PyTorchKR Papers

“Code Generation”

총 5,206편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2306.16741공식 코드 · Papers with Code

Foundation Model for Endoscopy Video Analysis via Large-scale Self-supervised Pre-train

Foundation models have exhibited remarkable success in various applications, such as disease diagnosis and text report generation. To date, a foundation model for endoscopic video analysis is still lacking. In this paper, we propose Endo-FM, a foundation model specifically developed using massive endoscopic video data. First, we build a video transformer, which captures both local and global long-range dependencies across spatial and temporal dimensions. Second, we pre-train our transformer model using global and local views via a self-supervised manner, aiming to make it robust to spatial-temporal variations and discriminative across different scenes. To develop the foundation model, we construct a large-scale endoscopy video dataset by combining 9 publicly available datasets and a privately collected dataset from Baoshan Branch of Renji Hospital in Shanghai, China. Our dataset overall consists of over 33K video clips with up to 5 million frames, encompassing various protocols, target organs, and disease types. Our pre-trained Endo-FM can be easily adopted for a given downstream task via fine-tuning by serving as the backbone. With experiments on 3 different types of downstream tasks, including classification, segmentation, and detection, our Endo-FM surpasses the current state-of-the-art (SOTA) self-supervised pre-training and adapter-based transfer learning methods by a significant margin, such as VCL (3.1% F1, 4.8% Dice, and 5.5% F1 for classification, segmentation, and detection) and ST-Adapter (5.9% F1, 9.6% Dice, and 9.9% F1 for classification, segmentation, and detection). Code, datasets, and models are released at https://github.com/med-air/Endo-FM.

arXiv:2403.16335공식 코드 · Papers with Code

MEDDAP: Medical Dataset Enhancement via Diversified Augmentation Pipeline

The effectiveness of Deep Neural Networks (DNNs) heavily relies on the abundance and accuracy of available training data. However, collecting and annotating data on a large scale is often both costly and time-intensive, particularly in medical cases where practitioners are already occupied with their duties. Moreover, ensuring that the model remains robust across various scenarios of image capture is crucial in medical domains, especially when dealing with ultrasound images that vary based on the settings of different devices and the manual operation of the transducer. To address this challenge, we introduce a novel pipeline called MEDDAP, which leverages Stable Diffusion (SD) models to augment existing small datasets by automatically generating new informative labeled samples. Pretrained checkpoints for SD are typically based on natural images, and training them for medical images requires significant GPU resources due to their heavy parameters. To overcome this challenge, we introduce USLoRA (Ultrasound Low-Rank Adaptation), a novel fine-tuning method tailored specifically for ultrasound applications. USLoRA allows for selective fine-tuning of weights within SD, requiring fewer than 0.1\% of parameters compared to fully fine-tuning only the UNet portion of SD. To enhance dataset diversity, we incorporate different adjectives into the generation process prompts, thereby desensitizing the classifiers to intensity changes across different images. This approach is inspired by clinicians' decision-making processes regarding breast tumors, where tumor shape often plays a more crucial role than intensity. In conclusion, our pipeline not only outperforms classifiers trained on the original dataset but also demonstrates superior performance when encountering unseen datasets. The source code is available at https://github.com/yasamin-med/MEDDAP.

arXiv:2406.12295공식 코드 · Papers with Code

Fast and Slow Generating: An Empirical Study on Large and Small Language Models Collaborative Decoding

Large Language Models (LLMs) exhibit impressive capabilities across various applications but encounter substantial challenges such as high inference latency, considerable training costs, and the generation of hallucinations. Collaborative decoding between large and small language models (SLMs) presents a promising strategy to mitigate these issues through methods including speculative decoding, contrastive decoding, and emulator or proxy fine-tuning. However, the specifics of such collaborations, particularly from a unified perspective, remain largely unexplored. Inspired by dual-process cognitive theory, we propose a unified framework in this paper, termed Fast and Slow Generating (FS-GEN). Within this framework, LLMs (sometimes along with SLMs) are categorized as System 2 (slow and deliberate), while independent SLMs are designated as System 1 (fast and intuitive). We provide a comprehensive analysis of these collaborative methodologies, elucidating their common properties and shedding light on the differential knowledge capabilities of System 2 versus System 1 through the FS-GEN framework. Our findings indicate that only a small proportion of collaborative interactions (approximately less than 20\% in most instances) are necessary across various methods. These interactions between System 1 and System 2 conform to a scaling law related to the parameter ratios, enabling predictable collaboration. Furthermore, we explore the specific conditions under which collaboration proves most effective, particularly from an uncertainty perspective, offering novel insights that may guide future optimization efforts. Our research underscores that the fundamental distinction between System 1 and System 2 lies in the uncertainty of next token predictions, where interventions by System 2 are crucial to support System 1. Code for Reproduction: https://github.com/TsinghuaC3I/FS-GEN

