In this paper, we introduce a new approach for high-quality multi-exposure image fusion (MEF). We show that the fusion weights of an exposure can be encoded into a 1D lookup table (LUT), which takes pixel intensity value as input and produces fusion weight as output. We learn one 1D LUT for each exposure, then all the pixels from different exposures can query 1D LUT of that exposure independently for high-quality and efficient fusion. Specifically, to learn these 1D LUTs, we involve attention mechanism in various dimensions including frame, channel and spatial ones into the MEF task so as to bring us significant quality improvement over the state-of-the-art (SOTA). In addition, we collect a new MEF dataset consisting of 960 samples, 155 of which are manually tuned by professionals as ground-truth for evaluation. Our network is trained by this dataset in an unsupervised manner. Extensive experiments are conducted to demonstrate the effectiveness of all the newly proposed components, and results show that our approach outperforms the SOTA in our and another representative dataset SICE, both qualitatively and quantitatively. Moreover, our 1D LUT approach takes less than 4ms to run a 4K image on a PC GPU. Given its high quality, efficiency and robustness, our method has been shipped into millions of Android mobiles across multiple brands world-wide. Code is available at: https://github.com/Hedlen/MEFLUT.
We present LongLoRA, an efficient fine-tuning approach that extends the context sizes of pre-trained large language models (LLMs), with limited computation cost. Typically, training LLMs with long context sizes is computationally expensive, requiring extensive training hours and GPU resources. For example, training on the context length of 8192 needs 16x computational costs in self-attention layers as that of 2048. In this paper, we speed up the context extension of LLMs in two aspects. On the one hand, although dense global attention is needed during inference, fine-tuning the model can be effectively and efficiently done by sparse local attention. The proposed shifted sparse attention effectively enables context extension, leading to non-trivial computation saving with similar performance to fine-tuning with vanilla attention. Particularly, it can be implemented with only two lines of code in training, while being optional in inference. On the other hand, we revisit the parameter-efficient fine-tuning regime for context expansion. Notably, we find that LoRA for context extension works well under the premise of trainable embedding and normalization. LongLoRA combines this improved LoRA with S^2-Attn. LongLoRA demonstrates strong empirical results on various tasks on Llama2 models from 7B/13B to 70B. LongLoRA extends Llama2 7B from 4k context to 100k, or Llama2 70B to 32k on a single 8x A100 machine. LongLoRA extends models' context while retaining their original architectures, and is compatible with most existing techniques, like Flash-Attention2. In addition, we further conduct supervised fine-tuning with LongLoRA and our long instruction-following LongAlpaca dataset.
Classic Graph Neural Network (GNN) inference approaches, designed for static graphs, are ill-suited for streaming graphs that evolve with time. The dynamism intrinsic to streaming graphs necessitates constant updates, posing unique challenges to acceleration on GPU. We address these challenges based on two key insights: (1) Inside the $k$-hop neighborhood, a significant fraction of the nodes is not impacted by the modified edges when the model uses min or max as aggregation function; (2) When the model weights remain static while the graph structure changes, node embeddings can incrementally evolve over time by computing only the impacted part of the neighborhood. With these insights, we propose a novel method, InkStream, designed for real-time inference with minimal memory access and computation, while ensuring an identical output to conventional methods. InkStream operates on the principle of propagating and fetching data only when necessary. It uses an event-based system to control inter-layer effect propagation and intra-layer incremental updates of node embedding. InkStream is highly extensible and easily configurable by allowing users to create and process customized events. We showcase that less than 10 lines of additional user code are needed to support popular GNN models such as GCN, GraphSAGE, and GIN. Our experiments with three GNN models on four large graphs demonstrate that InkStream accelerates by 2.5-427$\times$ on a CPU cluster and 2.4-343$\times$ on two different GPU clusters while producing identical outputs as GNN model inference on the latest graph snapshot.
