Multi-view 3D detection based on BEV (bird-eye-view) has recently achieved significant improvements. However, the huge memory consumption of state-of-the-art models makes it hard to deploy them on vehicles, and the non-trivial latency will affect the real-time perception of streaming applications. Despite the wide application of quantization to lighten models, we show in our paper that directly applying quantization in BEV tasks will 1) make the training unstable, and 2) lead to intolerable performance degradation. To solve these issues, our method QD-BEV enables a novel view-guided distillation (VGD) objective, which can stabilize the quantization-aware training (QAT) while enhancing the model performance by leveraging both image features and BEV features. Our experiments show that QD-BEV achieves similar or even better accuracy than previous methods with significant efficiency gains. On the nuScenes datasets, the 4-bit weight and 6-bit activation quantized QD-BEV-Tiny model achieves 37.2% NDS with only 15.8 MB model size, outperforming BevFormer-Tiny by 1.8% with an 8x model compression. On the Small and Base variants, QD-BEV models also perform superbly and achieve 47.9% NDS (28.2 MB) and 50.9% NDS (32.9 MB), respectively.
Diffusion models have achieved great success in synthesizing diverse and high-fidelity images. However, sampling speed and memory constraints remain a major barrier to the practical adoption of diffusion models, since the generation process for these models can be slow due to the need for iterative noise estimation using compute-intensive neural networks. We propose to tackle this problem by compressing the noise estimation network to accelerate the generation process through post-training quantization (PTQ). While existing PTQ approaches have not been able to effectively deal with the changing output distributions of noise estimation networks in diffusion models over multiple time steps, we are able to formulate a PTQ method that is specifically designed to handle the unique multi-timestep structure of diffusion models with a data calibration scheme using data sampled from different time steps. Experimental results show that our proposed method is able to directly quantize full-precision diffusion models into 8-bit or 4-bit models while maintaining comparable performance in a training-free manner, achieving a FID change of at most 1.88. Our approach can also be applied to text-guided image generation, and for the first time we can run stable diffusion in 4-bit weights without losing much perceptual quality, as shown in Figure 5 and Figure 9.
Ensemble learning has gain attention in resent deep learning research as a way to further boost the accuracy and generalizability of deep neural network (DNN) models. Recent ensemble training method explores different training algorithms or settings on multiple sub-models with the same model architecture, which lead to significant burden on memory and computation cost of the ensemble model. Meanwhile, the heurtsically induced diversity may not lead to significant performance gain. We propose a new prespective on exploring the intrinsic diversity within a model architecture to build efficient DNN ensemble. We make an intriguing observation that pruning and quantization, while both leading to efficient model architecture at the cost of small accuracy drop, leads to distinct behavior in the decision boundary. To this end, we propose Heterogeneously Compressed Ensemble (HCE), where we build an efficient ensemble with the pruned and quantized variants from a pretrained DNN model. An diversity-aware training objective is proposed to further boost the performance of the HCE ensemble. Experiemnt result shows that HCE achieves significant improvement in the efficiency-accuracy tradeoff comparing to both traditional DNN ensemble training methods and previous model compression methods.
Mixed-precision quantization has been widely applied on deep neural networks (DNNs) as it leads to significantly better efficiency-accuracy tradeoffs compared to uniform quantization. Meanwhile, determining the exact precision of each layer remains challenging. Previous attempts on bit-level regularization and pruning-based dynamic precision adjustment during training suffer from noisy gradients and unstable convergence. In this work, we propose Continuous Sparsification Quantization (CSQ), a bit-level training method to search for mixed-precision quantization schemes with improved stability. CSQ stabilizes the bit-level mixed-precision training process with a bi-level gradual continuous sparsification on both the bit values of the quantized weights and the bit selection in determining the quantization precision of each layer. The continuous sparsification scheme enables fully-differentiable training without gradient approximation while achieving an exact quantized model in the end.A budget-aware regularization of total model size enables the dynamic growth and pruning of each layer's precision towards a mixed-precision quantization scheme of the desired size. Extensive experiments show CSQ achieves better efficiency-accuracy tradeoff than previous methods on multiple models and datasets.
