Gongfan Fang

Gongfan Fang, Xinyin Ma, Michael Bi Mi et al.

Structured pruning reduces the computational overhead of deep neural networks by removing redundant sub-structures. However, assessing the relative importance of different sub-structures remains a significant challenge, particularly in advanced vision models featuring novel mechanisms and architectures like self-attention, depth-wise convolutions, or residual connections. These heterogeneous substructures usually exhibit diverged parameter scales, weight distributions, and computational topology, introducing considerable difficulty to importance comparison. To overcome this, we present Isomorphic Pruning, a simple approach that demonstrates effectiveness across a range of network architectures such as Vision Transformers and CNNs, and delivers competitive performance across different model sizes. Isomorphic Pruning originates from an observation that, when evaluated under a pre-defined importance criterion, heterogeneous sub-structures demonstrate significant divergence in their importance distribution, as opposed to isomorphic structures that present similar importance patterns. This inspires us to perform isolated ranking and comparison on different types of sub-structures for more reliable pruning. Our empirical results on ImageNet-1K demonstrate that Isomorphic Pruning surpasses several pruning baselines dedicatedly designed for Transformers or CNNs. For instance, we improve the accuracy of DeiT-Tiny from 74.52% to 77.50% by pruning an off-the-shelf DeiT-Base model. And for ConvNext-Tiny, we enhanced performance from 82.06% to 82.18%, while reducing the number of parameters and memory usage. Code is available at \url{https://github.com/VainF/Isomorphic-Pruning}.

Gongfan Fang, Hongxu Yin, Saurav Muralidharan et al.

Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at https://github.com/NVlabs/MaskLLM.

Gongfan Fang, Xinyin Ma, Mingli Song et al.

Structural pruning enables model acceleration by removing structurally-grouped parameters from neural networks. However, the parameter-grouping patterns vary widely across different models, making architecture-specific pruners, which rely on manually-designed grouping schemes, non-generalizable to new architectures. In this work, we study a highly-challenging yet barely-explored task, any structural pruning, to tackle general structural pruning of arbitrary architecture like CNNs, RNNs, GNNs and Transformers. The most prominent obstacle towards this goal lies in the structural coupling, which not only forces different layers to be pruned simultaneously, but also expects all removed parameters to be consistently unimportant, thereby avoiding structural issues and significant performance degradation after pruning. To address this problem, we propose a general and {fully automatic} method, \emph{Dependency Graph} (DepGraph), to explicitly model the dependency between layers and comprehensively group coupled parameters for pruning. In this work, we extensively evaluate our method on several architectures and tasks, including ResNe(X)t, DenseNet, MobileNet and Vision transformer for images, GAT for graph, DGCNN for 3D point cloud, alongside LSTM for language, and demonstrate that, even with a simple norm-based criterion, the proposed method consistently yields gratifying performances.

Xinyin Ma, Xinchao Wang, Gongfan Fang et al.

Data-free knowledge distillation (DFKD) conducts knowledge distillation via eliminating the dependence of original training data, and has recently achieved impressive results in accelerating pre-trained language models. At the heart of DFKD is to reconstruct a synthetic dataset by inverting the parameters of the uncompressed model. Prior DFKD approaches, however, have largely relied on hand-crafted priors of the target data distribution for the reconstruction, which can be inevitably biased and often incompetent to capture the intrinsic distributions. To address this problem, we propose a prompt-based method, termed as PromptDFD, that allows us to take advantage of learned language priors, which effectively harmonizes the synthetic sentences to be semantically and grammatically correct. Specifically, PromptDFD leverages a pre-trained generative model to provide language priors and introduces a reinforced topic prompter to control data synthesis, making the generated samples thematically relevant and semantically plausible, and thus friendly to downstream tasks. As shown in our experiments, the proposed method substantially improves the synthesis quality and achieves considerable improvements on distillation performance. In some cases, PromptDFD even gives rise to results on par with those from the data-driven knowledge distillation with access to the original training data.