The state-of-the-art neural video codecs have outperformed the most sophisticated traditional codecs in terms of RD performance in certain cases. However, utilizing them for practical applications is still challenging for two major reasons. 1) Cross-platform computational errors resulting from floating point operations can lead to inaccurate decoding of the bitstream. 2) The high computational complexity of the encoding and decoding process poses a challenge in achieving real-time performance. In this paper, we propose a real-time cross-platform neural video codec, which is capable of efficiently decoding of 720P video bitstream from other encoding platforms on a consumer-grade GPU. First, to solve the problem of inconsistency of codec caused by the uncertainty of floating point calculations across platforms, we design a calibration transmitting system to guarantee the consistent quantization of entropy parameters between the encoding and decoding stages. The parameters that may have transboundary quantization between encoding and decoding are identified in the encoding stage, and their coordinates will be delivered by auxiliary transmitted bitstream. By doing so, these inconsistent parameters can be processed properly in the decoding stage. Furthermore, to reduce the bitrate of the auxiliary bitstream, we rectify the distribution of entropy parameters using a piecewise Gaussian constraint. Second, to match the computational limitations on the decoding side for real-time video codec, we design a lightweight model. A series of efficiency techniques enable our model to achieve 25 FPS decoding speed on NVIDIA RTX 2080 GPU. Experimental results demonstrate that our model can achieve real-time decoding of 720P videos while encoding on another platform. Furthermore, the real-time model brings up to a maximum of 24.2\% BD-rate improvement from the perspective of PSNR with the anchor H.265.
With the fast growth of parameter size, it becomes increasingly challenging to deploy large generative models as they typically require large GPU memory consumption and massive computation. Unstructured model pruning has been a common approach to reduce both GPU memory footprint and the overall computation while retaining good model accuracy. However, the existing solutions do not provide a highly-efficient support for handling unstructured sparsity on modern GPUs, especially on the highly-structured Tensor Core hardware. Therefore, we propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference with the sophisticated support of unstructured sparsity on high-performance but highly restrictive Tensor Cores. Based on our key observation that the main bottleneck of generative model inference is the several skinny matrix multiplications for which Tensor Cores would be significantly under-utilized due to low computational intensity, we propose a general Load-as-Sparse and Compute-as-Dense methodology for unstructured sparse matrix multiplication. The basic insight is to address the significant memory bandwidth bottleneck while tolerating redundant computations that are not critical for end-to-end performance on Tensor Cores. Based on this, we design an effective software framework for Tensor Core based unstructured SpMM, leveraging on-chip resources for efficient sparse data extraction and computation/memory-access overlapping. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively. At end-to-end framework level on OPT-30B/66B/175B models, for tokens per GPU-second, Flash-LLM achieves up to 3.8x and 3.6x improvement over DeepSpeed and FasterTransformer, respectively, with significantly lower inference cost.
Semantic segmentation is an essential technology for self-driving cars to comprehend their surroundings. Currently, real-time semantic segmentation networks commonly employ either encoder-decoder architecture or two-pathway architecture. Generally speaking, encoder-decoder models tend to be quicker,whereas two-pathway models exhibit higher accuracy. To leverage both strengths, we present the Spatial-Assistant Encoder-Decoder Network (SANet) to fuse the two architectures. In the overall architecture, we uphold the encoder-decoder design while maintaining the feature maps in the middle section of the encoder and utilizing atrous convolution branches for same-resolution feature extraction. Toward the end of the encoder, we integrate the asymmetric pooling pyramid pooling module (APPPM) to optimize the semantic extraction of the feature maps. This module incorporates asymmetric pooling layers that extract features at multiple resolutions. In the decoder, we present a hybrid attention module, SAD, that integrates horizontal and vertical attention to facilitate the combination of various branches. To ascertain the effectiveness of our approach, our SANet model achieved competitive results on the real-time CamVid and cityscape datasets. By employing a single 2080Ti GPU, SANet achieved a 78.4 % mIOU at 65.1 FPS on the Cityscape test dataset and 78.8 % mIOU at 147 FPS on the CamVid test dataset. The training code and model for SANet are available at https://github.com/CuZaoo/SANet-main
Dynamic magnetic resonance imaging (DMRI) is an effective imaging tool for diagnosis tasks that require motion tracking of a certain anatomy. To speed up DMRI acquisition, k-space measurements are commonly undersampled along spatial or spatial-temporal domains. The difficulty of recovering useful information increases with increasing undersampling ratios. Compress sensing was invented for this purpose and has become the most popular method until deep learning (DL) based DMRI reconstruction methods emerged in the past decade. Nevertheless, existing DL networks are still limited in long-range sequential dependency understanding and computational efficiency and are not fully automated. Considering the success of Transformers positional embedding and "swin window" self-attention mechanism in the vision community, especially natural video understanding, we hereby propose a novel architecture named Reconstruction Swin Transformer (RST) for 4D MRI. RST inherits the backbone design of the Video Swin Transformer with a novel reconstruction head introduced to restore pixel-wise intensity. A convolution network called SADXNet is used for rapid initialization of 2D MR frames before RST learning to effectively reduce the model complexity, GPU hardware demand, and training time. Experimental results in the cardiac 4D MR dataset further substantiate the superiority of RST, achieving the lowest RMSE of 0.0286 +/- 0.0199 and 1 - SSIM of 0.0872 +/- 0.0783 on 9 times accelerated validation sequences.
