Large-scale pre-trained sequence-to-sequence models like BART and T5 achieve state-of-the-art performance on many generative NLP tasks. However, such models pose a great challenge in resource-constrained scenarios owing to their large memory requirements and high latency. To alleviate this issue, we propose to jointly distill and quantize the model, where knowledge is transferred from the full-precision teacher model to the quantized and distilled low-precision student model. Empirical analyses show that, despite the challenging nature of generative tasks, we were able to achieve a 16.5x model footprint compression ratio with little performance drop relative to the full-precision counterparts on multiple summarization and QA datasets. We further pushed the limit of compression ratio to 27.7x and presented the performance-efficiency trade-off for generative tasks using pre-trained models. To the best of our knowledge, this is the first work aiming to effectively distill and quantize sequence-to-sequence pre-trained models for language generation tasks.
When training neural networks with simulated quantization, we observe that quantized weights can, rather unexpectedly, oscillate between two grid-points. The importance of this effect and its impact on quantization-aware training (QAT) are not well-understood or investigated in literature. In this paper, we delve deeper into the phenomenon of weight oscillations and show that it can lead to a significant accuracy degradation due to wrongly estimated batch-normalization statistics during inference and increased noise during training. These effects are particularly pronounced in low-bit ($\leq$ 4-bits) quantization of efficient networks with depth-wise separable layers, such as MobileNets and EfficientNets. In our analysis we investigate several previously proposed QAT algorithms and show that most of these are unable to overcome oscillations. Finally, we propose two novel QAT algorithms to overcome oscillations during training: oscillation dampening and iterative weight freezing. We demonstrate that our algorithms achieve state-of-the-art accuracy for low-bit (3 & 4 bits) weight and activation quantization of efficient architectures, such as MobileNetV2, MobileNetV3, and EfficentNet-lite on ImageNet. Our source code is available at {https://github.com/qualcomm-ai-research/oscillations-qat}.
Modeling latent variables with priors and hyperpriors is an essential problem in variational image compression. Formally, trade-off between rate and distortion is handled well if priors and hyperpriors precisely describe latent variables. Current practices only adopt univariate priors and process each variable individually. However, we find inter-correlations and intra-correlations exist when observing latent variables in a vectorized perspective. These findings reveal visual redundancies to improve rate-distortion performance and parallel processing ability to speed up compression. This encourages us to propose a novel vectorized prior. Specifically, a multivariate Gaussian mixture is proposed with means and covariances to be estimated. Then, a novel probabilistic vector quantization is utilized to effectively approximate means, and remaining covariances are further induced to a unified mixture and solved by cascaded estimation without context models involved. Furthermore, codebooks involved in quantization are extended to multi-codebooks for complexity reduction, which formulates an efficient compression procedure. Extensive experiments on benchmark datasets against state-of-the-art indicate our model has better rate-distortion performance and an impressive $3.18\times$ compression speed up, giving us the ability to perform real-time, high-quality variational image compression in practice. Our source code is publicly available at \url{https://github.com/xiaosu-zhu/McQuic}.
In this paper, we introduce a novel algorithm, $\mathsf{CO}_3$, for communication-efficiency distributed Deep Neural Network (DNN) training. $\mathsf{CO}_3$ is a joint training/communication protocol, which encompasses three processing steps for the network gradients: (i) quantization through floating-point conversion, (ii) lossless compression, and (iii) error correction. These three components are crucial in the implementation of distributed DNN training over rate-constrained links. The interplay of these three steps in processing the DNN gradients is carefully balanced to yield a robust and high-performance scheme. The performance of the proposed scheme is investigated through numerical evaluations over CIFAR-10.