The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to 1.7%, 1.1% and 0.5% for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization.
With the recent demand of deploying neural network models on mobile and edge devices, it is desired to improve the model's generalizability on unseen testing data, as well as enhance the model's robustness under fixed-point quantization for efficient deployment. Minimizing the training loss, however, provides few guarantees on the generalization and quantization performance. In this work, we fulfill the need of improving generalization and quantization performance simultaneously by theoretically unifying them under the framework of improving the model's robustness against bounded weight perturbation and minimizing the eigenvalues of the Hessian matrix with respect to model weights. We therefore propose HERO, a Hessian-enhanced robust optimization method, to minimize the Hessian eigenvalues through a gradient-based training process, simultaneously improving the generalization and quantization performance. HERO enables up to a 3.8% gain on test accuracy, up to 30% higher accuracy under 80% training label perturbation, and the best post-training quantization accuracy across a wide range of precision, including a >10% accuracy improvement over SGD-trained models for common model architectures on various datasets.
Transformers yield state-of-the-art results across many tasks. However, they still impose huge computational costs during inference. We apply global, structural pruning with latency-aware regularization on all parameters of the Vision Transformer (ViT) model for latency reduction. Furthermore, we analyze the pruned architectures and find interesting regularities in the final weight structure. Our discovered insights lead to a new architecture called NViT (Novel ViT), with a redistribution of where parameters are used. This architecture utilizes parameters more efficiently and enables control of the latency-accuracy trade-off. On ImageNet-1K, we prune the DEIT-Base (Touvron et al., 2021) model to a 2.6x FLOPs reduction, 5.1x parameter reduction, and 1.9x run-time speedup with only 0.07% loss in accuracy. We achieve more than 1% accuracy gain when compressing the base model to the throughput of the Small/Tiny variants. NViT gains 0.1-1.1% accuracy over the hand-designed DEIT family when trained from scratch, while being faster.
We design blackbox transfer-based targeted adversarial attacks for an environment where the attacker's source model and the target blackbox model may have disjoint label spaces and training datasets. This scenario significantly differs from the "standard" blackbox setting, and warrants a unique approach to the attacking process. Our methodology begins with the construction of a class correspondence matrix between the whitebox and blackbox label sets. During the online phase of the attack, we then leverage representations of highly related proxy classes from the whitebox distribution to fool the blackbox model into predicting the desired target class. Our attacks are evaluated in three complex and challenging test environments where the source and target models have varying degrees of conceptual overlap amongst their unique categories. Ultimately, we find that it is indeed possible to construct targeted transfer-based adversarial attacks between models that have non-overlapping label spaces! We also analyze the sensitivity of attack success to properties of the clean data. Finally, we show that our transfer attacks serve as powerful adversarial priors when integrated with query-based methods, markedly boosting query efficiency and adversarial success.
Mixed-precision quantization can potentially achieve the optimal tradeoff between performance and compression rate of deep neural networks, and thus, have been widely investigated. However, it lacks a systematic method to determine the exact quantization scheme. Previous methods either examine only a small manually-designed search space or utilize a cumbersome neural architecture search to explore the vast search space. These approaches cannot lead to an optimal quantization scheme efficiently. This work proposes bit-level sparsity quantization (BSQ) to tackle the mixed-precision quantization from a new angle of inducing bit-level sparsity. We consider each bit of quantized weights as an independent trainable variable and introduce a differentiable bit-sparsity regularizer. BSQ can induce all-zero bits across a group of weight elements and realize the dynamic precision reduction, leading to a mixed-precision quantization scheme of the original model. Our method enables the exploration of the full mixed-precision space with a single gradient-based optimization process, with only one hyperparameter to tradeoff the performance and compression. BSQ achieves both higher accuracy and higher bit reduction on various model architectures on the CIFAR-10 and ImageNet datasets comparing to previous methods.