Transformers have demonstrated considerable success across various domains but are constrained by their significant computational and memory requirements. This poses challenges for deployment on resource-constrained devices. Quantization, as an effective model compression method, can significantly reduce the operational time of Transformers on edge devices. Notably, Transformers display more substantial outliers than convolutional neural networks, leading to uneven feature distribution among different channels and tokens. To address this issue, we propose an adaptive outlier correction quantization (AOCQ) method for Transformers, which significantly alleviates the adverse effects of these outliers. AOCQ adjusts the notable discrepancies in channels and tokens across three levels: operator level, framework level, and loss level. We introduce a new operator that equivalently balances the activations across different channels and insert an extra stage to optimize the activation quantization step on the framework level. Additionally, we transfer the imbalanced activations across tokens and channels to the optimization of model weights on the loss level. Based on the theoretical study, our method can reduce the quantization error. The effectiveness of the proposed method is verified on various benchmark models and tasks. Surprisingly, DeiT-Base with 8-bit post-training quantization (PTQ) can achieve 81.57% accuracy with a 0.28 percentage point drop while enjoying 4× faster runtime. Furthermore, the weights of Swin and DeiT on several tasks, including classification and object detection, can be post-quantized to ultra-low 4 bits, with a minimal accuracy loss of 2%, while requiring nearly 8× less memory.

Various attribution methods have been developed to explain deep neural networks (DNNs) by inferring the attribution/importance/contribution score of each input variable to the final output. However, existing attribution methods are often built upon different heuristics. There remains a lack of a unified theoretical understanding of why these methods are effective and how they are related. Furthermore, there is still no universally accepted criterion to compare whether one attribution method is preferable over another. In this paper, we resort to Taylor interactions and for the first time, we discover that fourteen existing attribution methods, which define attributions based on fully different heuristics, actually share the same core mechanism. Specifically, we prove that attribution scores of input variables estimated by the fourteen attribution methods can all be mathematically reformulated as a weighted allocation of two typical types of effects, i.e., independent effects of each input variable and interaction effects between input variables. The essential difference among these attribution methods lies in the weights of allocating different effects. Inspired by these insights, we propose three principles for fairly allocating the effects, which serve as new criteria to evaluate the faithfulness of attribution methods. In summary, this study can be considered as a new unified perspective to revisit fourteen attribution methods, which theoretically clarifies essential similarities and differences among these methods. Besides, the proposed new principles enable people to make a direct and fair comparison among different methods under the unified perspective.

Post-training quantization (PTQ) is an effective compression method to reduce the model size and computational cost. However, quantizing a model into a low-bit one, e.g., lower than 4, is difficult and often results in non-negligible performance degradation. To address this, we investigate the loss landscapes of quantized networks with various bit-widths. We show that the network with more ragged loss surface, is more easily trapped into bad local minima, which mostly appears in low-bit quantization. A deeper analysis indicates, the ragged surface is caused by the injection of excessive quantization noise. To this end, we detach a sharpness term from the loss which reflects the impact of quantization noise. To smooth the rugged loss surface, we propose to limit the sharpness term small and stable during optimization. Instead of directly optimizing the target bit network, we design a self-adapted shrinking scheduler for the bit-width in continuous domain from high bit-width to the target by limiting the increasing sharpness term within a proper range. It can be viewed as iteratively adding small “instant” quantization noise and adjusting the network to eliminate its impact. Widely experiments including classification and detection tasks demonstrate the effectiveness of the Bit-shrinking strategy in PTQ. On the Vision Transformer models, our INT8 and INT6 models drop within 0.5% and 1.5% Top-1 accuracy, respectively. On the traditional CNN networks, our INT4 quantized models drop within 1.3% and 3.5% Top-1 accuracy on ResNet18 and MobileNetV2 without fine-tuning, which achieves the state-of-the-art performance.

Although the impressive performance in visual grounding, the prevailing approaches usually exploit the visual backbone in a passive way, i.e., the visual backbone extracts features with fixed weights without expression-related hints. The passive perception may lead to mismatches (e.g., redundant and missing), limiting further performance improvement. Ideally, the visual backbone should actively extract visual features since the expressions already provide the blueprint of desired visual features. The active perception can take expressions as priors to extract relevant visual features, which can effectively alleviate the mismatches. Inspired by this, we propose an active perception Visual Grounding framework based on Language Adaptive Weights, called VG-LAW. The visual backbone serves as an expression-specific feature extractor through dynamic weights generated for various expressions. Benefiting from the specific and relevant visual features extracted from the language-aware visual backbone, VG-LAW does not require additional modules for cross-modal interaction. Along with a neat multi-task head, VG-LAW can be competent in referring expression comprehension and segmentation jointly. Extensive experiments on four representative datasets, i.e., RefCOCO, RefCOCO+, RefCOCOg, and ReferItGame, validate the effectiveness of the proposed framework and demonstrate state-of-the-art performance.