Photorealistic style transfer aims to apply stylization while preserving the realism and structure of input content. However, existing methods often encounter challenges such as color tone distortions, dependency on pair-wise pre-training, inefficiency with high-resolution inputs, and the need for additional constraints in video style transfer tasks. To address these issues, we propose a Universal Photorealistic Style Transfer (UPST) framework that delivers accurate photorealistic style transfer on high-resolution images and videos without relying on pre-training. Our approach incorporates a lightweight StyleNet for per-instance transfer, ensuring color tone accuracy while supporting high-resolution inputs, maintaining rapid processing speeds, and eliminating the need for pretraining. To further enhance photorealism and efficiency, we introduce instance-adaptive optimization, which features an adaptive coefficient to prioritize content image realism and employs early stopping to accelerate network convergence. Additionally, UPST enables seamless video style transfer without additional constraints due to its strong non-color information preservation ability. Experimental results show that UPST consistently produces photorealistic outputs and significantly reduces GPU memory usage, making it an effective and universal solution for various photorealistic style transfer tasks.
In language modeling based music generation, a generated waveform is represented by a sequence of hierarchical token stacks that can be decoded either in an auto-regressive manner or in parallel, depending on the codebook patterns. In particular, flattening the codebooks represents the highest quality decoding strategy, while being notoriously slow. To this end, we propose a novel stack-and-delay style of decoding strategy to improve upon the flat pattern decoding where generation speed is four times faster as opposed to vanilla flat decoding. This brings the inference time close to that of the delay decoding strategy, and allows for faster inference on GPU for small batch sizes. For the same inference efficiency budget as the delay pattern, we show that the proposed approach performs better in objective evaluations, almost closing the gap with the flat pattern in terms of quality. The results are corroborated by subjective evaluations which show that samples generated by the new model are slightly more often preferred to samples generated by the competing model given the same text prompts.
Tabular data is hard to acquire and is subject to missing values. This paper introduces a novel approach for generating and imputing mixed-type (continuous and categorical) tabular data utilizing score-based diffusion and conditional flow matching. In contrast to prior methods that rely on neural networks to learn the score function or the vector field, we adopt XGBoost, a widely used Gradient-Boosted Tree (GBT) technique. To test our method, we build one of the most extensive benchmarks for tabular data generation and imputation, containing 27 diverse datasets and 9 metrics. Through empirical evaluation across the benchmark, we demonstrate that our approach outperforms deep-learning generation methods in data generation tasks and remains competitive in data imputation. Notably, it can be trained in parallel using CPUs without requiring a GPU. Our Python and R code is available at https://github.com/SamsungSAILMontreal/ForestDiffusion.
Standard infinite-width limits of neural networks sacrifice the ability for intermediate layers to learn representations from data. Recent work (A theory of representation learning gives a deep generalisation of kernel methods, Yang et al. 2023) modified the Neural Network Gaussian Process (NNGP) limit of Bayesian neural networks so that representation learning is retained. Furthermore, they found that applying this modified limit to a deep Gaussian process gives a practical learning algorithm which they dubbed the deep kernel machine (DKM). However, they only considered the simplest possible setting: regression in small, fully connected networks with e.g. 10 input features. Here, we introduce convolutional deep kernel machines. This required us to develop a novel inter-domain inducing point approximation, as well as introducing and experimentally assessing a number of techniques not previously seen in DKMs, including analogues to batch normalisation, different likelihoods, and different types of top-layer. The resulting model trains in roughly 77 GPU hours, achieving around 99% test accuracy on MNIST, 72% on CIFAR-100, and 92.7% on CIFAR-10, which is SOTA for kernel methods.
High Performance Computing (HPC) benefits from different improvements during last decades, specially in terms of hardware platforms to provide more processing power while maintaining the power consumption at a reasonable level. The Intelligence Processing Unit (IPU) is a new type of massively parallel processor, designed to speedup parallel computations with huge number of processing cores and on-chip memory components connected with high-speed fabrics. IPUs mainly target machine learning applications, however, due to the architectural differences between GPUs and IPUs, especially significantly less memory capacity on an IPU, methods for reducing model size by sparsification have to be considered. Butterfly factorizations are well-known replacements for fully-connected and convolutional layers. In this paper, we examine how butterfly structures can be implemented on an IPU and study their behavior and performance compared to a GPU. Experimental results indicate that these methods can provide 98.5% compression ratio to decrease the immense need for memory, the IPU implementation can benefit from 1.3x and 1.6x performance improvement for butterfly and pixelated butterfly, respectively. We also reach to 1.62x training time speedup on a real-word dataset such as CIFAR10.