Donglin Xie, Ruonan Yu, Gongfan Fang et al.

In this paper, we explore a new knowledge-amalgamation problem, termed Federated Selective Aggregation (FedSA). The goal of FedSA is to train a student model for a new task with the help of several decentralized teachers, whose pre-training tasks and data are different and agnostic. Our motivation for investigating such a problem setup stems from a recent dilemma of model sharing. Many researchers or institutes have spent enormous resources on training large and competent networks. Due to the privacy, security, or intellectual property issues, they are, however, not able to share their own pre-trained models, even if they wish to contribute to the community. The proposed FedSA offers a solution to this dilemma and makes it one step further since, again, the learned student may specialize in a new task different from all of the teachers. To this end, we proposed a dedicated strategy for handling FedSA. Specifically, our student-training process is driven by a novel saliency-based approach that adaptively selects teachers as the participants and integrates their representative capabilities into the student. To evaluate the effectiveness of FedSA, we conduct experiments on both single-task and multi-task settings. Experimental results demonstrate that FedSA effectively amalgamates knowledge from decentralized models and achieves competitive performance to centralized baselines.

Haofei Zhang, Feng Mao, Mengqi Xue et al.

Knowledge amalgamation (KA) is a novel deep model reusing task aiming to transfer knowledge from several well-trained teachers to a multi-talented and compact student. Currently, most of these approaches are tailored for convolutional neural networks (CNNs). However, there is a tendency that transformers, with a completely different architecture, are starting to challenge the domination of CNNs in many computer vision tasks. Nevertheless, directly applying the previous KA methods to transformers leads to severe performance degradation. In this work, we explore a more effective KA scheme for transformer-based object detection models. Specifically, considering the architecture characteristics of transformers, we propose to dissolve the KA into two aspects: sequence-level amalgamation (SA) and task-level amalgamation (TA). In particular, a hint is generated within the sequence-level amalgamation by concatenating teacher sequences instead of redundantly aggregating them to a fixed-size one as previous KA works. Besides, the student learns heterogeneous detection tasks through soft targets with efficiency in the task-level amalgamation. Extensive experiments on PASCAL VOC and COCO have unfolded that the sequence-level amalgamation significantly boosts the performance of students, while the previous methods impair the students. Moreover, the transformer-based students excel in learning amalgamated knowledge, as they have mastered heterogeneous detection tasks rapidly and achieved superior or at least comparable performance to those of the teachers in their specializations.

Gongfan Fang, Xinyin Ma, Xinchao Wang

Reasoning Language Models, capable of extended chain-of-thought reasoning, have demonstrated remarkable performance on tasks requiring complex logical inference. However, applying elaborate reasoning for all queries often results in substantial computational inefficiencies, particularly when many problems admit straightforward solutions. This motivates an open question: Can LLMs learn when to think? To answer this, we propose Thinkless, a learnable framework that empowers an LLM to adaptively select between short-form and long-form reasoning, based on both task complexity and the model's ability. Thinkless is trained under a reinforcement learning paradigm and employs two control tokens, <short> for concise responses and <think> for detailed reasoning. At the core of our method is a Decoupled Group Relative Policy Optimization (DeGRPO) algorithm, which decomposes the learning objective of hybrid reasoning into two components: (1) a control token loss that governs the selection of the reasoning mode, and (2) a response loss that improves the accuracy of the generated answers. This decoupled formulation enables fine-grained control over the contributions of each objective, stabilizing training and effectively preventing collapse observed in vanilla GRPO. Empirically, on several benchmarks such as Minerva Algebra, MATH-500, and GSM8K, Thinkless is able to reduce the usage of long-chain thinking by 50% - 90%, significantly improving the efficiency of Reasoning Language Models. The code is available at https://github.com/VainF/Thinkless

Zigeng Chen, Gongfan Fang, Xinyin Ma et al.

