PyTorchKR Papers

최신 논문

총 1,512편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2402.12406공식 코드 · Papers with Code

Teacher as a Lenient Expert: Teacher-Agnostic Data-Free Knowledge Distillation

Data-free knowledge distillation (DFKD) aims to distill pretrained knowledge to a student model with the help of a generator without using original data. In such data-free scenarios, achieving stable performance of DFKD is essential due to the unavailability of validation data. Unfortunately, this paper has discovered that existing DFKD methods are quite sensitive to different teacher models, occasionally showing catastrophic failures of distillation, even when using well-trained teacher models. Our observation is that the generator in DFKD is not always guaranteed to produce precise yet diverse samples using the existing representative strategy of minimizing both class-prior and adversarial losses. Through our empirical study, we focus on the fact that class-prior not only decreases the diversity of generated samples, but also cannot completely address the problem of generating unexpectedly low-quality samples depending on teacher models. In this paper, we propose the teacher-agnostic data-free knowledge distillation (TA-DFKD) method, with the goal of more robust and stable performance regardless of teacher models. Our basic idea is to assign the teacher model a lenient expert role for evaluating samples, rather than a strict supervisor that enforces its class-prior on the generator. Specifically, we design a sample selection approach that takes only clean samples verified by the teacher model without imposing restrictions on the power of generating diverse samples. Through extensive experiments, we show that our method successfully achieves both robustness and training stability across various teacher models, while outperforming the existing DFKD methods.

arXiv:2402.12030공식 코드 · Papers with Code

Towards Cross-Tokenizer Distillation: the Universal Logit Distillation Loss for LLMs

Deploying large language models (LLMs) of several billion parameters can be impractical in most industrial use cases due to constraints such as cost, latency limitations, and hardware accessibility. Knowledge distillation (KD) offers a solution by compressing knowledge from resource-intensive large models to smaller ones. Various strategies exist, some relying on the text generated by the teacher model and optionally utilizing his logits to enhance learning. However, these methods based on logits often require both teacher and student models to share the same tokenizer, limiting their applicability across different LLM families. In this paper, we introduce Universal Logit Distillation (ULD) loss, grounded in optimal transport, to address this limitation. Our experimental results demonstrate the effectiveness of ULD loss in enabling distillation across models with different architectures and tokenizers, paving the way to a more widespread use of distillation techniques.

arXiv:2402.11890공식 코드 · Papers with Code

Revisiting Knowledge Distillation for Autoregressive Language Models

Knowledge distillation (KD) is a common approach to compress a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, in the context of autoregressive language models (LMs), we empirically find that larger teacher LMs might dramatically result in a poorer student. In response to this problem, we conduct a series of analyses and reveal that different tokens have different teaching modes, neglecting which will lead to performance degradation. Motivated by this, we propose a simple yet effective adaptive teaching approach (ATKD) to improve the KD. The core of ATKD is to reduce rote learning and make teaching more diverse and flexible. Extensive experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains (up to +3.04% average score) across all model types and sizes. More encouragingly, ATKD can improve the student model generalization effectively.

arXiv:2402.11401공식 코드 · Papers with Code

GraphKD: Exploring Knowledge Distillation Towards Document Object Detection with Structured Graph Creation

Object detection in documents is a key step to automate the structural elements identification process in a digital or scanned document through understanding the hierarchical structure and relationships between different elements. Large and complex models, while achieving high accuracy, can be computationally expensive and memory-intensive, making them impractical for deployment on resource constrained devices. Knowledge distillation allows us to create small and more efficient models that retain much of the performance of their larger counterparts. Here we present a graph-based knowledge distillation framework to correctly identify and localize the document objects in a document image. Here, we design a structured graph with nodes containing proposal-level features and edges representing the relationship between the different proposal regions. Also, to reduce text bias an adaptive node sampling strategy is designed to prune the weight distribution and put more weightage on non-text nodes. We encode the complete graph as a knowledge representation and transfer it from the teacher to the student through the proposed distillation loss by effectively capturing both local and global information concurrently. Extensive experimentation on competitive benchmarks demonstrates that the proposed framework outperforms the current state-of-the-art approaches. The code will be available at: https://github.com/ayanban011/GraphKD.

