The goal of Online Domain Adaptation for semantic segmentation is to handle unforeseeable domain changes that occur during deployment, like sudden weather events. However, the high computational costs associated with brute-force adaptation make this paradigm unfeasible for real-world applications. In this paper we propose HAMLET, a Hardware-Aware Modular Least Expensive Training framework for real-time domain adaptation. Our approach includes a hardware-aware back-propagation orchestration agent (HAMT) and a dedicated domain-shift detector that enables active control over when and how the model is adapted (LT). Thanks to these advancements, our approach is capable of performing semantic segmentation while simultaneously adapting at more than 29FPS on a single consumer-grade GPU. Our framework's encouraging accuracy and speed trade-off is demonstrated on OnDA and SHIFT benchmarks through experimental results.
Learned cardinality estimation methods have achieved high precision compared to traditional methods. Among learned methods, query-driven approaches have faced the workload drift problem for a long time. Although both data-driven and hybrid methods are proposed to avoid this problem, most of them suffer from high training and estimation costs, limited scalability, instability, and long-tail distribution problems on high-dimensional tables, which seriously affects the practical application of learned cardinality estimators. In this paper, we prove that most of these problems are directly caused by the widely used progressive sampling. We solve this problem by introducing predicate information into the autoregressive model and propose Duet, a stable, efficient, and scalable hybrid method to estimate cardinality directly without sampling or any non-differentiable process, which can not only reduce the inference complexity from $O(n)$ to $O(1)$ compared to Naru and UAE but also achieve higher accuracy on high cardinality and high-dimensional tables. Experimental results show that Duet can achieve all the design goals above and be much more practical. Besides, Duet even has a lower inference cost on CPU than that of most learned methods on GPU.
The curse-of-dimensionality taxes computational resources heavily with exponentially increasing computational cost as the dimension increases. This poses great challenges in solving high-dimensional PDEs, as Richard E. Bellman first pointed out over 60 years ago. While there has been some recent success in solving numerically partial differential equations (PDEs) in high dimensions, such computations are prohibitively expensive, and true scaling of general nonlinear PDEs to high dimensions has never been achieved. We develop a new method of scaling up physics-informed neural networks (PINNs) to solve arbitrary high-dimensional PDEs. The new method, called Stochastic Dimension Gradient Descent (SDGD), decomposes a gradient of PDEs into pieces corresponding to different dimensions and randomly samples a subset of these dimensional pieces in each iteration of training PINNs. We prove theoretically the convergence and other desired properties of the proposed method. We demonstrate in various diverse tests that the proposed method can solve many notoriously hard high-dimensional PDEs, including the Hamilton-Jacobi-Bellman (HJB) and the Schr\"{o}dinger equations in tens of thousands of dimensions very fast on a single GPU using the PINNs mesh-free approach. Notably, we solve nonlinear PDEs with nontrivial, anisotropic, and inseparable solutions in 100,000 effective dimensions in 12 hours on a single GPU using SDGD with PINNs. Since SDGD is a general training methodology of PINNs, it can be applied to any current and future variants of PINNs to scale them up for arbitrary high-dimensional PDEs.
Automatic License Plate Recognition (ALPR) is becoming a popular study area and is applied in many fields such as transportation or smart city. However, there are still several limitations when applying many current methods to practical problems due to the variation in real-world situations such as light changes, unclear License Plate (LP) characters, and image quality. Almost recent ALPR algorithms process on a single frame, which reduces accuracy in case of worse image quality. This paper presents methods to improve license plate recognition accuracy by tracking the license plate in multiple frames. First, the Adaptive License Plate Rotation algorithm is applied to correctly align the detected license plate. Second, we propose a method called Character Time-series Matching to recognize license plate characters from many consequence frames. The proposed method archives high performance in the UFPR-ALPR dataset which is \boldmath$96.7\%$ accuracy in real-time on RTX A5000 GPU card. We also deploy the algorithm for the Vietnamese ALPR system. The accuracy for license plate detection and character recognition are 0.881 and 0.979 $mAP^{test}$@.5 respectively. The source code is available at https://github.com/chequanghuy/Character-Time-series-Matching.git
Semantic segmentation is a common task in autonomous driving to understand the surrounding environment. Driveable Area Segmentation and Lane Detection are particularly important for safe and efficient navigation on the road. However, original semantic segmentation models are computationally expensive and require high-end hardware, which is not feasible for embedded systems in autonomous vehicles. This paper proposes a lightweight model for the driveable area and lane line segmentation. TwinLiteNet is designed cheaply but achieves accurate and efficient segmentation results. We evaluate TwinLiteNet on the BDD100K dataset and compare it with modern models. Experimental results show that our TwinLiteNet performs similarly to existing approaches, requiring significantly fewer computational resources. Specifically, TwinLiteNet achieves a mIoU score of 91.3% for the Drivable Area task and 31.08% IoU for the Lane Detection task with only 0.4 million parameters and achieves 415 FPS on GPU RTX A5000. Furthermore, TwinLiteNet can run in real-time on embedded devices with limited computing power, especially since it achieves 60FPS on Jetson Xavier NX, making it an ideal solution for self-driving vehicles. Code is available: url{https://github.com/chequanghuy/TwinLiteNet}.
