DiffRate : Differentiable Compression Rate for Efficient Vision Transformers

Token compression aims to speed up large-scale vision transformers (e.g. ViTs) by pruning (dropping) or merging tokens. It is an important but challenging task. Although recent advanced approaches achieved great success, they need to carefully handcraft a compression rate (i.e. number of tokens to remove), which is tedious and leads to sub-optimal performance. To tackle this problem, we propose Differentiable Compression Rate (DiffRate), a novel token compression method that has several appealing properties prior arts do not have. First, DiffRate enables propagating the loss function's gradient onto the compression ratio, which is considered as a non-differentiable hyperparameter in previous work. In this case, different layers can automatically learn different compression rates layer-wisely without extra overhead. Second, token pruning and merging can be naturally performed simultaneously in DiffRate, while they were isolated in previous works. Third, extensive experiments demonstrate that DiffRate achieves state-of-the-art performance. For example, by applying the learned layer-wise compression rates to an off-the-shelf ViT-H (MAE) model, we achieve a 40% FLOPs reduction and a 1.5x throughput improvement, with a minor accuracy drop of 0.16% on ImageNet without fine-tuning, even outperforming previous methods with fine-tuning. Codes and models are available at https://github.com/OpenGVLab/DiffRate.

Bi-ViT: Pushing the Limit of Vision Transformer Quantization

Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5x and 56.1x theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet.

MultiQuant: A Novel Multi-Branch Topology Method for Arbitrary Bit-width Network Quantization

Arbitrary bit-width network quantization has received significant attention due to its high adaptability to various bit-width requirements during runtime. However, in this paper, we investigate existing methods and observe a significant accumulation of quantization errors caused by frequent bit-width switching of weights and activations, leading to limited performance. To address this issue, we propose MultiQuant, a novel method that utilizes a multi-branch topology for arbitrary bit-width quantization. MultiQuant duplicates the network body into multiple independent branches and quantizes the weights of each branch to a fixed 2-bit while retaining the input activations in the expected bit-width. This approach maintains the computational cost as the same while avoiding the switching of weight bit-widths, thereby substantially reducing errors in weight quantization. Additionally, we introduce an amortization branch selection strategy to distribute quantization errors caused by activation bit-width switching among branches to enhance performance. Finally, we design an in-place distillation strategy that facilitates guidance between branches to further enhance MultiQuant's performance. Extensive experiments demonstrate that MultiQuant achieves significant performance gains compared to existing arbitrary bit-width quantization methods. Code is at \url{https://github.com/zysxmu/MultiQuant}.

Distribution-Flexible Subset Quantization for Post-Quantizing Super-Resolution Networks

This paper introduces Distribution-Flexible Subset Quantization (DFSQ), a post-training quantization method for super-resolution networks. Our motivation for developing DFSQ is based on the distinctive activation distributions of current super-resolution models, which exhibit significant variance across samples and channels. To address this issue, DFSQ conducts channel-wise normalization of the activations and applies distribution-flexible subset quantization (SQ), wherein the quantization points are selected from a universal set consisting of multi-word additive log-scale values. To expedite the selection of quantization points in SQ, we propose a fast quantization points selection strategy that uses K-means clustering to select the quantization points closest to the centroids. Compared to the common iterative exhaustive search algorithm, our strategy avoids the enumeration of all possible combinations in the universal set, reducing the time complexity from exponential to linear. Consequently, the constraint of time costs on the size of the universal set is greatly relaxed. Extensive evaluations of various super-resolution models show that DFSQ effectively retains performance even without fine-tuning. For example, when quantizing EDSRx2 on the Urban benchmark, DFSQ achieves comparable performance to full-precision counterparts on 6- and 8-bit quantization, and incurs only a 0.1 dB PSNR drop on 4-bit quantization. Code is at \url{https://github.com/zysxmu/DFSQ}

Q-DETR: An Efficient Low-Bit Quantized Detection Transformer

The recent detection transformer (DETR) has advanced object detection, but its application on resource-constrained devices requires massive computation and memory resources. Quantization stands out as a solution by representing the network in low-bit parameters and operations. However, there is a significant performance drop when performing low-bit quantized DETR (Q-DETR) with existing quantization methods. We find that the bottlenecks of Q-DETR come from the query information distortion through our empirical analyses. This paper addresses this problem based on a distribution rectification distillation (DRD). We formulate our DRD as a bi-level optimization problem, which can be derived by generalizing the information bottleneck (IB) principle to the learning of Q-DETR. At the inner level, we conduct a distribution alignment for the queries to maximize the self-information entropy. At the upper level, we introduce a new foreground-aware query matching scheme to effectively transfer the teacher information to distillation-desired features to minimize the conditional information entropy. Extensive experimental results show that our method performs much better than prior arts. For example, the 4-bit Q-DETR can theoretically accelerate DETR with ResNet-50 backbone by 6.6x and achieve 39.4% AP, with only 2.6% performance gaps than its real-valued counterpart on the COCO dataset.