The exponentially large discrete search space in mixed-precision quantization (MPQ) makes it hard to determine the optimal bit-width for each layer. Previous works usually resort to iterative search methods on the training set, which consume hundreds or even thousands of GPU-hours. In this study, we reveal that some unique learnable parameters in quantization, namely the scale factors in the quantizer, can serve as importance indicators of a layer, reflecting the contribution of that layer to the final accuracy at certain bit-widths. These importance indicators naturally perceive the numerical transformation during quantization-aware training, which can precisely provide quantization sensitivity metrics of layers. However, a deep network always contains hundreds of such indicators, and training them one by one would lead to an excessive time cost. To overcome this issue, we propose a joint training scheme that can obtain all indicators at once. It considerably speeds up the indicators training process by parallelizing the original sequential training processes. With these learned importance indicators, we formulate the MPQ search problem as a one-time integer linear programming (ILP) problem. That avoids the iterative search and significantly reduces search time without limiting the bit-width search space. For example, MPQ search on ResNet18 with our indicators takes only 0.06 s, which improves time efficiency exponentially compared to iterative search methods. Also, extensive experiments show our approach can achieve SOTA accuracy on ImageNet for far-ranging models with various constraints (e.g., BitOps, compress rate). Code is available on https://github.com/1hunters/LIMPQ.
Quantization has been applied to multiple domains in Deep Neural Networks (DNNs). We propose Depthwise Quantization (DQ) where $\textit{quantization}$ is applied to a decomposed sub-tensor along the $\textit{feature axis}$ of weak statistical dependence. The feature decomposition leads to an exponential increase in $\textit{representation capacity}$ with a linear increase in memory and parameter cost. In addition, DQ can be directly applied to existing encoder-decoder frameworks without modification of the DNN architecture. We use DQ in the context of Hierarchical Auto-Encoder and train end-to-end on an image feature representation. We provide an analysis on cross-correlation between spatial and channel features and we propose a decomposition of the image feature representation along the channel axis. The improved performance of the depthwise operator is due to the increased representation capacity from implicit feature decoupling. We evaluate DQ on the likelihood estimation task, where it outperforms the previous state-of-the-art on CIFAR-10, ImageNet-32 and ImageNet-64. We progressively train with increasing image size a single hierarchical model that uses 69% less parameters and has a faster convergence than the previous works.
Transformer-based language models have become a key building block for natural language processing. While these models are extremely accurate, they can be too large and computationally intensive to run on standard deployments. A variety of compression methods, including distillation, quantization, structured and unstructured pruning are known to decrease model size and increase inference speed, with low accuracy loss. In this context, this paper's contributions are two-fold. We perform an in-depth study of the accuracy-compression trade-off for unstructured weight pruning of BERT models. We introduce Optimal BERT Surgeon (oBERT), an efficient and accurate weight pruning method based on approximate second-order information, which we show to yield state-of-the-art results in both stages of language tasks: pre-training and fine-tuning. Specifically, oBERT extends existing work on unstructured second-order pruning by allowing for pruning blocks of weights, and by being applicable at the BERT scale. Second, we investigate the impact of this pruning method when compounding compression approaches to obtain highly compressed but accurate models for deployment on edge devices. These models significantly push boundaries of the current state-of-the-art sparse BERT models with respect to all metrics: model size, inference speed and task accuracy. For example, relative to the dense BERT-base, we obtain 10x model size compression (in MB) with < 1% accuracy drop, 10x CPU-inference speedup with < 2% accuracy drop, and 29x CPU-inference speedup with < 7.5% accuracy drop. Our code, fully integrated with Transformers and SparseML, is available at https://github.com/neuralmagic/sparseml/tree/main/research/optimal_BERT_surgeon_oBERT.
Mapping between discrete and continuous distributions is a difficult task and many have had to resort to heuristical approaches. We propose a tessellation-based approach that directly learns quantization boundaries in a continuous space, complete with exact likelihood evaluations. This is done through constructing normalizing flows on convex polytopes parameterized using a simple homeomorphism with an efficient log determinant Jacobian. We explore this approach in two application settings, mapping from discrete to continuous and vice versa. Firstly, a Voronoi dequantization allows automatically learning quantization boundaries in a multidimensional space. The location of boundaries and distances between regions can encode useful structural relations between the quantized discrete values. Secondly, a Voronoi mixture model has near-constant computation cost for likelihood evaluation regardless of the number of mixture components. Empirically, we show improvements over existing methods across a range of structured data modalities.