With its small size, low cost and all-weather operation, millimeter-wave radar can accurately measure the distance, azimuth and radial velocity of a target compared to other traffic sensors. However, in practice, millimeter-wave radars are plagued by various interferences, leading to a drop in target detection accuracy or even failure to detect targets. This is undesirable in autonomous vehicles and traffic surveillance, as it is likely to threaten human life and cause property damage. Therefore, interference mitigation is of great significance for millimeter-wave radar-based target detection. Currently, the development of deep learning is rapid, but existing deep learning-based interference mitigation models still have great limitations in terms of model size and inference speed. For these reasons, we propose Radar-STDA, a Radar-Spatial Temporal Denoising Autoencoder. Radar-STDA is an efficient nano-level denoising autoencoder that takes into account both spatial and temporal information of range-Doppler maps. Among other methods, it achieves a maximum SINR of 17.08 dB with only 140,000 parameters. It obtains 207.6 FPS on an RTX A4000 GPU and 56.8 FPS on an NVIDIA Jetson AGXXavier respectively when denoising range-Doppler maps for three consecutive frames. Moreover, we release a synthetic data set called Ra-inf for the task, which involves 384,769 range-Doppler maps with various clutters from objects of no interest and receiver noise in realistic scenarios. To the best of our knowledge, Ra-inf is the first synthetic dataset of radar interference. To support the community, our research is open-source via the link \url{https://github.com/GuanRunwei/rd_map_temporal_spatial_denoising_autoencoder}.
In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From 20000+ GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets.
Scaling Transformers to longer sequence lengths has been a major problem in the last several years, promising to improve performance in language modeling and high-resolution image understanding, as well as to unlock new applications in code, audio, and video generation. The attention layer is the main bottleneck in scaling to longer sequences, as its runtime and memory increase quadratically in the sequence length. FlashAttention exploits the asymmetric GPU memory hierarchy to bring significant memory saving (linear instead of quadratic) and runtime speedup (2-4$\times$ compared to optimized baselines), with no approximation. However, FlashAttention is still not nearly as fast as optimized matrix-multiply (GEMM) operations, reaching only 25-40\% of the theoretical maximum FLOPs/s. We observe that the inefficiency is due to suboptimal work partitioning between different thread blocks and warps on the GPU, causing either low-occupancy or unnecessary shared memory reads/writes. We propose FlashAttention-2, with better work partitioning to address these issues. In particular, we (1) tweak the algorithm to reduce the number of non-matmul FLOPs (2) parallelize the attention computation, even for a single head, across different thread blocks to increase occupancy, and (3) within each thread block, distribute the work between warps to reduce communication through shared memory. These yield around 2$\times$ speedup compared to FlashAttention, reaching 50-73\% of the theoretical maximum FLOPs/s on A100 and getting close to the efficiency of GEMM operations. We empirically validate that when used end-to-end to train GPT-style models, FlashAttention-2 reaches training speed of up to 225 TFLOPs/s per A100 GPU (72\% model FLOPs utilization).
In this work, we propose Retentive Network (RetNet) as a foundation architecture for large language models, simultaneously achieving training parallelism, low-cost inference, and good performance. We theoretically derive the connection between recurrence and attention. Then we propose the retention mechanism for sequence modeling, which supports three computation paradigms, i.e., parallel, recurrent, and chunkwise recurrent. Specifically, the parallel representation allows for training parallelism. The recurrent representation enables low-cost $O(1)$ inference, which improves decoding throughput, latency, and GPU memory without sacrificing performance. The chunkwise recurrent representation facilitates efficient long-sequence modeling with linear complexity, where each chunk is encoded parallelly while recurrently summarizing the chunks. Experimental results on language modeling show that RetNet achieves favorable scaling results, parallel training, low-cost deployment, and efficient inference. The intriguing properties make RetNet a strong successor to Transformer for large language models. Code will be available at https://aka.ms/retnet.
We design a novel algorithm for optimal transport by drawing from the entropic optimal transport, mirror descent and conjugate gradients literatures. Our scalable and GPU parallelizable algorithm is able to compute the Wasserstein distance with extreme precision, reaching relative error rates of $10^{-8}$ without numerical stability issues. Empirically, the algorithm converges to high precision solutions more quickly in terms of wall-clock time than a variety of algorithms including log-domain stabilized Sinkhorn's Algorithm. We provide careful ablations with respect to algorithm and problem parameters, and present benchmarking over upsampled MNIST images, comparing to various recent algorithms over high-dimensional problems. The results suggest that our algorithm can be a useful addition to the practitioner's optimal transport toolkit.
