNet-L. Code: https://github.com/zysxmu/UnDeM.

Post-training Sparsity (PTS) is a recently emerged av-enue that chases efficient network sparsity with limited data in need. Existing PTS methods, however, undergo significant performance degradation compared with traditional methods that retrain the sparse networks via the whole dataset, especially at high sparsity ratios. In this paper, we attempt to reconcile this disparity by transposing three cardinal factors that profoundly alter the performance of conventional sparsity into the context of PTS. Our endeav-ors particularly comprise (1) A base-decayed sparsity objective that promotes efficient knowledge transferring from dense network to the sparse counterpart. (2) A reducing-regrowing search algorithm designed to ascertain the op-timal sparsity distribution while circumventing overfitting to the small calibration set in PTS. (3) The employment of dynamic sparse training predicated on the preceding as-pects, aimed at comprehensively optimizing the sparsity structure while ensuring training stability. Our proposed framework, termed UniPTS, is validated to be much superior to existing PTS methods across extensive benchmarks. As an illustration, it amplifies the performance of POT, a recently proposed recipe, from 3.9% to 68.6% when pruning ResNet-50 at 90% sparsity ratio on ImageNet. We re-lease the code of our paper at https://github.com/xjjxmu/UniPTS.

Text-to-image (T2I) generative models have attracted significant attention and found extensive applications within and beyond academic research. For example, the Civitai community, a platform for T2I innovation, currently hosts an impressive array of 74,492 distinct models. However, this diversity presents a formidable challenge in selecting the most appropriate model and parameters, a process that typically requires numerous trials. Drawing inspiration from the tool usage research of large language models (LLMs), we introduce DiffAgent, an LLM agent designed to screen the accurate selection in seconds via API calls. DiffAgent leverages a novel two-stage training framework, SFTA, enabling it to accurately align T2I API responses with user input in accordance with human preferences. To train and evaluate DiffAgent's capabilities, we present DABench, a comprehensive dataset encompassing an extensive range of T2I APIs from the community. Our evaluations reveal that DiffAgent not only excels in identifying the appropriate T2I API but also underscores the effectiveness of the SFTA training framework. Codes are available at https://github.com/OpenGVLab/DiffAgent.

The ever-increasing large language models (LLMs), though opening a potential path for the upcoming artificial general intelligence, sadly drops a daunting obstacle on the way towards their on-device deployment. As one of the most well-established pre-LLMs approaches in reducing model complexity, network pruning appears to lag behind in the era of LLMs, due mostly to its costly fine-tuning (or re-training) necessity under the massive volumes of model parameter and training data. To close this industry-academia gap, we introduce Dynamic Sparse No Training ($\texttt{DSNT}$), a training-free fine-tuning approach that slightly updates sparse LLMs without the expensive backpropagation and any weight updates. Inspired by the Dynamic Sparse Training, $\texttt{DSNT}$ minimizes the reconstruction error between the dense and sparse LLMs, in the fashion of performing iterative weight pruning-and-growing on top of sparse LLMs. To accomplish this purpose, $\texttt{DSNT}$ particularly takes into account the anticipated reduction in reconstruction error for pruning and growing, as well as the variance w.r.t. different input data for growing each weight. This practice can be executed efficiently in linear time since its obviates the need of backpropagation for fine-tuning LLMs. Extensive experiments on LLaMA-V1/V2, Vicuna, and OPT across various benchmarks demonstrate the effectiveness of $\texttt{DSNT}$ in enhancing the performance of sparse LLMs, especially at high sparsity levels. For instance, $\texttt{DSNT}$ is able to outperform the state-of-the-art Wanda by 26.79 perplexity at 70% sparsity with LLaMA-7B. Our paper offers fresh insights into how to fine-tune sparse LLMs in an efficient training-free manner and open new venues to scale the great potential of sparsity to LLMs. Codes are available at https://github.com/zyxxmu/DSnoT.