arXiv:2408.15205공식 코드 · Papers with Code

Leveraging Hallucinations to Reduce Manual Prompt Dependency in Promptable Segmentation

Promptable segmentation typically requires instance-specific manual prompts to guide the segmentation of each desired object. To minimize such a need, task-generic promptable segmentation has been introduced, which employs a single task-generic prompt to segment various images of different objects in the same task. Current methods use Multimodal Large Language Models (MLLMs) to reason detailed instance-specific prompts from a task-generic prompt for improving segmentation accuracy. The effectiveness of this segmentation heavily depends on the precision of these derived prompts. However, MLLMs often suffer hallucinations during reasoning, resulting in inaccurate prompting. While existing methods focus on eliminating hallucinations to improve a model, we argue that MLLM hallucinations can reveal valuable contextual insights when leveraged correctly, as they represent pre-trained large-scale knowledge beyond individual images. In this paper, we utilize hallucinations to mine task-related information from images and verify its accuracy for enhancing precision of the generated prompts. Specifically, we introduce an iterative Prompt-Mask Cycle generation framework (ProMaC) with a prompt generator and a mask generator.The prompt generator uses a multi-scale chain of thought prompting, initially exploring hallucinations for extracting extended contextual knowledge on a test image.These hallucinations are then reduced to formulate precise instance-specific prompts, directing the mask generator to produce masks that are consistent with task semantics by mask semantic alignment. The generated masks iteratively induce the prompt generator to focus more on task-relevant image areas and reduce irrelevant hallucinations, resulting jointly in better prompts and masks. Experiments on 5 benchmarks demonstrate the effectiveness of ProMaC. Code given in https://lwpyh.github.io/ProMaC/.

arXiv:2410.17779공식 코드 · Papers with Code

ADEM-VL: Adaptive and Embedded Fusion for Efficient Vision-Language Tuning

Recent advancements in multimodal fusion have witnessed the remarkable success of vision-language (VL) models, which excel in various multimodal applications such as image captioning and visual question answering. However, building VL models requires substantial hardware resources, where efficiency is restricted by two key factors: the extended input sequence of the language model with vision features demands more computational operations, and a large number of additional learnable parameters increase memory complexity. These challenges significantly restrict the broader applicability of such models. To bridge this gap, we propose ADEM-VL, an efficient vision-language method that tunes VL models based on pretrained large language models (LLMs) by adopting a parameter-free cross-attention mechanism for similarity measurements in multimodal fusion. This approach only requires embedding vision features into the language space, significantly reducing the number of trainable parameters and accelerating both training and inference speeds. To enhance representation learning in fusion module, we introduce an efficient multiscale feature generation scheme that requires only a single forward pass through the vision encoder. Moreover, we propose an adaptive fusion scheme that dynamically discards less relevant visual information for each text token based on its attention score. This ensures that the fusion process prioritizes the most pertinent visual features. With experiments on various tasks including visual question answering, image captioning, and instruction-following, we demonstrate that our framework outperforms existing approaches. Specifically, our method surpasses existing methods by an average accuracy of 0.77% on ScienceQA dataset, with reduced training and inference latency, demonstrating the superiority of our framework. The code is available at https://github.com/Hao840/ADEM-VL.