Current approaches for compressing the Segment Anything Model (SAM) yield commendable results, yet necessitate extensive data to train a new network from scratch. Employing conventional pruning techniques can remarkably reduce data requirements but would suffer from a degradation in performance. To address this challenging trade-off, we introduce SlimSAM, a novel data-efficient SAM compression method that achieves superior performance with extremely less training data. The essence of SlimSAM is encapsulated in the alternate slimming framework which effectively enhances knowledge inheritance under severely limited training data availability and exceptional pruning ratio. Diverging from prior techniques, our framework progressively compresses the model by alternately pruning and distilling distinct, decoupled sub-structures. Disturbed Taylor pruning is also proposed to address the misalignment between the pruning objective and training target, thereby boosting the post-distillation after pruning. SlimSAM yields significant performance improvements while demanding over 10 times less training data than any other existing compression methods. Even when compared to the original SAM, SlimSAM achieves approaching performance while reducing parameter counts to merely 1.4% (9.1M), MACs to 0.8% (23G), and requiring only 0.1% (10k) of the SAM training data. The code is available at http://github.com/czg1225/SlimSAM.

Gongfan Fang, Xinyin Ma, Xinchao Wang

Transformer-based diffusion models have achieved significant advancements across a variety of generative tasks. However, producing high-quality outputs typically necessitates large transformer models, which result in substantial training and inference overhead. In this work, we investigate an alternative approach involving multiple experts for denoising, and introduce Remix-DiT, a novel method designed to enhance output quality at a low cost. The goal of Remix-DiT is to craft N diffusion experts for different denoising timesteps, yet without the need for expensive training of N independent models. To achieve this, Remix-DiT employs K basis models (where K < N) and utilizes learnable mixing coefficients to adaptively craft expert models. This design offers two significant advantages: first, although the total model size is increased, the model produced by the mixing operation shares the same architecture as a plain model, making the overall model as efficient as a standard diffusion transformer. Second, the learnable mixing adaptively allocates model capacity across timesteps, thereby effectively improving generation quality. Experiments conducted on the ImageNet dataset demonstrate that Remix-DiT achieves promising results compared to standard diffusion transformers and other multiple-expert methods. The code is available at https://github.com/VainF/Remix-DiT.

Song Wang, Xiaolu Liu, Lingdong Kong et al.

Self-supervised representation learning for point cloud has demonstrated effectiveness in improving pre-trained model performance across diverse tasks. However, as pre-trained models grow in complexity, fully fine-tuning them for downstream applications demands substantial computational and storage resources. Parameter-efficient fine-tuning (PEFT) methods offer a promising solution to mitigate these resource requirements, yet most current approaches rely on complex adapter and prompt mechanisms that increase tunable parameters. In this paper, we propose PointLoRA, a simple yet effective method that combines low-rank adaptation (LoRA) with multi-scale token selection to efficiently fine-tune point cloud models. Our approach embeds LoRA layers within the most parameter-intensive components of point cloud transformers, reducing the need for tunable parameters while enhancing global feature capture. Additionally, multi-scale token selection extracts critical local information to serve as prompts for downstream fine-tuning, effectively complementing the global context captured by LoRA. The experimental results across various pre-trained models and three challenging public datasets demonstrate that our approach achieves competitive performance with only 3.43% of the trainable parameters, making it highly effective for resource-constrained applications. Source code is available at: https://github.com/songw-zju/PointLoRA.

Zeqing Wang, Gongfan Fang, Xinyin Ma et al.

