Masked image modeling (MIM) pre-training for large-scale vision transformers (ViTs) in computer vision has enabled promising downstream performance on top of the learned self-supervised ViT features. In this paper, we question if the extremely simple ViTs' fine-tuning performance with a small-scale architecture can also benefit from this pre-training paradigm, which is considerably less studied yet in contrast to the well-established lightweight architecture design methodology with sophisticated components introduced. By carefully adapting various typical MIM pre-training methods to this lightweight regime and comparing them with the contrastive learning (CL) pre-training on various downstream image classification and dense prediction tasks, we systematically observe different behaviors between MIM and CL with respect to the downstream fine-tuning data scales. Furthermore, we analyze the frozen features under linear probing evaluation and also the layer representation similarities and attention maps across the obtained models, which clearly show the inferior learning of MIM pre-training on higher layers, leading to unsatisfactory fine-tuning performance on data-insufficient downstream tasks. This finding is naturally a guide to choosing appropriate distillation strategies during pre-training to solve the above deterioration problem. Extensive experiments on various vision tasks demonstrate the effectiveness of our observation-analysis-solution flow. In particular, our pre-training with distillation on pure lightweight ViTs with vanilla/hierarchical design (5.7M/6.5M) can achieve 79.4%/78.9% top-1 accuracy on ImageNet-1K. It also enables SOTA performance on the ADE20K semantic segmentation task (42.8% mIoU) and LaSOT visual tracking task (66.1% AUC) in the lightweight regime. The latter even surpasses all the current SOTA lightweight CPU-realtime trackers.
ChatGPT has been the most talked-about concept in recent months, captivating both professionals and the general public alike, and has sparked discussions about the changes that artificial intelligence (AI) will bring to the world. As physicists and astrophysicists, we are curious about if scientific data can be correctly analyzed by large language models (LLMs) and yield accurate physics. In this article, we fine-tune the generative pre-trained transformer (GPT) model by the astronomical data from the observations of galaxies, quasars, stars, gamma-ray bursts (GRBs), and the simulations of black holes (BHs), the fine-tuned model demonstrates its capability to classify astrophysical phenomena, distinguish between two types of GRBs, deduce the redshift of quasars, and estimate BH parameters. We regard this as a successful test, marking the LLM's proven efficacy in scientific research. With the ever-growing volume of multidisciplinary data and the advancement of AI technology, we look forward to the emergence of a more fundamental and comprehensive understanding of our universe. This article also shares some interesting thoughts on data collection and AI design. Using the approach of understanding the universe - looking outward at data and inward for fundamental building blocks - as a guideline, we propose a method of series expansion for AI, suggesting ways to train and control AI that is smarter than humans.
Many rescue missions demand effective perception and real-time decision making, which highly rely on effective data collection and processing. In this study, we propose a three-layer architecture of emergency caching networks focusing on data collection and reliable transmission, by leveraging efficient perception and edge caching technologies. Based on this architecture, we propose a disaster map collection framework that integrates coded caching technologies. Our framework strategically caches coded fragments of maps across unmanned aerial vehicles (UAVs), fostering collaborative uploading for augmented transmission reliability. Additionally, we establish a comprehensive probability model to assess the effective recovery area of disaster maps. Towards the goal of utility maximization, we propose a deep reinforcement learning (DRL) based algorithm that jointly makes decisions about cooperative UAVs selection, bandwidth allocation and coded caching parameter adjustment, accommodating the real-time map updates in a dynamic disaster situation. Our proposed scheme is more effective than the non-coding caching scheme, as validated by simulation.
Given a script, the challenge in Movie Dubbing (Visual Voice Cloning, V2C) is to generate speech that aligns well with the video in both time and emotion, based on the tone of a reference audio track. Existing state-of-the-art V2C models break the phonemes in the script according to the divisions between video frames, which solves the temporal alignment problem but leads to incomplete phoneme pronunciation and poor identity stability. To address this problem, we propose StyleDubber, which switches dubbing learning from the frame level to phoneme level. It contains three main components: (1) A multimodal style adaptor operating at the phoneme level to learn pronunciation style from the reference audio, and generate intermediate representations informed by the facial emotion presented in the video; (2) An utterance-level style learning module, which guides both the mel-spectrogram decoding and the refining processes from the intermediate embeddings to improve the overall style expression; And (3) a phoneme-guided lip aligner to maintain lip sync. Extensive experiments on two of the primary benchmarks, V2C and Grid, demonstrate the favorable performance of the proposed method as compared to the current state-of-the-art. The source code and trained models will be released to the public.
Diffusion-based text-to-image personalization have achieved great success in generating subjects specified by users among various contexts. Even though, existing finetuning-based methods still suffer from model overfitting, which greatly harms the generative diversity, especially when given subject images are few. To this end, we propose Pick-and-Draw, a training-free semantic guidance approach to boost identity consistency and generative diversity for personalization methods. Our approach consists of two components: appearance picking guidance and layout drawing guidance. As for the former, we construct an appearance palette with visual features from the reference image, where we pick local patterns for generating the specified subject with consistent identity. As for layout drawing, we outline the subject's contour by referring to a generative template from the vanilla diffusion model, and inherit the strong image prior to synthesize diverse contexts according to different text conditions. The proposed approach can be applied to any personalized diffusion models and requires as few as a single reference image. Qualitative and quantitative experiments show that Pick-and-Draw consistently improves identity consistency and generative diversity, pushing the trade-off between subject fidelity and image-text fidelity to a new Pareto frontier.