arXiv:2412.06674공식 코드 · Papers with Code

EMOv2: Pushing 5M Vision Model Frontier

This work focuses on developing parameter-efficient and lightweight models for dense predictions while trading off parameters, FLOPs, and performance. Our goal is to set up the new frontier of the 5M magnitude lightweight model on various downstream tasks. Inverted Residual Block (IRB) serves as the infrastructure for lightweight CNNs, but no counterparts have been recognized by attention-based design. Our work rethinks the lightweight infrastructure of efficient IRB and practical components in Transformer from a unified perspective, extending CNN-based IRB to attention-based models and abstracting a one-residual Meta Mobile Block (MMBlock) for lightweight model design. Following neat but effective design criterion, we deduce a modern Improved Inverted Residual Mobile Block (i2RMB) and improve a hierarchical Efficient MOdel (EMOv2) with no elaborate complex structures. Considering the imperceptible latency for mobile users when downloading models under 4G/5G bandwidth and ensuring model performance, we investigate the performance upper limit of lightweight models with a magnitude of 5M. Extensive experiments on various vision recognition, dense prediction, and image generation tasks demonstrate the superiority of our EMOv2 over state-of-the-art methods, e.g., EMOv2-1M/2M/5M achieve 72.3, 75.8, and 79.4 Top-1 that surpass equal-order CNN-/Attention-based models significantly. At the same time, EMOv2-5M equipped RetinaNet achieves 41.5 mAP for object detection tasks that surpasses the previous EMO-5M by +2.6. When employing the more robust training recipe, our EMOv2-5M eventually achieves 82.9 Top-1 accuracy, which elevates the performance of 5M magnitude models to a new level. Code is available at https://github.com/zhangzjn/EMOv2.

arXiv:2307.11604공식 코드 · Papers with Code

Consistency-guided Meta-Learning for Bootstrapping Semi-Supervised Medical Image Segmentation

Medical imaging has witnessed remarkable progress but usually requires a large amount of high-quality annotated data which is time-consuming and costly to obtain. To alleviate this burden, semi-supervised learning has garnered attention as a potential solution. In this paper, we present Meta-Learning for Bootstrapping Medical Image Segmentation (MLB-Seg), a novel method for tackling the challenge of semi-supervised medical image segmentation. Specifically, our approach first involves training a segmentation model on a small set of clean labeled images to generate initial labels for unlabeled data. To further optimize this bootstrapping process, we introduce a per-pixel weight mapping system that dynamically assigns weights to both the initialized labels and the model's own predictions. These weights are determined using a meta-process that prioritizes pixels with loss gradient directions closer to those of clean data, which is based on a small set of precisely annotated images. To facilitate the meta-learning process, we additionally introduce a consistency-based Pseudo Label Enhancement (PLE) scheme that improves the quality of the model's own predictions by ensembling predictions from various augmented versions of the same input. In order to improve the quality of the weight maps obtained through multiple augmentations of a single input, we introduce a mean teacher into the PLE scheme. This method helps to reduce noise in the weight maps and stabilize its generation process. Our extensive experimental results on public atrial and prostate segmentation datasets demonstrate that our proposed method achieves state-of-the-art results under semi-supervision. Our code is available at https://github.com/aijinrjinr/MLB-Seg.

arXiv:2310.00868공식 코드 · Papers with Code

RT-GAN: Recurrent Temporal GAN for Adding Lightweight Temporal Consistency to Frame-Based Domain Translation Approaches

Fourteen million colonoscopies are performed annually just in the U.S. However, the videos from these colonoscopies are not saved due to storage constraints (each video from a high-definition colonoscope camera can be in tens of gigabytes). Instead, a few relevant individual frames are saved for documentation/reporting purposes and these are the frames on which most current colonoscopy AI models are trained on. While developing new unsupervised domain translation methods for colonoscopy (e.g. to translate between real optical and virtual/CT colonoscopy), it is thus typical to start with approaches that initially work for individual frames without temporal consistency. Once an individual-frame model has been finalized, additional contiguous frames are added with a modified deep learning architecture to train a new model from scratch for temporal consistency. This transition to temporally-consistent deep learning models, however, requires significantly more computational and memory resources for training. In this paper, we present a lightweight solution with a tunable temporal parameter, RT-GAN (Recurrent Temporal GAN), for adding temporal consistency to individual frame-based approaches that reduces training requirements by a factor of 5. We demonstrate the effectiveness of our approach on two challenging use cases in colonoscopy: haustral fold segmentation (indicative of missed surface) and realistic colonoscopy simulator video generation. We also release a first-of-its kind temporal dataset for colonoscopy for the above use cases. The datasets, accompanying code, and pretrained models will be made available on our Computational Endoscopy Platform GitHub (https://github.com/nadeemlab/CEP). The supplementary video is available at https://youtu.be/UMVP-uIXwWk.

arXiv:2403.02534공식 코드 · Papers with Code

Towards Foundation Time Series Model: To Synthesize Or Not To Synthesize?