While diffusion language models (DLMs) offer a promising alternative to autoregressive models (ARs), existing open-source DLMs suffer from high inference latency. This bottleneck is mainly due to the attention's quadratic complexity with respect to context length in computing all query-key pairs. Intuitively, to reduce this complexity, a natural strategy is to restrict attention to sparse patterns that retain only the most relevant connections. Such approaches are well-established in ARs, where attention follows fixed and clearly defined sparse patterns. However, in DLMs, we observe distinct sparsity behaviors: (1) attention patterns vary across heads, (2) attention patterns in each head remain highly similar across denoising steps, and (3) early denoising steps are critical for generation. These findings render sparse attention methods designed for ARs largely incompatible with DLMs, as they fail to capture head-specific structures and risk degrading generation when applied in early denoising steps. To address these challenges, we propose SparseD, a novel sparse attention method for DLMs. Leveraging the observations, SparseD only requires pre-computing head-specific sparse patterns one time, and reuses them across all steps. This prevents recomputing sparse patterns at each denoising step. Meanwhile, SparseD uses full attention in the early steps, then switches to sparse attention later to maintain generation quality. Together, these establish SparseD as a practical and efficient solution for deploying DLMs in long-context applications. Experimental results demonstrate that SparseD achieves lossless acceleration, delivering up to $1.50\times$ speedup over FlashAttention at a 64k context length with 1,024 denoising steps.

Xinyin Ma, Gongfan Fang, Xinchao Wang

Large language models (LLMs) have shown remarkable capabilities in language understanding and generation. However, such impressive capability typically comes with a substantial model size, which presents significant challenges in both the deployment, inference, and training stages. With LLM being a general-purpose task solver, we explore its compression in a task-agnostic manner, which aims to preserve the multi-task solving and language generation ability of the original LLM. One challenge to achieving this is the enormous size of the training corpus of LLM, which makes both data transfer and model post-training over-burdensome. Thus, we tackle the compression of LLMs within the bound of two constraints: being task-agnostic and minimizing the reliance on the original training dataset. Our method, named LLM-Pruner, adopts structural pruning that selectively removes non-critical coupled structures based on gradient information, maximally preserving the majority of the LLM's functionality. To this end, the performance of pruned models can be efficiently recovered through tuning techniques, LoRA, in merely 3 hours, requiring only 50K data. We validate the LLM-Pruner on three LLMs, including LLaMA, Vicuna, and ChatGLM, and demonstrate that the compressed models still exhibit satisfactory capabilities in zero-shot classification and generation. The code is available at: https://github.com/horseee/LLM-Pruner

Gongfan Fang, Xinyin Ma, Xinchao Wang

Generative modeling has recently undergone remarkable advancements, primarily propelled by the transformative implications of Diffusion Probabilistic Models (DPMs). The impressive capability of these models, however, often entails significant computational overhead during both training and inference. To tackle this challenge, we present Diff-Pruning, an efficient compression method tailored for learning lightweight diffusion models from pre-existing ones, without the need for extensive re-training. The essence of Diff-Pruning is encapsulated in a Taylor expansion over pruned timesteps, a process that disregards non-contributory diffusion steps and ensembles informative gradients to identify important weights. Our empirical assessment, undertaken across several datasets highlights two primary benefits of our proposed method: 1) Efficiency: it enables approximately a 50\% reduction in FLOPs at a mere 10\% to 20\% of the original training expenditure; 2) Consistency: the pruned diffusion models inherently preserve generative behavior congruent with their pre-trained models. Code is available at \url{https://github.com/VainF/Diff-Pruning}.

Gongfan Fang, Kanya Mo, Xinchao Wang et al.

Data-free knowledge distillation (DFKD) has recently been attracting increasing attention from research communities, attributed to its capability to compress a model only using synthetic data. Despite the encouraging results achieved, state-of-the-art DFKD methods still suffer from the inefficiency of data synthesis, making the data-free training process extremely time-consuming and thus inapplicable for large-scale tasks. In this work, we introduce an efficacious scheme, termed as FastDFKD, that allows us to accelerate DFKD by a factor of orders of magnitude. At the heart of our approach is a novel strategy to reuse the shared common features in training data so as to synthesize different data instances. Unlike prior methods that optimize a set of data independently, we propose to learn a meta-synthesizer that seeks common features as the initialization for the fast data synthesis. As a result, FastDFKD achieves data synthesis within only a few steps, significantly enhancing the efficiency of data-free training. Experiments over CIFAR, NYUv2, and ImageNet demonstrate that the proposed FastDFKD achieves 10$\times$ and even 100$\times$ acceleration while preserving performances on par with state of the art. Code is available at \url{https://github.com/zju-vipa/Fast-Datafree}.