Knowledge distillation (KD) has shown potential for learning compact models in dense object detection. However, the commonly used softmax-based distillation ignores the absolute classification scores for individual categories. Thus, the optimum of the distillation loss does not necessarily lead to the optimal student classification scores for dense object detectors. This cross-task protocol inconsistency is critical, especially for dense object detectors, since the foreground categories are extremely imbalanced. To address the issue of protocol differences between distillation and classification, we propose a novel distillation method with cross-task consistent protocols, tailored for the dense object detection. For classification distillation, we address the cross-task protocol inconsistency problem by formulating the classification logit maps in both teacher and student models as multiple binary-classification maps and applying a binary-classification distillation loss to each map. For localization distillation, we design an IoU-based Localization Distillation Loss that is free from specific network structures and can be compared with existing localization distillation losses. Our proposed method is simple but effective, and experimental results demonstrate its superiority over existing methods. Code is available at https://github.com/TinyTigerPan/BCKD.

Current 6D pose estimation methods focus on handling objects that are previously trained, which limits their applications in real dynamic world. To this end, we propose a geometry correspondence-based framework, termed GCPose, to estimate 6D pose of arbitrary unseen objects without any re-training. Specifically, the proposed method draws the idea from point cloud registration and resorts to object-agnostic geometry features to establish the 3D-3D correspondences between the object-scene point cloud and object-model point cloud. Then the 6D pose parameters are solved by a least-squares fitting algorithm. Taking the symmetry properties of objects into consideration, we design a symmetry-aware matching loss to facilitate the learning of dense point-wise geometry features and improve the performance considerably. Moreover, we introduce an online training data generation with special data augmentation and normalization to empower the network to learn diverse geometry prior. With training on synthetic objects from ShapeNet, our method outperforms previous approaches for unseen object pose estimation by a large margin on T-LESS, LINEMOD, Occluded-LINEMOD, and TUD-L datasets. Code is available at https://github.com/hikvision-research/GCPose.

The Bourne-again shell (Bash) is a prevalent scripting language for orchestrating shell commands and managing resources in Unix-like environments. It is one of the mainstream shell dialects that is available on most GNU Linux systems. However, the unique syntax and semantics of Bash could easily lead to unintended behaviors if carelessly used. Prior studies primarily focused on improving the reliability of Bash scripts or facilitating writing Bash scripts; there is yet no empirical study on the characteristics of Bash programs written in reality, e.g., frequently used language features, common code smells, and bugs. In this article, we perform a large-scale empirical study of Bash usage, based on analyses over one million open source Bash scripts found in Github repositories. We identify and discuss which features and utilities of Bash are most often used. Using static analysis, we find that Bash scripts are often error-prone, and the error-proneness has a moderately positive correlation with the size of the scripts. We also find that the most common problem areas concern quoting, resource management, command options, permissions, and error handling. We envision that these findings can be beneficial for learning Bash and future research that aims to improve shell and command-line productivity and reliability.

Feature Pyramid Network (FPN) has been a generic feature extractor in computer vision tasks, which utilizes multi-level features to generate discriminative pyramidal representations. However, the way simply using Sum or Concatenate operation on features to integrate multi-scale information is not sufficient to obtain discriminative semantic representations. In this paper, we propose a dynamic feature pyramid network (DyFPN) to merge multi-scale information in both features and weights. DyFPN uses both high-level context features and low-level spatial structural features to obtain dynamic convolution kernel that contains multi-scale information. In this manner, each resolution in the pyramid performs unique and adaptive convolution directly, meanwhile strengthening the information flow. Specially, DyFPN can be regarded as a complementary enhancement to existing feature pyramid networks. We analyze the effective receptive field and attention map of DyFPN. It proves that our method contains more local information and global information compared with merging multi-scale information only on feature level. Benefit from multi-ways of integrating multi-scale information, our method outperforms other existing feature pyramid methods on COCO detection tasks by a large margin.