The industry is rich in cases when we are required to make forecasting for large amounts of time series at once. However, we might be in a situation where we can not afford to train a separate model for each of them. Such issue in time series modeling remains without due attention. The remedy for this setting is the establishment of a foundation model. Such a model is expected to work in zero-shot and few-shot regimes. However, what should we take as a training dataset for such kind of model? Witnessing the benefits from the enrichment of NLP datasets with artificially-generated data, we might want to adopt their experience for time series. In contrast to natural language, the process of generation of synthetic time series data is even more favorable because it provides full control of series patterns, time horizons, and number of samples. In this work, we consider the essential question if it is advantageous to train a foundation model on synthetic data or it is better to utilize only a limited number of real-life examples. Our experiments are conducted only for regular time series and speak in favor of leveraging solely the real time series. Moreover, the choice of the proper source dataset strongly influences the performance during inference. When provided access even to a limited quantity of short time series data, employing it within a supervised framework yields more favorable results than training on a larger volume of synthetic data. The code for our experiments is publicly available on Github \url{https://github.com/sb-ai-lab/synthesize_or_not}.

arXiv:2407.14086공식 코드 · Papers with Code

Temporal Correlation Meets Embedding: Towards a 2nd Generation of JDE-based Real-Time Multi-Object Tracking

Joint Detection and Embedding (JDE) trackers have demonstrated excellent performance in Multi-Object Tracking (MOT) tasks by incorporating the extraction of appearance features as auxiliary tasks through embedding Re-Identification task (ReID) into the detector, achieving a balance between inference speed and tracking performance. However, solving the competition between the detector and the feature extractor has always been a challenge. Meanwhile, the issue of directly embedding the ReID task into MOT has remained unresolved. The lack of high discriminability in appearance features results in their limited utility. In this paper, a new learning approach using cross-correlation to capture temporal information of objects is proposed. The feature extraction network is no longer trained solely on appearance features from each frame but learns richer motion features by utilizing feature heatmaps from consecutive frames, which addresses the challenge of inter-class feature similarity. Furthermore, our learning approach is applied to a more lightweight feature extraction network, and treat the feature matching scores as strong cues rather than auxiliary cues, with an appropriate weight calculation to reflect the compatibility between our obtained features and the MOT task. Our tracker, named TCBTrack, achieves state-of-the-art performance on multiple public benchmarks, i.e., MOT17, MOT20, and DanceTrack datasets. Specifically, on the DanceTrack test set, we achieve 56.8 HOTA, 58.1 IDF1 and 92.5 MOTA, making it the best online tracker capable of achieving real-time performance. Comparative evaluations with other trackers prove that our tracker achieves the best balance between speed, robustness and accuracy. Code is available at https://github.com/yfzhang1214/TCBTrack.

arXiv:2412.17241공식 코드 · Papers with Code

QTSeg: A Query Token-Based Architecture for Efficient 2D Medical Image Segmentation

Medical image segmentation is crucial in assisting medical doctors in making diagnoses and enabling accurate automatic diagnosis. While advanced convolutional neural networks (CNNs) excel in segmenting regions of interest with pixel-level precision, they often struggle with long-range dependencies, which is crucial for enhancing model performance. Conversely, transformer architectures leverage attention mechanisms to excel in handling long-range dependencies. However, the computational complexity of transformers grows quadratically, posing resource-intensive challenges, especially with high-resolution medical images. Recent research aims to combine CNN and transformer architectures to mitigate their drawbacks and enhance performance while keeping resource demands low. Nevertheless, existing approaches have not fully leveraged the strengths of both architectures to achieve high accuracy with low computational requirements. To address this gap, we propose a novel architecture for 2D medical image segmentation (QTSeg) that leverages a feature pyramid network (FPN) as the image encoder, a multi-level feature fusion (MLFF) as the adaptive module between encoder and decoder and a multi-query mask decoder (MQM Decoder) as the mask decoder. In the first step, an FPN model extracts pyramid features from the input image. Next, MLFF is incorporated between the encoder and decoder to adapt features from different encoder stages to the decoder. Finally, an MQM Decoder is employed to improve mask generation by integrating query tokens with pyramid features at all stages of the mask decoder. Our experimental results show that QTSeg outperforms state-of-the-art methods across all metrics with lower computational demands than the baseline and the existing methods. Code is available at https://github.com/tpnam0901/QTSeg (v0.1.0)

arXiv:2502.08807공식 코드 · Papers with Code

InTAR: Inter-Task Auto-Reconfigurable Accelerator Design for High Data Volume Variation in DNNs