We introduce CoTracker, a transformer-based model that tracks a large number of 2D points in long video sequences. Differently from most existing approaches that track points independently, CoTracker tracks them jointly, accounting for their dependencies. We show that joint tracking significantly improves tracking accuracy and robustness, and allows CoTracker to track occluded points and points outside of the camera view. We also introduce several innovations for this class of trackers, including using token proxies that significantly improve memory efficiency and allow CoTracker to track 70k points jointly and simultaneously at inference on a single GPU. CoTracker is an online algorithm that operates causally on short windows. However, it is trained utilizing unrolled windows as a recurrent network, maintaining tracks for long periods of time even when points are occluded or leave the field of view. Quantitatively, CoTracker substantially outperforms prior trackers on standard point-tracking benchmarks.
In out-of-distribution (OOD) generalization tasks, fine-tuning pre-trained models has become a prevalent strategy. Different from most prior work that has focused on advancing learning algorithms, we systematically examined how pre-trained model size, pre-training dataset size, and training strategies impact generalization and uncertainty calibration on downstream tasks. We evaluated 100 models across diverse pre-trained model sizes, \update{five} pre-training datasets, and five data augmentations through extensive experiments on four distribution shift datasets totaling over 120,000 GPU hours. Our results demonstrate the significant impact of pre-trained model selection, with optimal choices substantially improving OOD accuracy over algorithm improvement alone. We find larger models and bigger pre-training data improve OOD performance and calibration, in contrast to some prior studies that found modern deep networks to calibrate worse than classical shallow models. Our work underscores the overlooked importance of pre-trained model selection for out-of-distribution generalization and calibration.
The development of reliable and extensible molecular mechanics (MM) force fields -- fast, empirical models characterizing the potential energy surface of molecular systems -- is indispensable for biomolecular simulation and computer-aided drug design. Here, we introduce a generalized and extensible machine-learned MM force field, \texttt{espaloma-0.3}, and an end-to-end differentiable framework using graph neural networks to overcome the limitations of traditional rule-based methods. Trained in a single GPU-day to fit a large and diverse quantum chemical dataset of over 1.1M energy and force calculations, \texttt{espaloma-0.3} reproduces quantum chemical energetic properties of chemical domains highly relevant to drug discovery, including small molecules, peptides, and nucleic acids. Moreover, this force field maintains the quantum chemical energy-minimized geometries of small molecules and preserves the condensed phase properties of peptides, self-consistently parametrizing proteins and ligands to produce stable simulations leading to highly accurate predictions of binding free energies. This methodology demonstrates significant promise as a path forward for systematically building more accurate force fields that are easily extensible to new chemical domains of interest.
We propose the In-context Autoencoder (ICAE), leveraging the power of a large language model (LLM) to compress a long context into short compact memory slots that can be directly conditioned on by the LLM for various purposes. ICAE is first pretrained using both autoencoding and language modeling objectives on massive text data, enabling it to generate memory slots that accurately and comprehensively represent the original context. Then, it is fine-tuned on instruction data for producing desirable responses to various prompts. Experiments demonstrate that our lightweight ICAE, introducing about 1% additional parameters, effectively achieves $4\times$ context compression based on Llama, offering advantages in both improved latency and GPU memory cost during inference, and showing an interesting insight in memorization as well as potential for scalability. These promising results imply a novel perspective on the connection between working memory in cognitive science and representation learning in LLMs, revealing ICAE's significant implications in addressing the long context problem and suggesting further research in LLM context management. Our data, code and models are available at https://github.com/getao/icae.
Despite the dominance and effectiveness of scaling, resulting in large networks with hundreds of billions of parameters, the necessity to train overparameterized models remains poorly understood, while training costs grow exponentially. In this paper, we explore parameter-efficient training techniques as an approach to training large neural networks. We introduce a novel method called ReLoRA, which utilizes low-rank updates to train high-rank networks. We apply ReLoRA to training transformer language models with up to 1.3B parameters and demonstrate comparable performance to regular neural network training. ReLoRA saves up to 5.5Gb of RAM per GPU and improves training speed by 9-40% depending on the model size and hardware setup. Our findings show the potential of parameter-efficient techniques for large-scale pre-training.
Data-driven deep learning has been successfully applied to various computed tomographic reconstruction problems. The deep inference models may outperform existing analytical and iterative algorithms, especially in ill-posed CT reconstruction. However, those methods often predict images that do not agree with the measured projection data. This paper presents an accurate differentiable forward and back projection software library to ensure the consistency between the predicted images and the original measurements. The software library efficiently supports various projection geometry types while minimizing the GPU memory footprint requirement, which facilitates seamless integration with existing deep learning training and inference pipelines. The proposed software is available as open source: https://github.com/LLNL/LEAP.