Despite the exceptional performance of Large Language Models (LLMs), the substantial volume of key-value (KV) pairs cached during inference presents a barrier to their efficient deployment. To ameliorate this, recent works have aimed to selectively eliminate these caches, informed by the attention scores of associated tokens. However, such cache eviction invariably leads to output perturbation, regardless of the token choice. This perturbation escalates with the compression ratio, which can precipitate a marked deterioration in LLM inference performance. This paper introduces Cache Merging (CaM) as a solution to mitigate this challenge. CaM adaptively merges to-be-evicted caches into the remaining ones, employing a novel sampling strategy governed by the prominence of attention scores within discarded locations. In this manner, CaM enables memory-efficient LLMs to preserve critical token information, even obviating the need to maintain their corresponding caches. Extensive experiments utilizing LLaMA, OPT, and GPT-NeoX across various benchmarks corroborate CaM’s proficiency in bolstering the performance of memory-efficient LLMs. Code is released at https://github.com/zyxxmu/cam.

Post-training quantization (PTQ) for vision transformers (ViTs) has garnered significant attention due to its efficiency in compressing models. However, existing methods typically overlook the intricate interdependence between quantized weight and activation, leading to considerable quantization error. In this paper, we propose ERQ, a two-step PTQ approach meticulously crafted to sequentially reduce the quantization error arising from activation and weight quantization. ERQ first introduces Activation quantization error reduction (Aqer) that strategically formulates the minimization of activation quantization error as a Ridge Regression problem, tackling it by updating weights with full-precision. Subsequently, ERQ introduces Weight quantization error reduction (Wqer) that adopts an iterative approach to mitigate the quantization error induced by weight quantization. In each iteration, an empirically derived, efficient proxy is employed to refine the rounding directions of quantized weights, coupled with a Ridge Regression solver to curtail weight quantization error. Experimental results attest to the effectiveness of our approach. Notably, ERQ surpasses the state-of-the-art GPTQ by 22.36% in accuracy for W3A4 ViT-S.

This paper presents the first study to explore the potential of parameter quantization for multimodal large language models to alleviate the significant resource constraint encountered during vision-language instruction tuning. We introduce a Quantization-aware Scale LeArning method based on multimodal Warmup, termed QSLAW. This method is grounded in two key innovations: (1) The learning of group-wise scale factors for quantized LLM weights to mitigate the quantization error arising from activation outliers and achieve more effective vision-language instruction tuning; (2) The implementation of a multimodal warmup that progressively integrates linguistic and multimodal training samples, thereby preventing overfitting of the quantized model to multimodal data while ensuring stable adaptation of multimodal large language models to downstream vision-language tasks. Extensive experiments demonstrate that models quantized by QSLAW perform on par with, or even surpass, their full-precision counterparts, while facilitating up to 1.4 times reduction in VL tuning time and GPU consumption.

CutMix is a vital augmentation strategy that determines the performance and generalization ability of vision transformers (ViTs). However, the inconsistency between the mixed images and the corresponding labels harms its efficacy. Existing CutMix variants tackle this problem by generating more consistent mixed images or more precise mixed labels, but inevitably introduce heavy training overhead or require extra information, undermining ease of use. To this end, we propose an novel and effective Self-Motivated image Mixing method (SMMix), which motivates both image and label enhancement by the model under training itself. Specifically, we propose a max-min attention region mixing approach that enriches the attention-focused objects in the mixed images. Then, we introduce a fine-grained label assignment technique that co-trains the output tokens of mixed images with fine-grained supervision. Moreover, we devise a novel feature consistency constraint to align features from mixed and unmixed images. Due to the subtle designs of the self-motivated paradigm, our SMMix is significant in its smaller training overhead and better performance than other CutMix variants. In particular, SMMix improves the accuracy of DeiT-T/S/B, CaiT-XXS-24/36, and PVT-T/S/M/L by more than +1% on ImageNet-1k. The generalization capability of our method is also demonstrated on downstream tasks and out-of-distribution datasets. Our project is available at https://github.com/ChenMnZ/SMMix.