The rise of deep neural networks (DNNs) has driven an increased demand for computing power and memory. Modern DNNs exhibit high data volume variation (HDV) across tasks, which poses challenges for FPGA acceleration: conventional accelerators rely on fixed execution patterns (dataflow or sequential) that can lead to pipeline stalls or necessitate frequent off-chip memory accesses. To address these challenges, we introduce the Inter-Task Auto-Reconfigurable Accelerator (InTAR), a novel accelerator design methodology for HDV applications on FPGAs. InTAR combines the high computational efficiency of sequential execution with the reduced off-chip memory overhead of dataflow execution. It switches execution patterns automatically with a static schedule determined before circuit design based on resource constraints and problem sizes. Unlike previous reconfigurable accelerators, InTAR encodes reconfiguration schedules during circuit design, allowing model-specific optimizations that allocate only the necessary logic and interconnects. Thus, InTAR achieves a high clock frequency with fewer resources and low reconfiguration time. Furthermore, InTAR supports high-level tools such as HLS for fast design generation. We implement a set of multi-task HDV DNN kernels using InTAR. Compared with dataflow and sequential accelerators, InTAR exhibits $\mathbf{1.8\times}$ and $\mathbf{7.1 \times}$ speedups correspondingly. Moreover, we extend InTAR to GPT-2 medium as a more complex example, which is $\mathbf{3.65 \sim 39.14\times}$ faster and a $\mathbf{1.72 \sim 10.44\times}$ more DSP efficient than SoTA accelerators (Allo and DFX) on FPGAs. Additionally, this design demonstrates $\mathbf{1.66 \sim 7.17\times}$ better power efficiency than GPUs. Code: https://github.com/OswaldHe/InTAR

arXiv:2503.01739공식 코드 · Papers with Code

VideoUFO: A Million-Scale User-Focused Dataset for Text-to-Video Generation

Text-to-video generative models convert textual prompts into dynamic visual content, offering wide-ranging applications in film production, gaming, and education. However, their real-world performance often falls short of user expectations. One key reason is that these models have not been trained on videos related to some topics users want to create. In this paper, we propose VideoUFO, the first Video dataset specifically curated to align with Users' FOcus in real-world scenarios. Beyond this, our VideoUFO also features: (1) minimal (0.29%) overlap with existing video datasets, and (2) videos searched exclusively via YouTube's official API under the Creative Commons license. These two attributes provide future researchers with greater freedom to broaden their training sources. The VideoUFO comprises over 1.09 million video clips, each paired with both a brief and a detailed caption (description). Specifically, through clustering, we first identify 1,291 user-focused topics from the million-scale real text-to-video prompt dataset, VidProM. Then, we use these topics to retrieve videos from YouTube, split the retrieved videos into clips, and generate both brief and detailed captions for each clip. After verifying the clips with specified topics, we are left with about 1.09 million video clips. Our experiments reveal that (1) current 16 text-to-video models do not achieve consistent performance across all user-focused topics; and (2) a simple model trained on VideoUFO outperforms others on worst-performing topics. The dataset and code are publicly available at https://huggingface.co/datasets/WenhaoWang/VideoUFO and https://github.com/WangWenhao0716/BenchUFO under the CC BY 4.0 License.

arXiv:2506.20752공식 코드 · Papers with Code

Characterization and Mitigation of Training Instabilities in Microscaling Formats

