Inference Acceleration

A diffusion model which is formulated to produce an image using thousands of denoising steps usually suffers from a slow inference speed. Existing acceleration algorithms simplify the sampling by skipping most steps yet exhibit considerable performance degradation. By viewing the generation of diffusion models as a discretized integral process we argue that the quality drop is partly caused by applying an inaccurate integral direction to a timestep interval. To rectify this issue we propose a timestep tuner that helps find a more accurate integral direction for a particular interval at the minimum cost. Specifically at each denoising step we replace the original parameterization by conditioning the network on a new timestep enforcing the sampling distribution towards the real one. Extensive experiments show that our plug-in design can be trained efficiently and boost the inference performance of various state-of-the-art acceleration methods especially when there are few denoising steps. For example when using 10 denoising steps on LSUN Bedroom dataset we improve the FID of DDIM from 9.65 to 6.07 simply by adopting our method for a more appropriate set of timesteps. Code is available at \href https://github.com/THU-LYJ-Lab/time-tuner https://github.com/THU-LYJ-Lab/time-tuner .

Logit knowledge distillation attracts increasing attention due to its practicality in recent studies. However it often suffers inferior performance compared to the feature knowledge distillation. In this paper we argue that existing logit-based methods may be sub-optimal since they only leverage the global logit output that couples multiple semantic knowledge. This may transfer ambiguous knowledge to the student and mislead its learning. To this end we propose a simple but effective method i.e. Scale Decoupled Distillation (SDD) for logit knowledge distillation. SDD decouples the global logit output into multiple local logit outputs and establishes distillation pipelines for them. This helps the student to mine and inherit fine-grained and unambiguous logit knowledge. Moreover the decoupled knowledge can be further divided into consistent and complementary logit knowledge that transfers the semantic information and sample ambiguity respectively. By increasing the weight of complementary parts SDD can guide the student to focus more on ambiguous samples improving its discrimination ability. Extensive experiments on several benchmark datasets demonstrate the effectiveness of SDD for wide teacher-student pairs especially in the fine-grained classification task. Code is available at: \href https://github.com/shicaiwei123/SDD-CVPR2024 https://github.com/shicaiwei123/SDD-CVPR2024

Recently transformer-based methods have achieved state-of-the-art prediction quality on human pose estimation(HPE). Nonetheless most of these top-performing transformer-based models are too computation-consuming and storage-demanding to deploy on edge computing platforms. Those transformer-based models that require fewer resources are prone to under-fitting due to their smaller scale and thus perform notably worse than their larger counterparts. Given this conundrum we introduce SDPose a new self-distillation method for improving the performance of small transformer-based models. To mitigate the problem of under-fitting we design a transformer module named Multi-Cycled Transformer(MCT) based on multiple-cycled forwards to more fully exploit the potential of small model parameters. Further in order to prevent the additional inference compute-consuming brought by MCT we introduce a self-distillation scheme extracting the knowledge from the MCT module to a naive forward model. Specifically on the MSCOCO validation dataset SDPose-T obtains 69.7% mAP with 4.4M parameters and 1.8 GFLOPs. Furthermore SDPose-S-V2 obtains 73.5% mAP on the MSCOCO validation dataset with 6.2M parameters and 4.7 GFLOPs achieving a new state-of-the-art among predominant tiny neural network methods.

Quantization is of significance for compressing the over-parameterized deep neural models and deploying them on resource-limited devices. Fixed-precision quantization suffers from performance drop due to the limited numerical representation ability. Conversely mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. Previous works only focus on determining the optimal bit-width configuration in the first stage efficiently while ignoring the considerable time costs in the second stage. However retraining always consumes hundreds of GPU-hours on the cutting-edge GPUs thus hindering deployment efficiency significantly. In this paper we devise a one-shot training-searching paradigm for mixed-precision model compression. Specifically in the first stage all potential bit-width configurations are coupled and thus optimized simultaneously within a set of shared weights. However our observations reveal a previously unseen and severe bit-width interference phenomenon among highly coupled weights during optimization leading to considerable performance degradation under a high compression ratio. To tackle this problem we first design a bit-width scheduler to dynamically freeze the most turbulent bit-width of layers during training to ensure the rest bit-widths converged properly. Then taking inspiration from information theory we present an information distortion mitigation technique to align the behaviour of the bad-performing bit-widths to the well-performing ones. In the second stage an inference-only greedy search scheme is devised to evaluate the goodness of configurations without introducing any additional training costs. Extensive experiments on three representative models and three datasets demonstrate the effectiveness of the proposed method.