Gongfan Fang, Yifan Bao, Jie Song et al.

Knowledge distillation~(KD) aims to craft a compact student model that imitates the behavior of a pre-trained teacher in a target domain. Prior KD approaches, despite their gratifying results, have largely relied on the premise that \emph{in-domain} data is available to carry out the knowledge transfer. Such an assumption, unfortunately, in many cases violates the practical setting, since the original training data or even the data domain is often unreachable due to privacy or copyright reasons. In this paper, we attempt to tackle an ambitious task, termed as \emph{out-of-domain} knowledge distillation~(OOD-KD), which allows us to conduct KD using only OOD data that can be readily obtained at a very low cost. Admittedly, OOD-KD is by nature a highly challenging task due to the agnostic domain gap. To this end, we introduce a handy yet surprisingly efficacious approach, dubbed as~\textit{MosaicKD}. The key insight behind MosaicKD lies in that, samples from various domains share common local patterns, even though their global semantic may vary significantly; these shared local patterns, in turn, can be re-assembled analogous to mosaic tiling, to approximate the in-domain data and to further alleviating the domain discrepancy. In MosaicKD, this is achieved through a four-player min-max game, in which a generator, a discriminator, a student network, are collectively trained in an adversarial manner, partially under the guidance of a pre-trained teacher. We validate MosaicKD over {classification and semantic segmentation tasks} across various benchmarks, and demonstrate that it yields results much superior to the state-of-the-art counterparts on OOD data. Our code is available at \url{https://github.com/zju-vipa/MosaicKD}.

Gongfan Fang, Jie Song, Xinchao Wang et al.

Model inversion, whose goal is to recover training data from a pre-trained model, has been recently proved feasible. However, existing inversion methods usually suffer from the mode collapse problem, where the synthesized instances are highly similar to each other and thus show limited effectiveness for downstream tasks, such as knowledge distillation. In this paper, we propose Contrastive Model Inversion~(CMI), where the data diversity is explicitly modeled as an optimizable objective, to alleviate the mode collapse issue. Our main observation is that, under the constraint of the same amount of data, higher data diversity usually indicates stronger instance discrimination. To this end, we introduce in CMI a contrastive learning objective that encourages the synthesizing instances to be distinguishable from the already synthesized ones in previous batches. Experiments of pre-trained models on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CMI not only generates more visually plausible instances than the state of the arts, but also achieves significantly superior performance when the generated data are used for knowledge distillation. Code is available at \url{https://github.com/zju-vipa/DataFree}.

Gongfan Fang, Jie Song, Chengchao Shen et al.

Knowledge Distillation (KD) has made remarkable progress in the last few years and become a popular paradigm for model compression and knowledge transfer. However, almost all existing KD algorithms are data-driven, i.e., relying on a large amount of original training data or alternative data, which is usually unavailable in real-world scenarios. In this paper, we devote ourselves to this challenging problem and propose a novel adversarial distillation mechanism to craft a compact student model without any real-world data. We introduce a model discrepancy to quantificationally measure the difference between student and teacher models and construct an optimizable upper bound. In our work, the student and the teacher jointly act the role of the discriminator to reduce this discrepancy, when a generator adversarially produces some "hard samples" to enlarge it. Extensive experiments demonstrate that the proposed data-free method yields comparable performance to existing data-driven methods. More strikingly, our approach can be directly extended to semantic segmentation, which is more complicated than classification, and our approach achieves state-of-the-art results. Code and pretrained models are available at https://github.com/VainF/Data-Free-Adversarial-Distillation.