Training large language models is an expensive, compute-bound process that must be repeated as models scale, algorithms improve, and new data is collected. To address this, next-generation hardware accelerators increasingly support lower-precision arithmetic formats, such as the Microscaling (MX) formats introduced in NVIDIA's Blackwell architecture. These formats use a shared scale within blocks of parameters to extend representable range and perform forward/backward GEMM operations in reduced precision for efficiency gains. In this work, we investigate the challenges and viability of block-scaled precision formats during model training. Across nearly one thousand language models trained from scratch -- spanning compute budgets from $2 \times 10^{17}$ to $4.8 \times 10^{19}$ FLOPs and sweeping over a broad range of weight-activation precision combinations -- we consistently observe that training in MX formats exhibits sharp, stochastic instabilities in the loss, particularly at larger compute scales. To explain this phenomenon, we conduct controlled experiments and ablations on a smaller proxy model that exhibits similar behavior as the language model, sweeping across architectural settings, hyperparameters, and precision formats. These experiments motivate a simple model in which multiplicative gradient bias introduced by the quantization of layer-norm affine parameters and a small fraction of activations can trigger runaway divergence. Through \emph{in situ} intervention experiments on our proxy model, we demonstrate that instabilities can be averted or delayed by modifying precision schemes mid-training. Guided by these findings, we evaluate stabilization strategies in the LLM setting and show that certain hybrid configurations recover performance competitive with full-precision training. We release our code at https://github.com/Hither1/systems-scaling.

arXiv:1705.04932공식 코드 · Papers with Code

GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data

Object Transfiguration replaces an object in an image with another object from a second image. For example it can perform tasks like "putting exactly those eyeglasses from image A on the nose of the person in image B". Usage of exemplar images allows more precise specification of desired modifications and improves the diversity of conditional image generation. However, previous methods that rely on feature space operations, require paired data and/or appearance models for training or disentangling objects from background. In this work, we propose a model that can learn object transfiguration from two unpaired sets of images: one set containing images that "have" that kind of object, and the other set being the opposite, with the mild constraint that the objects be located approximately at the same place. For example, the training data can be one set of reference face images that have eyeglasses, and another set of images that have not, both of which spatially aligned by face landmarks. Despite the weak 0/1 labels, our model can learn an "eyeglasses" subspace that contain multiple representatives of different types of glasses. Consequently, we can perform fine-grained control of generated images, like swapping the glasses in two images by swapping the projected components in the "eyeglasses" subspace, to create novel images of people wearing eyeglasses. Overall, our deterministic generative model learns disentangled attribute subspaces from weakly labeled data by adversarial training. Experiments on CelebA and Multi-PIE datasets validate the effectiveness of the proposed model on real world data, in generating images with specified eyeglasses, smiling, hair styles, and lighting conditions etc. The code is available online.

arXiv:2304.02169공식 코드 · Papers with Code

Synthesize High-dimensional Longitudinal Electronic Health Records via Hierarchical Autoregressive Language Model

Synthetic electronic health records (EHRs) that are both realistic and preserve privacy can serve as an alternative to real EHRs for machine learning (ML) modeling and statistical analysis. However, generating high-fidelity and granular electronic health record (EHR) data in its original, highly-dimensional form poses challenges for existing methods due to the complexities inherent in high-dimensional data. In this paper, we propose Hierarchical Autoregressive Language mOdel (HALO) for generating longitudinal high-dimensional EHR, which preserve the statistical properties of real EHR and can be used to train accurate ML models without privacy concerns. Our HALO method, designed as a hierarchical autoregressive model, generates a probability density function of medical codes, clinical visits, and patient records, allowing for the generation of realistic EHR data in its original, unaggregated form without the need for variable selection or aggregation. Additionally, our model also produces high-quality continuous variables in a longitudinal and probabilistic manner. We conducted extensive experiments and demonstrate that HALO can generate high-fidelity EHR data with high-dimensional disease code probabilities (d > 10,000), disease co-occurrence probabilities within visits (d > 1,000,000), and conditional probabilities across consecutive visits (d > 5,000,000) and achieve above 0.9 R2 correlation in comparison to real EHR data. This performance then enables downstream ML models trained on its synthetic data to achieve comparable accuracy to models trained on real data (0.938 AUROC with HALO data vs. 0.943 with real data). Finally, using a combination of real and synthetic data enhances the accuracy of ML models beyond that achieved by using only real EHR data.