Recovering a high dynamic range (HDR) image from a single low dynamic range (LDR) image, namely inverse tone mapping (ITM), is challenging due to the lack of information in over- and under-exposed regions. Current methods focus exclusively on training high-performing but computationally inefficient ITM models, which in turn hinder deployment of the ITM models in resource-constrained environments with limited computing power such as edge and mobile device applications. To this end, we propose combining efficient operations of deep neural networks with a novel mixed quantization scheme to construct a well-performing but computationally efficient mixed quantization network (MQN) which can perform single image ITM on mobile platforms. In the ablation studies, we explore the effect of using different attention mechanisms, quantization schemes, and loss functions on the performance of MQN in ITM tasks. In the comparative analyses, ITM models trained using MQN perform on par with the state-of-the-art methods on benchmark datasets. MQN models provide up to 10 times improvement on latency and 25 times improvement on memory consumption.
Recently, post-training quantization (PTQ) has driven much attention to produce efficient neural networks without long-time retraining. Despite its low cost, current PTQ works tend to fail under the extremely low-bit setting. In this study, we pioneeringly confirm that properly incorporating activation quantization into the PTQ reconstruction benefits the final accuracy. To deeply understand the inherent reason, a theoretical framework is established, indicating that the flatness of the optimized low-bit model on calibration and test data is crucial. Based on the conclusion, a simple yet effective approach dubbed as QDROP is proposed, which randomly drops the quantization of activations during PTQ. Extensive experiments on various tasks including computer vision (image classification, object detection) and natural language processing (text classification and question answering) prove its superiority. With QDROP, the limit of PTQ is pushed to the 2-bit activation for the first time and the accuracy boost can be up to 51.49%. Without bells and whistles, QDROP establishes a new state of the art for PTQ. Our code is available at https://github.com/wimh966/QDrop and has been integrated into MQBench (https://github.com/ModelTC/MQBench)
Devices participating in federated learning (FL) typically have heterogeneous communication, computation, and memory resources. However, in synchronous FL, all devices need to finish training by the same deadline dictated by the server. Our results show that training a smaller subset of the neural network (NN) at constrained devices, i.e., dropping neurons/filters as proposed by state of the art, is inefficient, preventing these devices to make an effective contribution to the model. This causes unfairness w.r.t the achievable accuracies of constrained devices, especially in cases with a skewed distribution of class labels across devices. We present a novel FL technique, CoCoFL, which maintains the full NN structure on all devices. To adapt to the devices' heterogeneous resources, CoCoFL freezes and quantizes selected layers, reducing communication, computation, and memory requirements, whereas other layers are still trained in full precision, enabling to reach a high accuracy. Thereby, CoCoFL efficiently utilizes the available resources on devices and allows constrained devices to make a significant contribution to the FL system, increasing fairness among participants (accuracy parity) and significantly improving the final accuracy of the model.
Light-weight super-resolution (SR) models have received considerable attention for their serviceability in mobile devices. Many efforts employ network quantization to compress SR models. However, these methods suffer from severe performance degradation when quantizing the SR models to ultra-low precision (e.g., 2-bit and 3-bit) with the low-cost layer-wise quantizer. In this paper, we identify that the performance drop comes from the contradiction between the layer-wise symmetric quantizer and the highly asymmetric activation distribution in SR models. This discrepancy leads to either a waste on the quantization levels or detail loss in reconstructed images. Therefore, we propose a novel activation quantizer, referred to as Dynamic Dual Trainable Bounds (DDTB), to accommodate the asymmetry of the activations. Specifically, DDTB innovates in: 1) A layer-wise quantizer with trainable upper and lower bounds to tackle the highly asymmetric activations. 2) A dynamic gate controller to adaptively adjust the upper and lower bounds at runtime to overcome the drastically varying activation ranges over different samples.To reduce the extra overhead, the dynamic gate controller is quantized to 2-bit and applied to only part of the SR networks according to the introduced dynamic intensity. Extensive experiments demonstrate that our DDTB exhibits significant performance improvements in ultra-low precision. For example, our DDTB achieves a 0.70dB PSNR increase on Urban100 benchmark when quantizing EDSR to 2-bit and scaling up output images to x4. Code is at \url{https://github.com/zysxmu/DDTB}.
