PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2402.12736

CST: Calibration Side-Tuning for Parameter and Memory Efficient Transfer Learning

Achieving a universally high accuracy in object detection is quite challenging, and the mainstream focus in the industry currently lies on detecting specific classes of objects. However, deploying one or multiple object detection networks requires a certain amount of GPU memory for training and storage capacity for inference. This presents challenges in terms of how to effectively coordinate multiple object detection tasks under resource-constrained conditions. This paper introduces a lightweight fine-tuning strategy called Calibration side tuning, which integrates aspects of adapter tuning and side tuning to adapt the successful techniques employed in transformers for use with ResNet. The Calibration side tuning architecture that incorporates maximal transition calibration, utilizing a small number of additional parameters to enhance network performance while maintaining a smooth training process. Furthermore, this paper has conducted an analysis on multiple fine-tuning strategies and have implemented their application within ResNet, thereby expanding the research on fine-tuning strategies for object detection networks. Besides, this paper carried out extensive experiments using five benchmark datasets. The experimental results demonstrated that this method outperforms other compared state-of-the-art techniques, and a better balance between the complexity and performance of the finetune schemes is achieved.

arXiv:2402.12419공식 코드 · Papers with Code

EBFT: Effective and Block-Wise Fine-Tuning for Sparse LLMs

Existing methods for fine-tuning sparse LLMs often suffer from resource-intensive requirements and high retraining costs. Additionally, many fine-tuning methods often rely on approximations or heuristic optimization strategies, which may lead to suboptimal solutions. To address these issues, we propose an efficient and fast framework for fine-tuning sparse LLMs based on minimizing reconstruction error. Our approach involves sampling a small dataset for calibration and utilizing backpropagation to iteratively optimize block-wise reconstruction error, on a block-by-block basis, aiming for optimal solutions. Extensive experiments on various benchmarks consistently demonstrate the superiority of our method over other baselines. For instance, on the Wikitext2 dataset with LlamaV1-7B at 70% sparsity, our proposed EBFT achieves a perplexity of 16.88, surpassing the state-of-the-art DSnoT with a perplexity of 75.14. Moreover, with a structured sparsity ratio of 26\%, EBFT achieves a perplexity of 16.27, outperforming LoRA (perplexity 16.44). Furthermore, the fine-tuning process of EBFT for LlamaV1-7B only takes approximately 30 minutes, and the entire framework can be executed on a single 16GB GPU. The source code is available at https://github.com/sunggo/EBFT.

arXiv:2402.12399

Turn Waste into Worth: Rectifying Top-$k$ Router of MoE

Sparse Mixture of Experts (MoE) models are popular for training large language models due to their computational efficiency. However, the commonly used top-$k$ routing mechanism suffers from redundancy computation and memory costs due to the unbalanced routing. Some experts are overflow, where the exceeding tokens are dropped. While some experts are vacant, which are padded with zeros, negatively impacting model performance. To address the dropped tokens and padding, we propose the Rectify-Router, comprising the Intra-GPU Rectification and the Fill-in Rectification. The Intra-GPU Rectification handles dropped tokens, efficiently routing them to experts within the GPU where they are located to avoid inter-GPU communication. The Fill-in Rectification addresses padding by replacing padding tokens with the tokens that have high routing scores. Our experimental results demonstrate that the Intra-GPU Rectification and the Fill-in Rectification effectively handle dropped tokens and padding, respectively. Furthermore, the combination of them achieves superior performance, surpassing the accuracy of the vanilla top-1 router by 4.7%.

arXiv:2402.13077공식 코드 · Papers with Code

Mechanistic Neural Networks for Scientific Machine Learning

This paper presents Mechanistic Neural Networks, a neural network design for machine learning applications in the sciences. It incorporates a new Mechanistic Block in standard architectures to explicitly learn governing differential equations as representations, revealing the underlying dynamics of data and enhancing interpretability and efficiency in data modeling. Central to our approach is a novel Relaxed Linear Programming Solver (NeuRLP) inspired by a technique that reduces solving linear ODEs to solving linear programs. This integrates well with neural networks and surpasses the limitations of traditional ODE solvers enabling scalable GPU parallel processing. Overall, Mechanistic Neural Networks demonstrate their versatility for scientific machine learning applications, adeptly managing tasks from equation discovery to dynamic systems modeling. We prove their comprehensive capabilities in analyzing and interpreting complex scientific data across various applications, showing significant performance against specialized state-of-the-art methods.