arXiv:2305.08275공식 코드 · Papers with Code

ULIP-2: Towards Scalable Multimodal Pre-training for 3D Understanding

Recent advancements in multimodal pre-training have shown promising efficacy in 3D representation learning by aligning multimodal features across 3D shapes, their 2D counterparts, and language descriptions. However, the methods used by existing frameworks to curate such multimodal data, in particular language descriptions for 3D shapes, are not scalable, and the collected language descriptions are not diverse. To address this, we introduce ULIP-2, a simple yet effective tri-modal pre-training framework that leverages large multimodal models to automatically generate holistic language descriptions for 3D shapes. It only needs 3D data as input, eliminating the need for any manual 3D annotations, and is therefore scalable to large datasets. ULIP-2 is also equipped with scaled-up backbones for better multimodal representation learning. We conduct experiments on two large-scale 3D datasets, Objaverse and ShapeNet, and augment them with tri-modal datasets of 3D point clouds, images, and language for training ULIP-2. Experiments show that ULIP-2 demonstrates substantial benefits in three downstream tasks: zero-shot 3D classification, standard 3D classification with fine-tuning, and 3D captioning (3D-to-language generation). It achieves a new SOTA of 50.6% (top-1) on Objaverse-LVIS and 84.7% (top-1) on ModelNet40 in zero-shot classification. In the ScanObjectNN benchmark for standard fine-tuning, ULIP-2 reaches an overall accuracy of 91.5% with a compact model of only 1.4 million parameters. ULIP-2 sheds light on a new paradigm for scalable multimodal 3D representation learning without human annotations and shows significant improvements over existing baselines. The code and datasets are released at https://github.com/salesforce/ULIP.

arXiv:2307.08691공식 코드 · Papers with Code

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

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).

arXiv:2410.20348공식 코드 · Papers with Code

UTSRMorph: A Unified Transformer and Superresolution Network for Unsupervised Medical Image Registration

Complicated image registration is a key issue in medical image analysis, and deep learning-based methods have achieved better results than traditional methods. The methods include ConvNet-based and Transformer-based methods. Although ConvNets can effectively utilize local information to reduce redundancy via small neighborhood convolution, the limited receptive field results in the inability to capture global dependencies. Transformers can establish long-distance dependencies via a self-attention mechanism; however, the intense calculation of the relationships among all tokens leads to high redundancy. We propose a novel unsupervised image registration method named the unified Transformer and superresolution (UTSRMorph) network, which can enhance feature representation learning in the encoder and generate detailed displacement fields in the decoder to overcome these problems. We first propose a fusion attention block to integrate the advantages of ConvNets and Transformers, which inserts a ConvNet-based channel attention module into a multihead self-attention module. The overlapping attention block, a novel cross-attention method, uses overlapping windows to obtain abundant correlations with match information of a pair of images. Then, the blocks are flexibly stacked into a new powerful encoder. The decoder generation process of a high-resolution deformation displacement field from low-resolution features is considered as a superresolution process. Specifically, the superresolution module was employed to replace interpolation upsampling, which can overcome feature degradation. UTSRMorph was compared to state-of-the-art registration methods in the 3D brain MR (OASIS, IXI) and MR-CT datasets. The qualitative and quantitative results indicate that UTSRMorph achieves relatively better performance. The code and datasets are publicly available at https://github.com/Runshi-Zhang/UTSRMorph.

arXiv:2412.08821공식 코드 · Papers with Code

Large Concept Models: Language Modeling in a Sentence Representation Space

LLMs have revolutionized the field of artificial intelligence and have emerged as the de-facto tool for many tasks. The current established technology of LLMs is to process input and generate output at the token level. This is in sharp contrast to humans who operate at multiple levels of abstraction, well beyond single words, to analyze information and to generate creative content. In this paper, we present an attempt at an architecture which operates on an explicit higher-level semantic representation, which we name a concept. Concepts are language- and modality-agnostic and represent a higher level idea or action in a flow. Hence, we build a "Large Concept Model". In this study, as proof of feasibility, we assume that a concept corresponds to a sentence, and use an existing sentence embedding space, SONAR, which supports up to 200 languages in both text and speech modalities. The Large Concept Model is trained to perform autoregressive sentence prediction in an embedding space. We explore multiple approaches, namely MSE regression, variants of diffusion-based generation, and models operating in a quantized SONAR space. These explorations are performed using 1.6B parameter models and training data in the order of 1.3T tokens. We then scale one architecture to a model size of 7B parameters and training data of about 2.7T tokens. We perform an experimental evaluation on several generative tasks, namely summarization and a new task of summary expansion. Finally, we show that our model exhibits impressive zero-shot generalization performance to many languages, outperforming existing LLMs of the same size. The training code of our models is freely available.