The recent advancements in transformer-based visual trackers have led to significant progress, attributed to their strong modeling capabilities. However, as performance improves, running latency correspondingly increases, presenting a challenge for real-time robotics applications, especially on edge devices with computational constraints. In response to this, we introduce LiteTrack, an efficient transformer-based tracking model optimized for high-speed operations across various devices. It achieves a more favorable trade-off between accuracy and efficiency than the other lightweight trackers. The main innovations of LiteTrack encompass: 1) asynchronous feature extraction and interaction between the template and search region for better feature fushion and cutting redundant computation, and 2) pruning encoder layers from a heavy tracker to refine the balnace between performance and speed. As an example, our fastest variant, LiteTrack-B4, achieves 65.2% AO on the GOT-10k benchmark, surpassing all preceding efficient trackers, while running over 100 fps with ONNX on the Jetson Orin NX edge device. Moreover, our LiteTrack-B9 reaches competitive 72.2% AO on GOT-10k and 82.4% AUC on TrackingNet, and operates at 171 fps on an NVIDIA 2080Ti GPU. The code and demo materials will be available at https://github.com/TsingWei/LiteTrack.
We present ExBluRF, a novel view synthesis method for extreme motion blurred images based on efficient radiance fields optimization. Our approach consists of two main components: 6-DOF camera trajectory-based motion blur formulation and voxel-based radiance fields. From extremely blurred images, we optimize the sharp radiance fields by jointly estimating the camera trajectories that generate the blurry images. In training, multiple rays along the camera trajectory are accumulated to reconstruct single blurry color, which is equivalent to the physical motion blur operation. We minimize the photo-consistency loss on blurred image space and obtain the sharp radiance fields with camera trajectories that explain the blur of all images. The joint optimization on the blurred image space demands painfully increasing computation and resources proportional to the blur size. Our method solves this problem by replacing the MLP-based framework to low-dimensional 6-DOF camera poses and voxel-based radiance fields. Compared with the existing works, our approach restores much sharper 3D scenes from challenging motion blurred views with the order of 10 times less training time and GPU memory consumption.
This paper presents an innovative methodology for improving the robustness and computational efficiency of Spiking Neural Networks (SNNs), a critical component in neuromorphic computing. The proposed approach integrates astrocytes, a type of glial cell prevalent in the human brain, into SNNs, creating astrocyte-augmented networks. To achieve this, we designed and implemented an astrocyte model in two distinct platforms: CPU/GPU and FPGA. Our FPGA implementation notably utilizes Dynamic Function Exchange (DFX) technology, enabling real-time hardware reconfiguration and adaptive model creation based on current operating conditions. The novel approach of leveraging astrocytes significantly improves the fault tolerance of SNNs, thereby enhancing their robustness. Notably, our astrocyte-augmented SNN displays near-zero latency and theoretically infinite throughput, implying exceptional computational efficiency. Through comprehensive comparative analysis with prior works, it's established that our model surpasses others in terms of neuron and synapse count while maintaining an efficient power consumption profile. These results underscore the potential of our methodology in shaping the future of neuromorphic computing, by providing robust and energy-efficient systems.
Apache TVM (Tensor Virtual Machine), an open source machine learning compiler framework designed to optimize computations across various hardware platforms, provides an opportunity to improve the performance of dense matrix factorizations such as LU (Lower Upper) decomposition and Cholesky decomposition on GPUs and AI (Artificial Intelligence) accelerators. In this paper, we propose a new TVM autotuning framework using Bayesian Optimization and use the TVM tensor expression language to implement linear algebra kernels such as LU, Cholesky, and 3mm. We use these scientific computation kernels to evaluate the effectiveness of our methods on a GPU cluster, called Swing, at Argonne National Laboratory. We compare the proposed autotuning framework with the TVM autotuning framework AutoTVM with four tuners and find that our framework outperforms AutoTVM in most cases.