arXiv:2402.11339공식 코드 · Papers with Code

Expressive Higher-Order Link Prediction through Hypergraph Symmetry Breaking

A hypergraph consists of a set of nodes along with a collection of subsets of the nodes called hyperedges. Higher-order link prediction is the task of predicting the existence of a missing hyperedge in a hypergraph. A hyperedge representation learned for higher order link prediction is fully expressive when it does not lose distinguishing power up to an isomorphism. Many existing hypergraph representation learners, are bounded in expressive power by the Generalized Weisfeiler Lehman-1 (GWL-1) algorithm, a generalization of the Weisfeiler Lehman-1 algorithm. However, GWL-1 has limited expressive power. In fact, induced subhypergraphs with identical GWL-1 valued nodes are indistinguishable. Furthermore, message passing on hypergraphs can already be computationally expensive, especially on GPU memory. To address these limitations, we devise a preprocessing algorithm that can identify certain regular subhypergraphs exhibiting symmetry. Our preprocessing algorithm runs once with complexity the size of the input hypergraph. During training, we randomly replace subhypergraphs identified by the algorithm with covering hyperedges to break symmetry. We show that our method improves the expressivity of GWL-1. Our extensive experiments also demonstrate the effectiveness of our approach for higher-order link prediction on both graph and hypergraph datasets with negligible change in computation.

arXiv:2402.12369

Short-Period Variables in TESS Full-Frame Image Light Curves Identified via Convolutional Neural Networks

The Transiting Exoplanet Survey Satellite (TESS) mission measured light from stars in ~85% of the sky throughout its two-year primary mission, resulting in millions of TESS 30-minute cadence light curves to analyze in the search for transiting exoplanets. To search this vast dataset, we aim to provide an approach that is both computationally efficient, produces highly performant predictions, and minimizes the required human search effort. We present a convolutional neural network that we train to identify short period variables. To make a prediction for a given light curve, our network requires no prior target parameters identified using other methods. Our network performs inference on a TESS 30-minute cadence light curve in ~5ms on a single GPU, enabling large scale archival searches. We present a collection of 14156 short-period variables identified by our network. The majority of our identified variables fall into two prominent populations, one of short-period main sequence binaries and another of Delta Scuti stars. Our neural network model and related code is additionally provided as open-source code for public use and extension.

arXiv:2402.11322

SpikeNAS: A Fast Memory-Aware Neural Architecture Search Framework for Spiking Neural Network-based Autonomous Agents

Autonomous mobile agents (e.g., UAVs and UGVs) are typically expected to incur low power/energy consumption for solving machine learning tasks (such as object recognition), as these mobile agents are usually powered by portable batteries. These requirements can be fulfilled by Spiking Neural Networks (SNNs), since their bio-inspired spike-based operations offer high accuracy and ultra low-power/energy computation. Currently, most of the SNN architectures are derived from Artificial Neural Networks whose neurons' architectures and operations are different from SNNs, or developed without considering memory budgets from the underlying processing hardware of autonomous mobile agents. These limitations hinder SNNs from reaching their full potential in accuracy and efficiency. Toward this, we propose SpikeNAS, a novel fast memory-aware neural architecture search (NAS) framework for SNNs that quickly finds an appropriate SNN architecture with high accuracy under the given memory budgets from autonomous mobile agents. To do this, our SpikeNAS employs several key steps: analyzing the impacts of network operations on the accuracy, enhancing the network architecture to improve the learning quality, and developing a fast memory-aware search algorithm. The experimental results show that our SpikeNAS improves the searching time and maintains high accuracy as compared to state-of-the-art while meeting the given memory budgets (e.g., 4.4x faster search with 1.3% accuracy improvement for CIFAR100, using an Nvidia RTX 6000 Ada GPU machine), thereby quickly providing the appropriate SNN architecture for the memory-constrained autonomous mobile agents.