Due to highly constrained computing power and memory, deploying 3D lidar-based detectors on edge devices equipped in autonomous vehicles and robots poses a crucial challenge. Being a convenient and straightforward model compression approach, Post-Training Quantization (PTQ) has been widely adopted in 2D vision tasks. However, applying it directly to 3D lidar-based tasks inevitably leads to performance degradation. As a remedy, we propose an effective PTQ method called LiDAR-PTQ, which is particularly curated for 3D lidar detection (both SPConv-based and SPConv-free). Our LiDAR-PTQ features three main components, \textbf{(1)} a sparsity-based calibration method to determine the initialization of quantization parameters, \textbf{(2)} a Task-guided Global Positive Loss (TGPL) to reduce the disparity between the final predictions before and after quantization, \textbf{(3)} an adaptive rounding-to-nearest operation to minimize the layerwise reconstruction error. Extensive experiments demonstrate that our LiDAR-PTQ can achieve state-of-the-art quantization performance when applied to CenterPoint (both Pillar-based and Voxel-based). To our knowledge, for the very first time in lidar-based 3D detection tasks, the PTQ INT8 model's accuracy is almost the same as the FP32 model while enjoying $3\times$ inference speedup. Moreover, our LiDAR-PTQ is cost-effective being $30\times$ faster than the quantization-aware training method. Code will be released at \url{https://github.com/StiphyJay/LiDAR-PTQ}.
Data-to-text (D2T) generation aims to transform structured data into natural language text. Data-to-text pre-training has proved to be powerful in enhancing D2T generation and yields impressive performances. However, previous pre-training methods either oversimplified structured data into a sequence without considering input structures or designed training objectives tailored for a specific data structure (e.g., table or knowledge graph). In this paper, we unify different types of structured data (i.e., table, key-value data, knowledge graph) into the graph format and cast different data-to-text generation tasks as graph-to-text generation. To effectively exploit the structural information of the input graph, we propose a structure-enhanced pre-training method for D2T generation by designing a structure-enhanced Transformer. Concretely, we devise a position matrix for the Transformer, encoding relative positional information of connected nodes in the input graph. In addition, we propose a new attention matrix to incorporate graph structures into the original Transformer by taking the available explicit connectivity structure into account. Extensive experiments on six benchmark datasets show the effectiveness of our model. Our source codes are available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/unid2t.
Visual grounding (VG) aims to locate a specific target in an image based on a given language query. The discriminative information from context is important for distinguishing the target from other objects, particularly for the targets that have the same category as others. However, most previous methods underestimate such information. Moreover, they are usually designed for the standard scene (without any novel object), which limits their generalization to the open-vocabulary scene. In this paper, we propose a novel framework with context disentangling and prototype inheriting for robust visual grounding to handle both scenes. Specifically, the context disentangling disentangles the referent and context features, which achieves better discrimination between them. The prototype inheriting inherits the prototypes discovered from the disentangled visual features by a prototype bank to fully utilize the seen data, especially for the open-vocabulary scene. The fused features, obtained by leveraging Hadamard product on disentangled linguistic and visual features of prototypes to avoid sharp adjusting the importance between the two types of features, are then attached with a special token and feed to a vision Transformer encoder for bounding box regression. Extensive experiments are conducted on both standard and open-vocabulary scenes. The performance comparisons indicate that our method outperforms the state-of-the-art methods in both scenarios. {The code is available at https://github.com/WayneTomas/TransCP.
The large language model era urges faster and less costly inference. Prior model compression works on LLMs tend to undertake a software-centric approach primarily focused on the simulated quantization performance. By neglecting the feasibility of deployment, these approaches are typically disabled in real practice. They used to drastically push down the quantization bit range for a reduced computation which might not be supported by the mainstream hardware, or involve sophisticated algorithms that introduce extra computation or memory access overhead. We argue that pursuing a hardware-centric approach in the construction of quantization algorithms is crucial. In this regard, we are driven to build our compression method on top of hardware awareness, eliminating impractical algorithm choices while maximizing the benefit of hardware acceleration. Our method, OdysseyLLM, comes with a novel W4A8 kernel implementation called FastGEMM and a combined recipe of quantization strategies. Extensive experiments manifest the superiority of our W4A8 method which brings the actual speed boosting up to \textbf{4$\times$} compared to Hugging Face FP16 inference and \textbf{2.23$\times$} vs. the state-of-the-art inference engine TensorRT-LLM in FP16, and \textbf{1.45$\times$} vs. TensorRT-LLM in INT8, yet without substantially harming the performance.
Recent text-to-image (T2I) diffusion models have achieved remarkable progress in generating high-quality images given text-prompts as input. However, these models fail to convey appropriate spatial composition specified by a layout instruction. In this work, we probe into zero-shot grounded T2I generation with diffusion models, that is, generating images corresponding to the input layout information without training auxiliary modules or finetuning diffusion models. We propose a Region and Boundary (R&B) aware cross-attention guidance approach that gradually modulates the attention maps of diffusion model during generative process, and assists the model to synthesize images (1) with high fidelity, (2) highly compatible with textual input, and (3) interpreting layout instructions accurately. Specifically, we leverage the discrete sampling to bridge the gap between consecutive attention maps and discrete layout constraints, and design a region-aware loss to refine the generative layout during diffusion process. We further propose a boundary-aware loss to strengthen object discriminability within the corresponding regions. Experimental results show that our method outperforms existing state-of-the-art zero-shot grounded T2I generation methods by a large margin both qualitatively and quantitatively on several benchmarks.