A thriving trend for domain adaptive segmentation endeavors to generate the high-quality pseudo labels for target domain and retrain the segmentor on them. Under this self-training paradigm, some competitive methods have sought to the latent-space information, which establishes the feature centroids (a.k.a prototypes) of the semantic classes and determines the pseudo label candidates by their distances from these centroids. In this paper, we argue that the latent space contains more information to be exploited thus taking one step further to capitalize on it. Firstly, instead of merely using the source-domain prototypes to determine the target pseudo labels as most of the traditional methods do, we bidirectionally produce the target-domain prototypes to degrade those source features which might be too hard or disturbed for the adaptation. Secondly, existing attempts simply model each category as a single and isotropic prototype while ignoring the variance of the feature distribution, which could lead to the confusion of similar categories. To cope with this issue, we propose to represent each category with multiple and anisotropic prototypes via Gaussian Mixture Model, in order to fit the de facto distribution of source domain and estimate the likelihood of target samples based on the probability density. We apply our method on GTA5->Cityscapes and Synthia->Cityscapes tasks and achieve 61.2% and 62.8% respectively in terms of mean IoU, substantially outperforming other competitive self-training methods. Noticeably, in some categories which severely suffer from the categorical confusion such as "truck" and "bus", our method achieves 56.4% and 68.8% respectively, which further demonstrates the effectiveness of our design. The code and model are available at https://github.com/luyvlei/BiSMAPs.

Solid results from Transformers have made them prevailing architectures in various natural language and vision tasks. As a default component in Transformers, Layer Normalization (LN) normalizes activations within each token to boost the robustness. However, LN requires on-the-fly statistics calculation in inference as well as division and square root operations, leading to inefficiency on hardware. What is more, replacing LN with other hardware-efficient normalization schemes (e.g., Batch Normalization) results in inferior performance, even collapse in training. We find that this dilemma is caused by abnormal behaviors of activation statistics, including large fluctuations over iterations and extreme outliers across layers. To tackle these issues, we propose Unified Normalization (UN), which can speed up the inference by being fused with other linear operations and achieve comparable performance on par with LN. UN strives to boost performance by calibrating the activation and gradient statistics with a tailored fluctuation smoothing strategy. Meanwhile, an adaptive outlier filtration strategy is applied to avoid collapse in training whose effectiveness is theoretically proved and experimentally verified in this paper. We demonstrate that UN can be an efficient drop-in alternative to LN by conducting extensive experiments on language and vision tasks. Besides, we evaluate the efficiency of our method on GPU. Transformers equipped with UN enjoy about 31% inference speedup and nearly 18% memory reduction. Code will be released at https://github.com/hikvision-research/Unified-Normalization.

Vision Transformers (ViTs) yield impressive performance across various vision tasks. However, heavy computation and memory footprint make them inaccessible for edge devices. Previous works apply importance criteria determined independently by each individual component to prune ViTs. Considering that heterogeneous components in ViTs play distinct roles, these approaches lead to suboptimal performance. In this paper, we introduce joint importance, which integrates essential structural-aware interactions between components for the first time, to perform collaborative pruning. Based on the theoretical analysis, we construct a Taylor-based approximation to evaluate the joint importance. This guides pruning toward a more balanced reduction across all components. To further reduce the algorithm complexity, we incorporate the interactions into the optimization function under some mild assumptions. Moreover, the proposed method can be seamlessly applied to various tasks including object detection. Extensive experiments demonstrate the effectiveness of our method. Notably, the proposed approach outperforms the existing state-of-the-art approaches on ImageNet, increasing accuracy by 0.7% over the DeiT-Base baseline while saving 50% FLOPs. On COCO, we are the first to show that 70% FLOPs of FasterRCNN with ViT backbone can be removed with only 0.3% mAP drop. The code is available at https://github.com/hikvision-research/SAViT.

Low bit-width model quantization is highly desirable when deploying a deep neural network on mobile and edge devices. Quantization is an effective way to reduce the model size with low bit-width weight representation. However, the unacceptable accuracy drop hinders the development of this approach. One possible reason for this is that the weights in quantization intervals are directly assigned to the center. At the same time, some quantization applications are limited by the various of different network models. Accordingly, in this paper, we propose Multiple Phase Adaptations (MPA), a framework designed to address these two problems. Firstly, weights in the target interval are assigned to center by gradually spreading the quantization range. During the MPA process, the accuracy drop can be compensated for the unquantized parts. Moreover, as MPA does not introduce hyperparameters that depend on different models or bit-width, the framework can be conveniently applied to various models. Extensive experiments demonstrate that MPA achieves higher accuracy than most existing methods on classification tasks for AlexNet, VGG-16 and ResNet.