arXiv:2402.12061

All Language Models Large and Small

Many leading language models (LMs) use high-intensity computational resources both during training and execution. This poses the challenge of lowering resource costs for deployment and faster execution of decision-making tasks among others. We introduce a novel plug-and-play LM framework named Language Optimising Network Distribution (LONDI) framework. LONDI learns to selectively employ large LMs only where complex decision-making and reasoning are required while using low-resource LMs (i.e. LMs require less GPU usage, but may not be able to solve the problem alone) everywhere else. LONDI consists of a system of two (off-)policy networks, an LM, a large LM (LLM), and a reinforcement learning module that uses switching controls to quickly learn which system states to call the LLM. We then introduce a variant of LONDI that maintains budget constraints on LLM calls and hence its resource usage. Theoretically, we prove LONDI learns the subset of system states to activate the LLM required to solve the task. We then prove that LONDI converges to optimal solutions while also preserving budgetary constraints on LLM calls almost surely enabling it to solve various tasks while significantly lowering computational costs. We test LONDI's performance in a range of tasks in ScienceWorld and BabyAI-Text and demonstrate that LONDI can solve tasks only solvable by resource-intensive LLMs while reducing GPU usage by up to 30%.

arXiv:2402.12373

LTL learning on GPUs

Linear temporal logic (LTL) is widely used in industrial verification. LTL formulae can be learned from traces. Scaling LTL formula learning is an open problem. We implement the first GPU-based LTL learner using a novel form of enumerative program synthesis. The learner is sound and complete. Our benchmarks indicate that it handles traces at least 2048 times more numerous, and on average at least 46 times faster than existing state-of-the-art learners. This is achieved with, among others, novel branch-free LTL semantics that has $O(\log n)$ time complexity, where $n$ is trace length, while previous implementations are $O(n^2)$ or worse (assuming bitwise boolean operations and shifts by powers of 2 have unit costs -- a realistic assumption on modern processors).

arXiv:2402.10462

QDyLoRA: Quantized Dynamic Low-Rank Adaptation for Efficient Large Language Model Tuning

Finetuning large language models requires huge GPU memory, restricting the choice to acquire Larger models. While the quantized version of the Low-Rank Adaptation technique, named QLoRA, significantly alleviates this issue, finding the efficient LoRA rank is still challenging. Moreover, QLoRA is trained on a pre-defined rank and, therefore, cannot be reconfigured for its lower ranks without requiring further fine-tuning steps. This paper proposes QDyLoRA -Quantized Dynamic Low-Rank Adaptation-, as an efficient quantization approach for dynamic low-rank adaptation. Motivated by Dynamic LoRA, QDyLoRA is able to efficiently finetune LLMs on a set of pre-defined LoRA ranks. QDyLoRA enables fine-tuning Falcon-40b for ranks 1 to 64 on a single 32 GB V100-GPU through one round of fine-tuning. Experimental results show that QDyLoRA is competitive to QLoRA and outperforms when employing its optimal rank.

arXiv:2402.10344

Evaluating Neural Radiance Fields (NeRFs) for 3D Plant Geometry Reconstruction in Field Conditions

We evaluate different Neural Radiance Fields (NeRFs) techniques for the 3D reconstruction of plants in varied environments, from indoor settings to outdoor fields. Traditional methods usually fail to capture the complex geometric details of plants, which is crucial for phenotyping and breeding studies. We evaluate the reconstruction fidelity of NeRFs in three scenarios with increasing complexity and compare the results with the point cloud obtained using LiDAR as ground truth. In the most realistic field scenario, the NeRF models achieve a 74.6% F1 score after 30 minutes of training on the GPU, highlighting the efficacy of NeRFs for 3D reconstruction in challenging environments. Additionally, we propose an early stopping technique for NeRF training that almost halves the training time while achieving only a reduction of 7.4% in the average F1 score. This optimization process significantly enhances the speed and efficiency of 3D reconstruction using NeRFs. Our findings demonstrate the potential of NeRFs in detailed and realistic 3D plant reconstruction and suggest practical approaches for enhancing the speed and efficiency of NeRFs in the 3D reconstruction process.

arXiv:2402.10747

Fully Differentiable Lagrangian Convolutional Neural Network for Continuity-Consistent Physics-Informed Precipitation Nowcasting

This paper presents a convolutional neural network model for precipitation nowcasting that combines data-driven learning with physics-informed domain knowledge. We propose LUPIN, a Lagrangian Double U-Net for Physics-Informed Nowcasting, that draws from existing extrapolation-based nowcasting methods and implements the Lagrangian coordinate system transformation of the data in a fully differentiable and GPU-accelerated manner to allow for real-time end-to-end training and inference. Based on our evaluation, LUPIN matches and exceeds the performance of the chosen benchmark, opening the door for other Lagrangian machine learning models.