Bi-directional Masks for Efficient N:M Sparse Training

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 uni-directional 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 \url{https://github.com/zyxxmu/Bi-Mask}.

Resilient Binary Neural Network

Binary neural networks (BNNs) have received ever-increasing popularity for their great capability of reducing storage burden as well as quickening inference time. However, there is a severe performance drop compared with real-valued networks, due to its intrinsic frequent weight oscillation during training. In this paper, we introduce a Resilient Binary Neural Network (ReBNN) to mitigate the frequent oscillation for better BNNs' training. We identify that the weight oscillation mainly stems from the non-parametric scaling factor. To address this issue, we propose to parameterize the scaling factor and introduce a weighted reconstruction loss to build an adaptive training objective. For the first time, we show that the weight oscillation is controlled by the balanced parameter attached to the reconstruction loss, which provides a theoretical foundation to parameterize it in back propagation. Based on this, we learn our ReBNN by calculating the balanced parameter based on its maximum magnitude, which can effectively mitigate the weight oscillation with a resilient training process. Extensive experiments are conducted upon various network models, such as ResNet and Faster-RCNN for computer vision, as well as BERT for natural language processing. The results demonstrate the overwhelming performance of our ReBNN over prior arts. For example, our ReBNN achieves 66.9% Top-1 accuracy with ResNet-18 backbone on the ImageNet dataset, surpassing existing state-of-the-arts by a significant margin. Our code is open-sourced at https://github.com/SteveTsui/ReBNN.

Real-Time Image Demoireing on Mobile Devices

Moire patterns appear frequently when taking photos of digital screens, drastically degrading the image quality. Despite the advance of CNNs in image demoireing, existing networks are with heavy design, causing redundant computation burden for mobile devices. In this paper, we launch the first study on accelerating demoireing networks and propose a dynamic demoireing acceleration method (DDA) towards a real-time deployment on mobile devices. Our stimulus stems from a simple-yet-universal fact that moire patterns often unbalancedly distribute across an image. Consequently, excessive computation is wasted upon non-moire areas. Therefore, we reallocate computation costs in proportion to the complexity of image patches. In order to achieve this aim, we measure the complexity of an image patch by designing a novel moire prior that considers both colorfulness and frequency information of moire patterns. Then, we restore image patches with higher-complexity using larger networks and the ones with lower-complexity are assigned with smaller networks to relieve the computation burden. At last, we train all networks in a parameter-shared supernet paradigm to avoid additional parameter burden. Extensive experiments on several benchmarks demonstrate the efficacy of our proposed DDA. In addition, the acceleration evaluated on the VIVO X80 Pro smartphone equipped with a chip of Snapdragon 8 Gen 1 shows that our method can drastically reduce the inference time, leading to a real-time image demoireing on mobile devices. Source codes and models are released at https://github.com/zyxxmu/DDA

Low-Rank Winograd Transformation for 3D Convolutional Neural Networks

This paper focuses on Winograd transformation in 3D convolutional neural networks (CNNs) that are more over-parameterized compared with the 2D version. The over-increasing Winograd parameters not only exacerbate training complexity but also barricade the practical speedups due simply to the volume of element-wise products in the Winograd domain. We attempt to reduce trainable parameters by introducing a low-rank Winograd transformation, a novel training paradigm that decouples the original large tensor into two less storage-required trainable tensors, leading to a significant complexity reduction. Built upon our low-rank Winograd transformation, we take one step ahead by proposing a low-rank oriented sparse granularity that measures column-wise parameter importance. By simply involving the non-zero columns in the element-wise product, our sparse granularity is empowered with the ability to produce a very regular sparse pattern to acquire effectual Winograd speedups. To better understand the efficacy of our method, we perform extensive experiments on 3D CNNs. Results manifest that our low-rank Winograd transformation well outperforms the vanilla Winograd transformation. We also show that our proposed low-rank oriented sparse granularity permits practical Winograd acceleration compared with the vanilla counterpart.

Discriminator-Cooperated Feature Map Distillation for GAN Compression

Despite excellent performance in image generation, Generative Adversarial Networks (GANs) are notorious for its requirements of enormous storage and intensive computation. As an awesome ''performance maker'', knowledge distillation is demonstrated to be particularly efficacious in exploring low-priced GANs. In this paper, we investigate the irreplaceability of teacher discriminator and present an inventive discriminator-cooperated distillation, abbreviated as DCD, towards refining better feature maps from the generator. In contrast to conventional pixel-to-pixel match methods in feature map distillation, our DCD utilizes teacher discriminator as a transformation to drive intermediate results of the student generator to be perceptually close to corresponding outputs of the teacher generator. Furthermore, in order to mitigate mode collapse in GAN compression, we construct a collaborative adversarial training paradigm where the teacher discriminator is from scratch established to co-train with student generator in company with our DCD. Our DCD shows superior results compared with existing GAN compression methods. For instance, after reducing over 40x MACs and 80x parameters of CycleGAN, we well decrease FID metric from 61.53 to 48.24 while the current SoTA method merely has 51.92. This work's source code has been made accessible at https://github.com/poopit/DCD-official.