Graph neural networks are widely used in machine learning applied to chemistry, and in particular for material science discovery. For crystalline materials, however, generating graph-based representation from geometrical information for neural networks is not a trivial task. The periodicity of crystalline needs efficient implementations to be processed in real-time under a massively parallel environment. With the aim of training graph-based generative models of new material discovery, we propose an efficient tool to generate cutoff graphs and k-nearest-neighbours graphs of periodic structures within GPU optimization. We provide pyMatGraph a Pytorch-compatible framework to generate graphs in real-time during the training of neural network architecture. Our tool can update a graph of a structure, making generative models able to update the geometry and process the updated graph during the forward propagation on the GPU side. Our code is publicly available at https://github.com/aklipf/mat-graph.
Recent work has explored Large Language Models (LLMs) to overcome the lack of training data for Information Retrieval (IR) tasks. The generalization abilities of these models have enabled the creation of synthetic in-domain data by providing instructions and a few examples on a prompt. InPars and Promptagator have pioneered this approach and both methods have demonstrated the potential of using LLMs as synthetic data generators for IR tasks. This makes them an attractive solution for IR tasks that suffer from a lack of annotated data. However, the reproducibility of these methods was limited, because InPars' training scripts are based on TPUs -- which are not widely accessible -- and because the code for Promptagator was not released and its proprietary LLM is not publicly accessible. To fully realize the potential of these methods and make their impact more widespread in the research community, the resources need to be accessible and easy to reproduce by researchers and practitioners. Our main contribution is a unified toolkit for end-to-end reproducible synthetic data generation research, which includes generation, filtering, training and evaluation. Additionally, we provide an interface to IR libraries widely used by the community and support for GPU. Our toolkit not only reproduces the InPars method and partially reproduces Promptagator, but also provides a plug-and-play functionality allowing the use of different LLMs, exploring filtering methods and finetuning various reranker models on the generated data. We also made available all the synthetic data generated in this work for the 18 different datasets in the BEIR benchmark which took more than 2,000 GPU hours to be generated as well as the reranker models finetuned on the synthetic data. Code and data are available at https://github.com/zetaalphavector/InPars
Robotic navigation in unknown, cluttered environments with limited sensing capabilities poses significant challenges in robotics. Local trajectory optimization methods, such as Model Predictive Path Intergal (MPPI), are a promising solution to this challenge. However, global guidance is required to ensure effective navigation, especially when encountering challenging environmental conditions or navigating beyond the planning horizon. This study presents the GP-MPPI, an online learning-based control strategy that integrates MPPI with a local perception model based on Sparse Gaussian Process (SGP). The key idea is to leverage the learning capability of SGP to construct a variance (uncertainty) surface, which enables the robot to learn about the navigable space surrounding it, identify a set of suggested subgoals, and ultimately recommend the optimal subgoal that minimizes a predefined cost function to the local MPPI planner. Afterward, MPPI computes the optimal control sequence that satisfies the robot and collision avoidance constraints. Such an approach eliminates the necessity of a global map of the environment or an offline training process. We validate the efficiency and robustness of our proposed control strategy through both simulated and real-world experiments of 2D autonomous navigation tasks in complex unknown environments, demonstrating its superiority in guiding the robot safely towards its desired goal while avoiding obstacles and escaping entrapment in local minima. The GPU implementation of GP-MPPI, including the supplementary video, is available at https://github.com/IhabMohamed/GP-MPPI.
Large medical imaging datasets can be cheaply and quickly annotated with low-confidence, weak labels (e.g., radiological scores). Access to high-confidence labels, such as histology-based diagnoses, is rare and costly. Pretraining strategies, like contrastive learning (CL) methods, can leverage unlabeled or weakly-annotated datasets. These methods typically require large batch sizes, which poses a difficulty in the case of large 3D images at full resolution, due to limited GPU memory. Nevertheless, volumetric positional information about the spatial context of each 2D slice can be very important for some medical applications. In this work, we propose an efficient weakly-supervised positional (WSP) contrastive learning strategy where we integrate both the spatial context of each 2D slice and a weak label via a generic kernel-based loss function. We illustrate our method on cirrhosis prediction using a large volume of weakly-labeled images, namely radiological low-confidence annotations, and small strongly-labeled (i.e., high-confidence) datasets. The proposed model improves the classification AUC by 5% with respect to a baseline model on our internal dataset, and by 26% on the public LIHC dataset from the Cancer Genome Atlas. The code is available at: https://github.com/Guerbet-AI/wsp-contrastive.