arXiv:2402.10739공식 코드 · Papers with Code

PointMamba: A Simple State Space Model for Point Cloud Analysis

Transformers have become one of the foundational architectures in point cloud analysis tasks due to their excellent global modeling ability. However, the attention mechanism has quadratic complexity, making the design of a linear complexity method with global modeling appealing. In this paper, we propose PointMamba, transferring the success of Mamba, a recent representative state space model (SSM), from NLP to point cloud analysis tasks. Unlike traditional Transformers, PointMamba employs a linear complexity algorithm, presenting global modeling capacity while significantly reducing computational costs. Specifically, our method leverages space-filling curves for effective point tokenization and adopts an extremely simple, non-hierarchical Mamba encoder as the backbone. Comprehensive evaluations demonstrate that PointMamba achieves superior performance across multiple datasets while significantly reducing GPU memory usage and FLOPs. This work underscores the potential of SSMs in 3D vision-related tasks and presents a simple yet effective Mamba-based baseline for future research. The code will be made available at \url{https://github.com/LMD0311/PointMamba}.

arXiv:2402.09709

ME-ViT: A Single-Load Memory-Efficient FPGA Accelerator for Vision Transformers

Vision Transformers (ViTs) have emerged as a state-of-the-art solution for object classification tasks. However, their computational demands and high parameter count make them unsuitable for real-time inference, prompting the need for efficient hardware implementations. Existing hardware accelerators for ViTs suffer from frequent off-chip memory access, restricting the achievable throughput by memory bandwidth. In devices with a high compute-to-communication ratio (e.g., edge FPGAs with limited bandwidth), off-chip memory access imposes a severe bottleneck on overall throughput. This work proposes ME-ViT, a novel \underline{M}emory \underline{E}fficient FPGA accelerator for \underline{ViT} inference that minimizes memory traffic. We propose a \textit{single-load policy} in designing ME-ViT: model parameters are only loaded once, intermediate results are stored on-chip, and all operations are implemented in a single processing element. To achieve this goal, we design a memory-efficient processing element (ME-PE), which processes multiple key operations of ViT inference on the same architecture through the reuse of \textit{multi-purpose buffers}. We also integrate the Softmax and LayerNorm functions into the ME-PE, minimizing stalls between matrix multiplications. We evaluate ME-ViT on systolic array sizes of 32 and 16, achieving up to a 9.22$\times$ and 17.89$\times$ overall improvement in memory bandwidth, and a 2.16$\times$ improvement in throughput per DSP for both designs over state-of-the-art ViT accelerators on FPGA. ME-ViT achieves a power efficiency improvement of up to 4.00$\times$ (1.03$\times$) over a GPU (FPGA) baseline. ME-ViT enables up to 5 ME-PE instantiations on a Xilinx Alveo U200, achieving a 5.10$\times$ improvement in throughput over the state-of-the art FPGA baseline, and a 5.85$\times$ (1.51$\times$) improvement in power efficiency over the GPU (FPGA) baseline.

arXiv:2402.09780

TinyCL: An Efficient Hardware Architecture for Continual Learning on Autonomous Systems

The Continuous Learning (CL) paradigm consists of continuously evolving the parameters of the Deep Neural Network (DNN) model to progressively learn to perform new tasks without reducing the performance on previous tasks, i.e., avoiding the so-called catastrophic forgetting. However, the DNN parameter update in CL-based autonomous systems is extremely resource-hungry. The existing DNN accelerators cannot be directly employed in CL because they only support the execution of the forward propagation. Only a few prior architectures execute the backpropagation and weight update, but they lack the control and management for CL. Towards this, we design a hardware architecture, TinyCL, to perform CL on resource-constrained autonomous systems. It consists of a processing unit that executes both forward and backward propagation, and a control unit that manages memory-based CL workload. To minimize the memory accesses, the sliding window of the convolutional layer moves in a snake-like fashion. Moreover, the Multiply-and-Accumulate units can be reconfigured at runtime to execute different operations. As per our knowledge, our proposed TinyCL represents the first hardware accelerator that executes CL on autonomous systems. We synthesize the complete TinyCL architecture in a 65 nm CMOS technology node with the conventional ASIC design flow. It executes 1 epoch of training on a Conv + ReLU + Dense model on the CIFAR10 dataset in 1.76 s, while 1 training epoch of the same model using an Nvidia Tesla P100 GPU takes 103 s, thus achieving a 58x speedup, consuming 86 mW in a 4.74 mm2 die.