Changhao Li, Haoling Li, Mengqi Xue et al.

Structural pruning has emerged as a promising approach for producing more efficient models. Nevertheless, the community suffers from a lack of standardized benchmarks and metrics, leaving the progress in this area not fully comprehended. To fill this gap, we present the first comprehensive benchmark, termed PruningBench, for structural pruning. PruningBench showcases the following three characteristics: 1) PruningBench employs a unified and consistent framework for evaluating the effectiveness of diverse structural pruning techniques; 2) PruningBench systematically evaluates 16 existing pruning methods, encompassing a wide array of models (e.g., CNNs and ViTs) and tasks (e.g., classification and detection); 3) PruningBench provides easily implementable interfaces to facilitate the implementation of future pruning methods, and enables the subsequent researchers to incorporate their work into our leaderboards. We provide an online pruning platform for customizing pruning tasks and reproducing all results in this paper. Leaderboard results can also be available.

Xinyin Ma, Yongliang Shen, Gongfan Fang et al.

Large pre-trained transformer-based language models have achieved impressive results on a wide range of NLP tasks. In the past few years, Knowledge Distillation(KD) has become a popular paradigm to compress a computationally expensive model to a resource-efficient lightweight model. However, most KD algorithms, especially in NLP, rely on the accessibility of the original training dataset, which may be unavailable due to privacy issues. To tackle this problem, we propose a novel two-stage data-free distillation method, named Adversarial self-Supervised Data-Free Distillation (AS-DFD), which is designed for compressing large-scale transformer-based models (e.g., BERT). To avoid text generation in discrete space, we introduce a Plug & Play Embedding Guessing method to craft pseudo embeddings from the teacher's hidden knowledge. Meanwhile, with a self-supervised module to quantify the student's ability, we adapt the difficulty of pseudo embeddings in an adversarial training manner. To the best of our knowledge, our framework is the first data-free distillation framework designed for NLP tasks. We verify the effectiveness of our method on several text classification datasets.

Gongfan Fang, Xinchao Wang, Haofei Zhang et al.

It is an innate ability for humans to imagine something only according to their impression, without having to memorize all the details of what they have seen. In this work, we would like to demonstrate that a trained convolutional neural network also has the capability to "remember" its input images. To achieve this, we propose a simple but powerful framework to establish an {\emph{Impression Space}} upon an off-the-shelf pretrained network. This network is referred to as the {\emph{Template Network}} because its filters will be used as templates to reconstruct images from the impression. In our framework, the impression space and image space are bridged by a layer-wise encoding and iterative decoding process. It turns out that the impression space indeed captures the salient features from images, and it can be directly applied to tasks such as unpaired image translation and image synthesis through impression matching without further network training. Furthermore, the impression naturally constructs a high-level common space for different data. Based on this, we propose a mechanism to model the data relations inside the impression space, which is able to reveal the feature similarity between images. Our code will be released.

Sihui Luo, Xinchao Wang, Gongfan Fang et al.

An increasing number of well-trained deep networks have been released online by researchers and developers, enabling the community to reuse them in a plug-and-play way without accessing the training annotations. However, due to the large number of network variants, such public-available trained models are often of different architectures, each of which being tailored for a specific task or dataset. In this paper, we study a deep-model reusing task, where we are given as input pre-trained networks of heterogeneous architectures specializing in distinct tasks, as teacher models. We aim to learn a multitalented and light-weight student model that is able to grasp the integrated knowledge from all such heterogeneous-structure teachers, again without accessing any human annotation. To this end, we propose a common feature learning scheme, in which the features of all teachers are transformed into a common space and the student is enforced to imitate them all so as to amalgamate the intact knowledge. We test the proposed approach on a list of benchmarks and demonstrate that the learned student is able to achieve very promising performance, superior to those of the teachers in their specialized tasks.