Audiovisual segmentation (AVS) is a challenging task that aims to segment visual objects in videos according to their associated acoustic cues. With multiple sound sources and background disturbances involved establishing robust correspondences between audio and visual contents poses unique challenges due to (1) complex entanglement across sound sources and (2) frequent changes in the occurrence of distinct sound events. Assuming sound events occur independently the multi-source semantic space can be represented as the Cartesian product of single-source sub-spaces. We are motivated to decompose the multi-source audio semantics into single-source semantics for more effective interactions with visual content. We propose a semantic decomposition method based on product quantization where the multi-source semantics can be decomposed and represented by several disentangled and noise-suppressed single-source semantics. Furthermore we introduce a global-to-local quantization mechanism which distills knowledge from stable global (clip-level) features into local (frame-level) ones to handle frequent changes in audio semantics. Extensive experiments demonstrate that our semantically decomposed audio representation significantly improves AVS performance eg +21.2% mIoU on the challenging AVS-Semantic benchmark with ResNet50 backbone.

Adversarial robustness of the neural network is a significant concern when it is applied to security-critical domains. In this situation adversarial distillation is a promising option which aims to distill the robustness of the teacher network to improve the robustness of a small student network. Previous works pretrain the teacher network to make it robust against the adversarial examples aimed at itself. However the adversarial examples are dependent on the parameters of the target network. The fixed teacher network inevitably degrades its robustness against the unseen transferred adversarial examples which target the parameters of the student network in the adversarial distillation process. We propose PeerAiD to make a peer network learn the adversarial examples of the student network instead of adversarial examples aimed at itself. PeerAiD is an adversarial distillation that trains the peer network and the student network simultaneously in order to specialize the peer network for defending the student network. We observe that such peer networks surpass the robustness of the pretrained robust teacher model against adversarial examples aimed at the student network. With this peer network and adversarial distillation PeerAiD achieves significantly higher robustness of the student network with AutoAttack (AA) accuracy by up to 1.66%p and improves the natural accuracy of the student network by up to 4.72%p with ResNet-18 on TinyImageNet dataset. Code is available at https://github.com/jaewonalive/PeerAiD.

Diffusion models generate high-quality images but require dozens of forward passes. We introduce Distribution Matching Distillation (DMD) a procedure to transform a diffusion model into a one-step image generator with minimal impact on image quality. We enforce the one-step image generator match the diffusion model at distribution level by minimizing an approximate KL divergence whose gradient can be expressed as the difference between 2 score functions one of the target distribution and the other of the synthetic distribution being produced by our one-step generator. The score functions are parameterized as two diffusion models trained separately on each distribution. Combined with a simple regression loss matching the large-scale structure of the multi-step diffusion outputs our method outperforms all published few-step diffusion approaches reaching 2.62 FID on ImageNet 64x64 and 11.49 FID on zero-shot COCO-30k comparable to Stable Diffusion but orders of magnitude faster. Utilizing FP16 inference our model can generate images at 20 FPS on modern hardware.

Model quantization is widely used to compress and accelerate deep neural networks. However recent studies have revealed the feasibility of weaponizing model quantization via implanting quantization-conditioned backdoors (QCBs). These special backdoors stay dormant on released full-precision models but will come into effect after standard quantization. Due to the peculiarity of QCBs existing defenses have minor effects on reducing their threats or are even infeasible. In this paper we conduct the first in-depth analysis of QCBs. We reveal that the activation of existing QCBs primarily stems from the nearest rounding operation and is closely related to the norms of neuron-wise truncation errors (i.e. the difference between the continuous fullprecision weights and its quantized version). Motivated by these insights we propose Error-guided Flipped Rounding with Activation Preservation (EFRAP) an effective and practical defense against QCBs. Specifically EFRAP learns a non-nearest rounding strategy with neuron-wise error norm and layer-wise activation preservation guidance flipping the rounding strategies of neurons crucial for backdoor effects but with minimal impact on clean accuracy. Extensive evaluations on benchmark datasets demonstrate that our EFRAP can defeat state-of-the-art QCB attacks under various settings. Code is available here.