arXiv:2402.09589

MLTCP: Congestion Control for DNN Training

We present MLTCP, a technique to augment today's congestion control algorithms to accelerate DNN training jobs in shared GPU clusters. MLTCP enables the communication phases of jobs that compete for network bandwidth to interleave with each other, thereby utilizing the network efficiently. At the heart of MLTCP lies a very simple principle based on a key conceptual insight: DNN training flows should scale their congestion window size based on the number of bytes sent at each training iteration. We show that integrating this principle into today's congestion control protocols is straightforward: by adding 30-60 lines of code to Reno, CUBIC, or DCQCN, MLTCP stabilizes flows of different jobs into an interleaved state within a few training iterations, regardless of the number of competing flows or the start time of each flow. Our experiments with popular DNN training jobs demonstrate that enabling MLTCP accelerates the average and 99th percentile training iteration time by up to 2x and 4x, respectively.

arXiv:2402.10207공식 코드 · Papers with Code

Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment

We consider the problem of multi-objective alignment of foundation models with human preferences, which is a critical step towards helpful and harmless AI systems. However, it is generally costly and unstable to fine-tune large foundation models using reinforcement learning (RL), and the multi-dimensionality, heterogeneity, and conflicting nature of human preferences further complicate the alignment process. In this paper, we introduce Rewards-in-Context (RiC), which conditions the response of a foundation model on multiple rewards in its prompt context and applies supervised fine-tuning for alignment. The salient features of RiC are simplicity and adaptivity, as it only requires supervised fine-tuning of a single foundation model and supports dynamic adjustment for user preferences during inference time. Inspired by the analytical solution of an abstracted convex optimization problem, our dynamic inference-time adjustment method approaches the Pareto-optimal solution for multiple objectives. Empirical evidence demonstrates the efficacy of our method in aligning both Large Language Models (LLMs) and diffusion models to accommodate diverse rewards with only around 10% GPU hours compared with multi-objective RL baseline.

arXiv:2402.10193공식 코드 · Papers with Code

BitDelta: Your Fine-Tune May Only Be Worth One Bit

Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it's intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10x, which can also be translated to enhanced generation latency in multi-tenant settings. We validate BitDelta through experiments across Llama-2 and Mistral model families, and on models up to 70B parameters, showcasing minimal performance degradation over all tested settings.

arXiv:2402.10076공식 코드 · Papers with Code

QUICK: Quantization-aware Interleaving and Conflict-free Kernel for efficient LLM inference

We introduce QUICK, a group of novel optimized CUDA kernels for the efficient inference of quantized Large Language Models (LLMs). QUICK addresses the shared memory bank-conflict problem of state-of-the-art mixed precision matrix multiplication kernels. Our method interleaves the quantized weight matrices of LLMs offline to skip the shared memory write-back after the dequantization. We demonstrate up to 1.91x speedup over existing kernels of AutoAWQ on larger batches and up to 1.94x throughput gain on representative LLM models on various NVIDIA GPU devices.

arXiv:2402.09759

Efficient Language Adaptive Pre-training: Extending State-of-the-Art Large Language Models for Polish

This study explores the potential of fine-tuning foundational English Large Language Models (LLMs) for generating Polish text. The first step involves Language Adaptive Pre-training (LAPT) on a high-quality dataset of 3.11 GB, consisting of 276 million Polish tokens. The LAPT is followed by additional fine-tuning aimed at solving nine KLEJ challenges. Our trained model Curie-7B-v1 not only generates Polish text with the lowest perplexity of 3.02 among decoder-based Polish models but also closely rivals the performance of the best Polish encoder-decoder models with a less than 2% gap on 8 out of 9 tasks. Curie-7B-v1 used approximately 2-3% of a typical dataset size to learn Polish. The LAPT was completed in less than five days using a consumer GPU, highlighting the method's efficiency. The proficiency of the model in Polish was significantly enhanced, demonstrating the viability of this approach for adding new languages to existing LLMs by training just 1.2% of its parameters. To contribute to the community's collaborative progress, the model has been released as open-source.