arXiv:2402.11187공식 코드 · Papers with Code

LaCo: Large Language Model Pruning via Layer Collapse

Large language models (LLMs) based on transformer are witnessing a notable trend of size expansion, which brings considerable costs to both model training and inference. However, existing methods such as model quantization, knowledge distillation, and model pruning are constrained by various issues, including hardware support limitations, the need for extensive training, and alterations to the model internal structure. In this paper, we propose a concise layer-wise structured pruner called \textit{Layer Collapse (LaCo)}, in which rear model layers collapse into a prior layer, enabling a rapid reduction in model size while preserving the model structure. Comprehensive experiments show that our method maintains an average task performance of over 80\% at pruning ratios of 25-30\%, significantly outperforming existing state-of-the-art structured pruning methods. We also conduct post-training experiments to confirm that the \textit{LaCo} effectively inherits the parameters of the original model. Additionally, we perform ablation studies on various settings of \textit{LaCo}. Finally, we discuss our motivation from the perspective of layer-wise similarity and evaluate the performance of the pruned LLMs across various pruning ratios\footnote{\url{https://github.com/yangyifei729/LaCo}}.

arXiv:2402.12265공식 코드 · Papers with Code

On the Byzantine-Resilience of Distillation-Based Federated Learning

Federated Learning (FL) algorithms using Knowledge Distillation (KD) have received increasing attention due to their favorable properties with respect to privacy, non-i.i.d. data and communication cost. These methods depart from transmitting model parameters and instead communicate information about a learning task by sharing predictions on a public dataset. In this work, we study the performance of such approaches in the byzantine setting, where a subset of the clients act in an adversarial manner aiming to disrupt the learning process. We show that KD-based FL algorithms are remarkably resilient and analyze how byzantine clients can influence the learning process. Based on these insights, we introduce two new byzantine attacks and demonstrate their ability to break existing byzantine-resilient methods. Additionally, we propose a novel defence method which enhances the byzantine resilience of KD-based FL algorithms. Finally, we provide a general framework to obfuscate attacks, making them significantly harder to detect, thereby improving their effectiveness. Our findings serve as an important building block in the analysis of byzantine FL, contributing through the development of new attacks and new defence mechanisms, further advancing the robustness of KD-based FL algorithms.

arXiv:2402.11148공식 코드 · Papers with Code

Knowledge Distillation Based on Transformed Teacher Matching

As a technique to bridge logit matching and probability distribution matching, temperature scaling plays a pivotal role in knowledge distillation (KD). Conventionally, temperature scaling is applied to both teacher's logits and student's logits in KD. Motivated by some recent works, in this paper, we drop instead temperature scaling on the student side, and systematically study the resulting variant of KD, dubbed transformed teacher matching (TTM). By reinterpreting temperature scaling as a power transform of probability distribution, we show that in comparison with the original KD, TTM has an inherent R\'enyi entropy term in its objective function, which serves as an extra regularization term. Extensive experiment results demonstrate that thanks to this inherent regularization, TTM leads to trained students with better generalization than the original KD. To further enhance student's capability to match teacher's power transformed probability distribution, we introduce a sample-adaptive weighting coefficient into TTM, yielding a novel distillation approach dubbed weighted TTM (WTTM). It is shown, by comprehensive experiments, that although WTTM is simple, it is effective, improves upon TTM, and achieves state-of-the-art accuracy performance. Our source code is available at https://github.com/zkxufo/TTM.

arXiv:2402.10631공식 코드 · Papers with Code

BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation

The upscaling of Large Language Models (LLMs) has yielded impressive advances in natural language processing, yet it also poses significant deployment challenges. Weight quantization has emerged as a widely embraced solution to reduce memory and computational demands. This paper introduces BitDistiller, a framework that synergizes Quantization-Aware Training (QAT) with Knowledge Distillation (KD) to boost the performance of LLMs at ultra-low precisions (sub-4-bit). Specifically, BitDistiller first incorporates a tailored asymmetric quantization and clipping technique to maximally preserve the fidelity of quantized weights, and then proposes a novel Confidence-Aware Kullback-Leibler Divergence (CAKLD) objective, which is employed in a self-distillation manner to enable faster convergence and superior model performance. Empirical evaluations demonstrate that BitDistiller significantly surpasses existing methods in both 3-bit and 2-bit configurations on general language understanding and complex reasoning benchmarks. Notably, BitDistiller is shown to be more cost-effective, demanding fewer data and training resources. The code is available at https://github.com/DD-DuDa/BitDistiller.