We focus on addressing the dense backward propagation issue for training efficiency of N:M fine-grained sparsity that preserves at most N out of M consecutive weights and achieves practical speedups supported by the N:M sparse tensor core. Therefore, we present a novel method of Bi-directional Masks (Bi-Mask) with its two central innovations in: 1) Separate sparse masks in the two directions of forward and backward propagation to obtain training acceleration. It disentangles the forward and backward weight sparsity and overcomes the very dense gradient computation. 2) An efficient weight row permutation method to maintain performance. It picks up the permutation candidate with the most eligible N:M weight blocks in the backward to minimize the gradient gap between traditional unidirectional masks and our bi-directional masks. Compared with existing uni-directional scenario that applies a transposable mask and enables backward acceleration, our Bi-Mask is experimentally demonstrated to be more superior in performance. Also, our Bi-Mask performs on par with or even better than methods that fail to achieve backward acceleration. Project of this paper is available at https://github.com/zyxxmu/Bi-Mask.

Though network pruning receives popularity in reducing the complexity of convolutional neural networks (CNNs), it remains an open issue to concurrently maintain model accuracy as well as achieve significant speedups on general CPUs. In this paper, we propose a novel 1×N pruning pattern to break this limitation. In particular, consecutive N output kernels with the same input channel index are grouped into one block, which serves as a basic pruning granularity of our pruning pattern. Our 1×N pattern prunes these blocks considered unimportant. We also provide a workflow of filter rearrangement that first rearranges the weight matrix in the output channel dimension to derive more influential blocks for accuracy improvements and then applies similar rearrangement to the next-layer weights in the input channel dimension to ensure correct convolutional operations. Moreover, the output computation after our 1×N pruning can be realized via a parallelized block-wise vectorized operation, leading to significant speedups on general CPUs. The efficacy of our pruning pattern is proved with experiments on ILSVRC-2012. For example, given the pruning rate of 50% and N=4, our pattern obtains about 3.0% improvements over filter pruning in the top-1 accuracy of MobileNet-V2. Meanwhile, it obtains 56.04ms inference savings on Cortex-A7 CPU over weight pruning. Our project is made available at https://github.com/lmbxmu/1xN.

Network pruning is an effective approach to reduce network complexity with acceptable performance compromise. Existing studies achieve the sparsity of neural networks via time-consuming weight training or complex searching on networks with expanded width, which greatly limits the applications of network pruning. In this paper, we show that high-performing and sparse sub-networks without the involvement of weight training, termed "lottery jackpots", exist in pre-trained models with unexpanded width. Our presented lottery jackpots are traceable through empirical and theoretical outcomes. For example, we obtain a lottery jackpot that has only 10% parameters and still reaches the performance of the original dense VGGNet-19 without any modifications on the pre-trained weights on CIFAR-10. Furthermore, we improve the efficiency for searching lottery jackpots from two perspectives. First, we observe that the sparse masks derived from many existing pruning criteria have a high overlap with the searched mask of our lottery jackpot, among which, the magnitude-based pruning results in the most similar mask with ours. In compliance with this insight, we initialize our sparse mask using the magnitude-based pruning, resulting in at least 3× cost reduction on the lottery jackpot searching while achieving comparable or even better performance. Second, we conduct an in-depth analysis of the searching process for lottery jackpots. Our theoretical result suggests that the decrease in training loss during weight searching can be disturbed by the dependency between weights in modern networks. To mitigate this, we propose a novel short restriction method to restrict change of masks that may have potential negative impacts on the training loss, which leads to a faster convergence and reduced oscillation for searching lottery jackpots. Consequently, our searched lottery jackpot removes 90% weights in ResNet-50, while it easily obtains more than 70% top-1 accuracy using only 5 searching epochs on ImageNet.

This article focuses on filter-level network pruning. A novel pruning method, termed CLR-RNF, is proposed. We first reveal a "long-tail" pruning problem in magnitude-based weight pruning methods and then propose a computation-aware measurement for individual weight importance, followed by a cross-layer ranking (CLR) of weights to identify and remove the bottom-ranked weights. Consequently, the per-layer sparsity makes up the pruned network structure in our filter pruning. Then, we introduce a recommendation-based filter selection scheme where each filter recommends a group of its closest filters. To pick the preserved filters from these recommended groups, we further devise a k -reciprocal nearest filter (RNF) selection scheme where the selected filters fall into the intersection of these recommended groups. Both our pruned network structure and the filter selection are nonlearning processes, which, thus, significantly reduces the pruning complexity and differentiates our method from existing works. We conduct image classification on CIFAR-10 and ImageNet to demonstrate the superiority of our CLR-RNF over the state-of-the-arts. For example, on CIFAR-10, CLR-RNF removes 74.1% FLOPs and 95.0% parameters from VGGNet-16 with even 0.3% accuracy improvements. On ImageNet, it removes 70.2% FLOPs and 64.8% parameters from ResNet-50 with only 1.7% top-five accuracy drops. Our project is available at https://github.com/lmbxmu/CLR-RNF.