Efficient detection and description of geometric regions in images is a prerequisite in visual systems for localization and mapping. Such systems still rely on traditional hand-crafted methods for efficient generation of lightweight descriptors, a common limitation of the more powerful neural network models that come with high compute and specific hardware requirements. In this paper, we focus on the adaptations required by detection and description neural networks to enable their use in computationally limited platforms such as robots, mobile, and augmented reality devices. To that end, we investigate and adapt network quantization techniques to accelerate inference and enable its use on compute limited platforms. In addition, we revisit common practices in descriptor quantization and propose the use of a binary descriptor normalization layer, enabling the generation of distinctive binary descriptors with a constant number of ones. ZippyPoint, our efficient quantized network with binary descriptors, improves the network runtime speed, the descriptor matching speed, and the 3D model size, by at least an order of magnitude when compared to full-precision counterparts. These improvements come at a minor performance degradation as evaluated on the tasks of homography estimation, visual localization, and map-free visual relocalization. Code and models are available at https://github.com/menelaoskanakis/ZippyPoint.
Vision transformers have recently gained great success on various computer vision tasks; nevertheless, their high model complexity makes it challenging to deploy on resource-constrained devices. Quantization is an effective approach to reduce model complexity, and data-free quantization, which can address data privacy and security concerns during model deployment, has received widespread interest. Unfortunately, all existing methods, such as BN regularization, were designed for convolutional neural networks and cannot be applied to vision transformers with significantly different model architectures. In this paper, we propose PSAQ-ViT, a Patch Similarity Aware data-free Quantization framework for Vision Transformers, to enable the generation of "realistic" samples based on the vision transformer's unique properties for calibrating the quantization parameters. Specifically, we analyze the self-attention module's properties and reveal a general difference (patch similarity) in its processing of Gaussian noise and real images. The above insights guide us to design a relative value metric to optimize the Gaussian noise to approximate the real images, which are then utilized to calibrate the quantization parameters. Extensive experiments and ablation studies are conducted on various benchmarks to validate the effectiveness of PSAQ-ViT, which can even outperform the real-data-driven methods. Code is available at: https://github.com/zkkli/PSAQ-ViT.
For autoregressive (AR) modeling of high-resolution images, vector quantization (VQ) represents an image as a sequence of discrete codes. A short sequence length is important for an AR model to reduce its computational costs to consider long-range interactions of codes. However, we postulate that previous VQ cannot shorten the code sequence and generate high-fidelity images together in terms of the rate-distortion trade-off. In this study, we propose the two-stage framework, which consists of Residual-Quantized VAE (RQ-VAE) and RQ-Transformer, to effectively generate high-resolution images. Given a fixed codebook size, RQ-VAE can precisely approximate a feature map of an image and represent the image as a stacked map of discrete codes. Then, RQ-Transformer learns to predict the quantized feature vector at the next position by predicting the next stack of codes. Thanks to the precise approximation of RQ-VAE, we can represent a 256$\times$256 image as 8$\times$8 resolution of the feature map, and RQ-Transformer can efficiently reduce the computational costs. Consequently, our framework outperforms the existing AR models on various benchmarks of unconditional and conditional image generation. Our approach also has a significantly faster sampling speed than previous AR models to generate high-quality images.
An object detection pipeline comprises a camera that captures the scene and an object detector that processes these images. The quality of the images directly affects the performance of the object detector. Many works nowadays focus either on improving the image quality or improving the object detection models independently, but neglect the importance of joint optimization of the two subsystems. The goal of this paper is to tune the detection throughput and accuracy of existing object detectors in the remote sensing scenario by focusing on optimizing the input images tailored to the object detector. To achieve this, we empirically analyze the influence of two selected camera calibration parameters (camera distortion correction and gamma correction) and five image parameters (quantization, compression, resolution, color model, additional channels) for these applications. For our experiments, we utilize three UAV data sets from different domains and a mixture of large and small state-of-the-art object detector models to provide an extensive evaluation of the influence of the pipeline parameters. Finally, we realize an object detection pipeline prototype on an embedded platform for an UAV and give a best practice recommendation for building object detection pipelines based on our findings. We show that not all parameters have an equal impact on detection accuracy and data throughput, and that by using a suitable compromise between parameters we are able to achieve higher detection accuracy for lightweight object detection models, while keeping the same data throughput.