arXiv:2402.10447공식 코드 · Papers with Code

Incremental Sequence Labeling: A Tale of Two Shifts

The incremental sequence labeling task involves continuously learning new classes over time while retaining knowledge of the previous ones. Our investigation identifies two significant semantic shifts: E2O (where the model mislabels an old entity as a non-entity) and O2E (where the model labels a non-entity or old entity as a new entity). Previous research has predominantly focused on addressing the E2O problem, neglecting the O2E issue. This negligence results in a model bias towards classifying new data samples as belonging to the new class during the learning process. To address these challenges, we propose a novel framework, Incremental Sequential Labeling without Semantic Shifts (IS3). Motivated by the identified semantic shifts (E2O and O2E), IS3 aims to mitigate catastrophic forgetting in models. As for the E2O problem, we use knowledge distillation to maintain the model's discriminative ability for old entities. Simultaneously, to tackle the O2E problem, we alleviate the model's bias towards new entities through debiased loss and optimization levels. Our experimental evaluation, conducted on three datasets with various incremental settings, demonstrates the superior performance of IS3 compared to the previous state-of-the-art method by a significant margin.The data, code, and scripts are publicly available at https://github.com/zzz47zzz/codebase-for-incremental-learning-with-llm.

arXiv:2402.09773공식 코드 · Papers with Code

NutePrune: Efficient Progressive Pruning with Numerous Teachers for Large Language Models

The considerable size of Large Language Models (LLMs) presents notable deployment challenges, particularly on resource-constrained hardware. Structured pruning, offers an effective means to compress LLMs, thereby reducing storage costs and enhancing inference speed for more efficient utilization. In this work, we study data-efficient and resource-efficient structure pruning methods to obtain smaller yet still powerful models. Knowledge Distillation is well-suited for pruning, as the intact model can serve as an excellent teacher for pruned students. However, it becomes challenging in the context of LLMs due to memory constraints. To address this, we propose an efficient progressive Numerous-teacher pruning method (NutePrune). NutePrune mitigates excessive memory costs by loading only one intact model and integrating it with various masks and LoRA modules, enabling it to seamlessly switch between teacher and student roles. This approach allows us to leverage numerous teachers with varying capacities to progressively guide the pruned model, enhancing overall performance. Extensive experiments across various tasks demonstrate the effectiveness of NutePrune. In LLaMA-7B zero-shot experiments, NutePrune retains 97.17% of the performance of the original model at 20% sparsity and 95.07% at 25% sparsity. Our code is available at https://github.com/Lucius-lsr/NutePrune.

arXiv:2402.09095공식 코드 · Papers with Code

FedSiKD: Clients Similarity and Knowledge Distillation: Addressing Non-i.i.d. and Constraints in Federated Learning

In recent years, federated learning (FL) has emerged as a promising technique for training machine learning models in a decentralized manner while also preserving data privacy. The non-independent and identically distributed (non-i.i.d.) nature of client data, coupled with constraints on client or edge devices, presents significant challenges in FL. Furthermore, learning across a high number of communication rounds can be risky and potentially unsafe for model exploitation. Traditional FL approaches may suffer from these challenges. Therefore, we introduce FedSiKD, which incorporates knowledge distillation (KD) within a similarity-based federated learning framework. As clients join the system, they securely share relevant statistics about their data distribution, promoting intra-cluster homogeneity. This enhances optimization efficiency and accelerates the learning process, effectively transferring knowledge between teacher and student models and addressing device constraints. FedSiKD outperforms state-of-the-art algorithms by achieving higher accuracy, exceeding by 25\% and 18\% for highly skewed data at $\alpha = {0.1,0.5}$ on the HAR and MNIST datasets, respectively. Its faster convergence is illustrated by a 17\% and 20\% increase in accuracy within the first five rounds on the HAR and MNIST datasets, respectively, highlighting its early-stage learning proficiency. Code is publicly available and hosted on GitHub (https://github.com/SimuEnv/FedSiKD)

arXiv:2402.07059공식 코드 · Papers with Code

Domain Adaptable Fine-Tune Distillation Framework For Advancing Farm Surveillance