Channel pruning has been long studied to compress convolutional neural networks (CNNs), which significantly reduces the overall computation. Prior works implement channel pruning in an unexplainable manner, which tends to reduce the final classification errors while failing to consider the internal influence of each channel. In this article, we conduct channel pruning in a white box. Through deep visualization of feature maps activated by different channels, we observe that different channels have a varying contribution to different categories in image classification. Inspired by this, we choose to preserve channels contributing to most categories. Specifically, to model the contribution of each channel to differentiating categories, we develop a class-wise mask for each channel, implemented in a dynamic training manner with respect to the input image's category. On the basis of the learned class-wise mask, we perform a global voting mechanism to remove channels with less category discrimination. Lastly, a fine-tuning process is conducted to recover the performance of the pruned model. To our best knowledge, it is the first time that CNN interpretability theory is considered to guide channel pruning. Extensive experiments on representative image classification tasks demonstrate the superiority of our White-Box over many state-of-the-arts (SOTAs). For instance, on CIFAR-10, it reduces 65.23% floating point operations per seconds (FLOPs) with even 0.62% accuracy improvement for ResNet-110. On ILSVRC-2012, White-Box achieves a 45.6% FLOP reduction with only a small loss of 0.83% in the top-1 accuracy for ResNet-50. Code is available at https://github.com/zyxxmu/White-Box.

By forcing N out of M consecutive weights to be non-zero, the recent N:M fine-grained network sparsity has received increasing attention with its two attractive advantages over traditional irregular network sparsity methods: 1) Promising performance at a high sparsity. 2) Significant speedups when performed on NVIDIA A100 GPUs. Current implementation on N:M sparsity requires a tedious pre-training phase or computationally heavy from-scratch training. To circumvent these problems, this paper presents an efficient solution for achieving N:M fine-grained sparsity from scratch. Specifically, we first make a re-formulation to convert the N:M fine-grained sparsity into a combinatorial problem, in which, the object falls into choosing the best weight combination among $C_M^N$ candidates. Then, we equip each combination with a learnable importance score, which can be jointly optimized along with its associated weights. Through rigorous proof, we demonstrate that the magnitude of the optimized score well reflects the importance of its corresponding weights combination to the training loss. Therefore, by gradually removing combinations with smaller scores till the best one is left, N:M fine-grained sparsity can be efficiently optimized during the normal training phase without any extra expenditure. Comprehensive experimental results have demonstrated that our proposed method for learning best combination, dubbed as LBC, consistently increases the efficacy of the off-the-shelf N:M methods across varying networks and datasets. Our project is released at https://github.com/zyxxmu/LBC.

Channel pruning is among the predominant approaches to compress deep neural networks. To this end, most existing pruning methods focus on selecting channels (filters) by importance/optimization or regularization based on rule-of-thumb designs, which defects in sub-optimal pruning. In this paper, we propose a new channel pruning method based on artificial bee colony algorithm (ABC), dubbed as ABCPruner, which aims to efficiently find optimal pruned structure, i.e., channel number in each layer, rather than selecting "important" channels as previous works did. To solve the intractably huge combinations of pruned structure for deep networks, we first propose to shrink the combinations where the preserved channels are limited to a specific space, thus the combinations of pruned structure can be significantly reduced. And then, we formulate the search of optimal pruned structure as an optimization problem and integrate the ABC algorithm to solve it in an automatic manner to lessen human interference. ABCPruner has been demonstrated to be more effective, which also enables the fine-tuning to be conducted efficiently in an end-to-end manner. The source codes can be available at https: //github.com/lmbxmu/ABCPruner.