Learning policies that effectively utilize language instructions in complex, multi-task environments is an important problem in sequential decision-making. While it is possible to condition on the entire language instruction directly, such an approach could suffer from generalization issues. In our work, we propose \emph{Learning Interpretable Skill Abstractions (LISA)}, a hierarchical imitation learning framework that can learn diverse, interpretable primitive behaviors or skills from language-conditioned demonstrations to better generalize to unseen instructions. LISA uses vector quantization to learn discrete skill codes that are highly correlated with language instructions and the behavior of the learned policy. In navigation and robotic manipulation environments, LISA outperforms a strong non-hierarchical Decision Transformer baseline in the low data regime and is able to compose learned skills to solve tasks containing unseen long-range instructions. Our method demonstrates a more natural way to condition on language in sequential decision-making problems and achieve interpretable and controllable behavior with the learned skills.
Cardiovascular disease (CVDs) is one of the universal deadly diseases, and the detection of it in the early stage is a challenging task to tackle. Recently, deep learning and convolutional neural networks have been employed widely for the classification of objects. Moreover, it is promising that lots of networks can be deployed on wearable devices. An increasing number of methods can be used to realize ECG signal classification for the sake of arrhythmia detection. However, the existing neural networks proposed for arrhythmia detection are not hardware-friendly enough due to a remarkable quantity of parameters resulting in memory and power consumption. In this paper, we present a 1-D adaptive loss-aware quantization, achieving a high compression rate that reduces memory consumption by 23.36 times. In order to adapt to our compression method, we need a smaller and simpler network. We propose a 17 layer end-to-end neural network classifier to classify 17 different rhythm classes trained on the MIT-BIH dataset, realizing a classification accuracy of 93.5%, which is higher than most existing methods. Due to the adaptive bitwidth method making important layers get more attention and offered a chance to prune useless parameters, the proposed quantization method avoids accuracy degradation. It even improves the accuracy rate, which is 95.84%, 2.34% higher than before. Our study achieves a 1-D convolutional neural network with high performance and low resources consumption, which is hardware-friendly and illustrates the possibility of deployment on wearable devices to realize a real-time arrhythmia diagnosis.
This paper addresses the unsupervised learning of content-style decomposed representation. We first give a definition of style and then model the content-style representation as a token-level bipartite graph. An unsupervised framework, named Retriever, is proposed to learn such representations. First, a cross-attention module is employed to retrieve permutation invariant (P.I.) information, defined as style, from the input data. Second, a vector quantization (VQ) module is used, together with man-induced constraints, to produce interpretable content tokens. Last, an innovative link attention module serves as the decoder to reconstruct data from the decomposed content and style, with the help of the linking keys. Being modal-agnostic, the proposed Retriever is evaluated in both speech and image domains. The state-of-the-art zero-shot voice conversion performance confirms the disentangling ability of our framework. Top performance is also achieved in the part discovery task for images, verifying the interpretability of our representation. In addition, the vivid part-based style transfer quality demonstrates the potential of Retriever to support various fascinating generative tasks. Project page at https://ydcustc.github.io/retriever-demo/.
Cross-modal retrieval aims to search for data with similar semantic meanings across different content modalities. However, cross-modal retrieval requires huge amounts of storage and retrieval time since it needs to process data in multiple modalities. Existing works focused on learning single-source compact features such as binary hash codes that preserve similarities between different modalities. In this work, we propose a jointly learned deep hashing and quantization network (HQ) for cross-modal retrieval. We simultaneously learn binary hash codes and quantization codes to preserve semantic information in multiple modalities by an end-to-end deep learning architecture. At the retrieval step, binary hashing is used to retrieve a subset of items from the search space, then quantization is used to re-rank the retrieved items. We theoretically and empirically show that this two-stage retrieval approach provides faster retrieval results while preserving accuracy. Experimental results on the NUS-WIDE, MIR-Flickr, and Amazon datasets demonstrate that HQ achieves boosts of more than 7% in precision compared to supervised neural network-based compact coding models.