In this study, we propose an automated framework for camel farm monitoring, introducing two key contributions: the Unified Auto-Annotation framework and the Fine-Tune Distillation framework. The Unified Auto-Annotation approach combines two models, GroundingDINO (GD), and Segment-Anything-Model (SAM), to automatically annotate raw datasets extracted from surveillance videos. Building upon this foundation, the Fine-Tune Distillation framework conducts fine-tuning of student models using the auto-annotated dataset. This process involves transferring knowledge from a large teacher model to a student model, resembling a variant of Knowledge Distillation. The Fine-Tune Distillation framework aims to be adaptable to specific use cases, enabling the transfer of knowledge from the large models to the small models, making it suitable for domain-specific applications. By leveraging our raw dataset collected from Al-Marmoom Camel Farm in Dubai, UAE, and a pre-trained teacher model, GroundingDINO, the Fine-Tune Distillation framework produces a lightweight deployable model, YOLOv8. This framework demonstrates high performance and computational efficiency, facilitating efficient real-time object detection. Our code is available at \href{https://github.com/Razaimam45/Fine-Tune-Distillation}{https://github.com/Razaimam45/Fine-Tune-Distillation}

arXiv:2402.07295공식 코드 · Papers with Code

Training Heterogeneous Client Models using Knowledge Distillation in Serverless Federated Learning

Federated Learning (FL) is an emerging machine learning paradigm that enables the collaborative training of a shared global model across distributed clients while keeping the data decentralized. Recent works on designing systems for efficient FL have shown that utilizing serverless computing technologies, particularly Function-as-a-Service (FaaS) for FL, can enhance resource efficiency, reduce training costs, and alleviate the complex infrastructure management burden on data holders. However, existing serverless FL systems implicitly assume a uniform global model architecture across all participating clients during training. This assumption fails to address fundamental challenges in practical FL due to the resource and statistical data heterogeneity among FL clients. To address these challenges and enable heterogeneous client models in serverless FL, we utilize Knowledge Distillation (KD) in this paper. Towards this, we propose novel optimized serverless workflows for two popular conventional federated KD techniques, i.e., FedMD and FedDF. We implement these workflows by introducing several extensions to an open-source serverless FL system called FedLess. Moreover, we comprehensively evaluate the two strategies on multiple datasets across varying levels of client data heterogeneity using heterogeneous client models with respect to accuracy, fine-grained training times, and costs. Results from our experiments demonstrate that serverless FedDF is more robust to extreme non-IID data distributions, is faster, and leads to lower costs than serverless FedMD. In addition, compared to the original implementation, our optimizations for particular steps in FedMD and FedDF lead to an average speedup of 3.5x and 1.76x across all datasets.

arXiv:2402.05942공식 코드 · Papers with Code

Cooperative Knowledge Distillation: A Learner Agnostic Approach

Knowledge distillation is a simple but powerful way to transfer knowledge between a teacher model to a student model. Existing work suffers from at least one of the following key limitations in terms of direction and scope of transfer which restrict its use: all knowledge is transferred from teacher to student regardless of whether or not that knowledge is useful, the student is the only one learning in this exchange, and typically distillation transfers knowledge only from a single teacher to a single student. We formulate a novel form of knowledge distillation in which many models can act as both students and teachers which we call cooperative distillation. The models cooperate as follows: a model (the student) identifies specific deficiencies in it's performance and searches for another model (the teacher) who encodes learned knowledge into instructional virtual instances via counterfactual instance generation. Because different models may have different strengths and weaknesses, all models can act as either students or teachers (cooperation) when appropriate and only distill knowledge in areas specific to their strengths (focus). Since counterfactuals as a paradigm are not tied to any specific algorithm, we can use this method to distill knowledge between learners of different architectures, algorithms, and even feature spaces. We demonstrate that our approach not only outperforms baselines such as transfer learning, self-supervised learning, and multiple knowledge distillation algorithms on several datasets, but it can also be used in settings where the aforementioned techniques cannot.