Post-training quantization (PTQ) converts a pre-trained full-precision (FP) model into a quantized model in a training-free manner. Determining suitable quantization parameters such as scaling factors and weight rounding is the primary strategy for mitigating the impact of quantization noise (calibration) and restoring the performance of the quantized models. However the existing activation calibration methods have never considered information degradation between pre- (FP) and post-quantized activations. In this study we introduce a well-defined distributional metric from information theory mutual information into PTQ calibration. We aim to calibrate the quantized activations by maximizing the mutual information between the pre- and post-quantized activations. To realize this goal we establish a contrastive learning (CL) framework for the PTQ calibration where the quantization parameters are optimized through a self-supervised proxy task. Specifically by leveraging CL during the PTQ process we can benefit from pulling the positive pairs of quantized and FP activations collected from the same input samples while pushing negative pairs from different samples. Thanks to the ingeniously designed critic function we avoid the unwanted but often-encountered collision solution in CL especially in calibration scenarios where the amount of calibration data is limited. Additionally we provide a theoretical guarantee that minimizing our designed loss is equivalent to maximizing the desired mutual information. Consequently the quantized activations retain more information which ultimately enhances the performance of the quantized network. Experimental results show that our method can effectively serve as an add-on module to existing SoTA PTQ methods.

Quantization for model compression can efficiently reduce the network complexity and storage requirement but the original training data is necessary to remedy the performance loss caused by quantization. The Data-Free Quantization (DFQ) methods have been proposed to handle the absence of original training data with synthetic data. However there are differences between the synthetic and original training data which affects the performance of the quantized network but none of the existing methods considers the differences. In this paper we propose an efficient data-free quantization via pseudo-label filtering which is the first to evaluate the synthetic data before quantization. We design a new metric for evaluating synthetic data using self-entropy which indicates the reliability of synthetic data. The synthetic data can be categorized with the metric into high- and low-reliable datasets for the following training process. Besides the multiple pseudo-labels are designed to label the synthetic data with different reliability which can provide valuable supervision information and avoid misleading training by low-reliable samples. Extensive experiments are implemented on several datasets including CIFAR-10 CIFAR-100 and ImageNet with various models. The experimental results show that our method can perform excellently and outperform existing methods in accuracy.

Contrastive Language-Image Pre-training (CLIP) has become a promising language-supervised visual pre-training framework. This paper aims to distill small CLIP models supervised by a large teacher CLIP model. We propose several distillation strategies including relation feature gradient and contrastive paradigms to examine the effectiveness of CLIP-Knowledge Distillation (KD). We show that a simple feature mimicry with Mean Squared Error loss works surprisingly well. Moreover interactive contrastive learning across teacher and student encoders is also effective in performance improvement. We explain that the success of CLIP-KD can be attributed to maximizing the feature similarity between teacher and student. The unified method is applied to distill several student models trained on CC3M+12M. CLIP-KD improves student CLIP models consistently over zero-shot ImageNet classification and cross-modal retrieval benchmarks. When using ViT-L/14 pretrained on Laion-400M as the teacher CLIP-KD achieves 57.5% and 55.4% zero-shot top-1 ImageNet accuracy over ViT-B/16 and ResNet-50 surpassing the original CLIP without KD by 20.5% and 20.1% margins respectively. Our code is released on https://github.com/winycg/CLIP-KD.

Spiking Neural Networks (SNNs) have been widely praised for their high energy efficiency and immense potential. However comprehensive research that critically contrasts and correlates SNNs with quantized Artificial Neural Networks (ANNs) remains scant often leading to skewed comparisons lacking fairness towards ANNs. This paper introduces a unified perspective illustrating that the time steps in SNNs and quantized bit-widths of activation values present analogous representations. Building on this we present a more pragmatic and rational approach to estimating the energy consumption of SNNs. Diverging from the conventional Synaptic Operations (SynOps) we champion the "Bit Budget" concept. This notion permits an intricate discourse on strategically allocating computational and storage resources between weights activation values and temporal steps under stringent hardware constraints. Guided by the Bit Budget paradigm we discern that pivoting efforts towards spike patterns and weight quantization rather than temporal attributes elicits profound implications for model performance. Utilizing the Bit Budget for holistic design consideration of SNNs elevates model performance across diverse data types encompassing static imagery and neuromorphic datasets. Our revelations bridge the theoretical chasm between SNNs and quantized ANNs and illuminate a pragmatic trajectory for future endeavors in energy-efficient neural computations.

We present 3D Paintbrush a technique for automatically texturing local semantic regions on meshes via text descriptions. Our method is designed to operate directly on meshes producing texture maps which seamlessly integrate into standard graphics pipelines. We opt to simultaneously produce a localization map (to specify the edit region) and a texture map which conforms to it. This approach improves the quality of both the localization and the stylization. To enhance the details and resolution of the textured area we leverage multiple stages of a cascaded diffusion model to supervise our local editing technique with generative priors learned from images at different resolutions. Our technique referred to as Cascaded Score Distillation (CSD) simultaneously distills scores at multiple resolutions in a cascaded fashion enabling control over both the granularity and global understanding of the supervision. We demonstrate the effectiveness of 3D Paintbrush to locally texture different semantic regions on a variety of shapes.