Our knowledge of the organisation of the human brain at the population-level is yet to translate into power to predict functional differences at the individual-level, limiting clinical applications, and casting doubt on the generalisability of inferred mechanisms. It remains unknown whether the difficulty arises from the absence of individuating biological patterns within the brain, or from limited power to access them with the models and compute at our disposal. Here we comprehensively investigate the resolvability of such patterns with data and compute at unprecedented scale. Across 23 810 unique participants from UK Biobank, we systematically evaluate the predictability of 25 individual biological characteristics, from all available combinations of structural and functional neuroimaging data. Over 4526 GPU hours of computation, we train, optimize, and evaluate out-of-sample 700 individual predictive models, including fully-connected feed-forward neural networks of demographic, psychological, serological, chronic disease, and functional connectivity characteristics, and both uni- and multi-modal 3D convolutional neural network models of macro- and micro-structural brain imaging. We find a marked discrepancy between the high predictability of sex (balanced accuracy 99.7%), age (mean absolute error 2.048 years, R2 0.859), and weight (mean absolute error 2.609Kg, R2 0.625), for which we set new state-of-the-art performance, and the surprisingly low predictability of other characteristics. Neither structural nor functional imaging predicted psychology better than the coincidence of chronic disease (p<0.05). Serology predicted chronic disease (p<0.05) and was best predicted by it (p<0.001), followed by structural neuroimaging (p<0.05). Our findings suggest either more informative imaging or more powerful models are needed to decipher individual level characteristics from the human brain.
Shampoo is an online and stochastic optimization algorithm belonging to the AdaGrad family of methods for training neural networks. It constructs a block-diagonal preconditioner where each block consists of a coarse Kronecker product approximation to full-matrix AdaGrad for each parameter of the neural network. In this work, we provide a complete description of the algorithm as well as the performance optimizations that our implementation leverages to train deep networks at-scale in PyTorch. Our implementation enables fast multi-GPU distributed data-parallel training by distributing the memory and computation associated with blocks of each parameter via PyTorch's DTensor data structure and performing an AllGather primitive on the computed search directions at each iteration. This major performance enhancement enables us to achieve at most a 10% performance reduction in per-step wall-clock time compared against standard diagonal-scaling-based adaptive gradient methods. We validate our implementation by performing an ablation study on training ImageNet ResNet50, demonstrating Shampoo's superiority over standard training recipes with minimal hyperparameter tuning.
Diffusion models have revolutionized text-to-image generation with its exceptional quality and creativity. However, its multi-step sampling process is known to be slow, often requiring tens of inference steps to obtain satisfactory results. Previous attempts to improve its sampling speed and reduce computational costs through distillation have been unsuccessful in achieving a functional one-step model. In this paper, we explore a recent method called Rectified Flow, which, thus far, has only been applied to small datasets. The core of Rectified Flow lies in its \emph{reflow} procedure, which straightens the trajectories of probability flows, refines the coupling between noises and images, and facilitates the distillation process with student models. We propose a novel text-conditioned pipeline to turn Stable Diffusion (SD) into an ultra-fast one-step model, in which we find reflow plays a critical role in improving the assignment between noise and images. Leveraging our new pipeline, we create, to the best of our knowledge, the first one-step diffusion-based text-to-image generator with SD-level image quality, achieving an FID (Frechet Inception Distance) of $23.3$ on MS COCO 2017-5k, surpassing the previous state-of-the-art technique, progressive distillation, by a significant margin ($37.2$ $\rightarrow$ $23.3$ in FID). By utilizing an expanded network with 1.7B parameters, we further improve the FID to $22.4$. We call our one-step models \emph{InstaFlow}. On MS COCO 2014-30k, InstaFlow yields an FID of $13.1$ in just $0.09$ second, the best in $\leq 0.1$ second regime, outperforming the recent StyleGAN-T ($13.9$ in $0.1$ second). Notably, the training of InstaFlow only costs 199 A100 GPU days. Codes and pre-trained models are available at \url{github.com/gnobitab/InstaFlow}.
The introduction of robust backbones, such as Vision Transformers, has improved the performance of object tracking algorithms in recent years. However, these state-of-the-art trackers are computationally expensive since they have a large number of model parameters and rely on specialized hardware (e.g., GPU) for faster inference. On the other hand, recent lightweight trackers are fast but are less accurate, especially on large-scale datasets. We propose a lightweight, accurate, and fast tracking algorithm using Mobile Vision Transformers (MobileViT) as the backbone for the first time. We also present a novel approach of fusing the template and search region representations in the MobileViT backbone, thereby generating superior feature encoding for target localization. The experimental results show that our MobileViT-based Tracker, MVT, surpasses the performance of recent lightweight trackers on the large-scale datasets GOT10k and TrackingNet, and with a high inference speed. In addition, our method outperforms the popular DiMP-50 tracker despite having 4.7 times fewer model parameters and running at 2.8 times its speed on a GPU. The tracker code and models are available at https://github.com/goutamyg/MVT