arXiv:2402.06213공식 코드 · Papers with Code

Multi-source-free Domain Adaptation via Uncertainty-aware Adaptive Distillation

Source-free domain adaptation (SFDA) alleviates the domain discrepancy among data obtained from domains without accessing the data for the awareness of data privacy. However, existing conventional SFDA methods face inherent limitations in medical contexts, where medical data are typically collected from multiple institutions using various equipment. To address this problem, we propose a simple yet effective method, named Uncertainty-aware Adaptive Distillation (UAD) for the multi-source-free unsupervised domain adaptation (MSFDA) setting. UAD aims to perform well-calibrated knowledge distillation from (i) model level to deliver coordinated and reliable base model initialisation and (ii) instance level via model adaptation guided by high-quality pseudo-labels, thereby obtaining a high-performance target domain model. To verify its general applicability, we evaluate UAD on two image-based diagnosis benchmarks among two multi-centre datasets, where our method shows a significant performance gain compared with existing works. The code will be available soon.

arXiv:2402.04616공식 코드 · Papers with Code

Beyond Answers: Transferring Reasoning Capabilities to Smaller LLMs Using Multi-Teacher Knowledge Distillation

Transferring the reasoning capability from stronger large language models (LLMs) to smaller ones has been quite appealing, as smaller LLMs are more flexible to deploy with less expense. Among the existing solutions, knowledge distillation stands out due to its outstanding efficiency and generalization. However, existing methods suffer from several drawbacks, including limited knowledge diversity and the lack of rich contextual information. To solve the problems and facilitate the learning of compact language models, we propose TinyLLM, a new knowledge distillation paradigm to learn a small student LLM from multiple large teacher LLMs. In particular, we encourage the student LLM to not only generate the correct answers but also understand the rationales behind these answers. Given that different LLMs possess diverse reasoning skills, we guide the student model to assimilate knowledge from various teacher LLMs. We further introduce an in-context example generator and a teacher-forcing Chain-of-Thought strategy to ensure that the rationales are accurate and grounded in contextually appropriate scenarios. Extensive experiments on six datasets across two reasoning tasks demonstrate the superiority of our method. Results show that TinyLLM can outperform large teacher LLMs significantly, despite a considerably smaller model size. The source code is available at: https://github.com/YikunHan42/TinyLLM.

arXiv:2402.05008공식 코드 · Papers with Code

EfficientViT-SAM: Accelerated Segment Anything Model Without Accuracy Loss

We present EfficientViT-SAM, a new family of accelerated segment anything models. We retain SAM's lightweight prompt encoder and mask decoder while replacing the heavy image encoder with EfficientViT. For the training, we begin with the knowledge distillation from the SAM-ViT-H image encoder to EfficientViT. Subsequently, we conduct end-to-end training on the SA-1B dataset. Benefiting from EfficientViT's efficiency and capacity, EfficientViT-SAM delivers 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing performance. Our code and pre-trained models are released at https://github.com/mit-han-lab/efficientvit.

arXiv:2402.03749공식 코드 · Papers with Code

Vision Superalignment: Weak-to-Strong Generalization for Vision Foundation Models

Recent advancements in large language models have sparked interest in their extraordinary and near-superhuman capabilities, leading researchers to explore methods for evaluating and optimizing these abilities, which is called superalignment. In this context, our paper delves into the realm of vision foundation models, focusing on the concept of weak-to-strong generalization, which involves using a weaker model to supervise a stronger one, aiming to enhance the latter's capabilities beyond the former's limits. We introduce a novel and adaptively adjustable loss function for weak-to-strong supervision. Our comprehensive experiments span various scenarios, including few-shot learning, transfer learning, noisy label learning, and common knowledge distillation settings. The results are striking: our approach not only exceeds the performance benchmarks set by strong-to-strong generalization but also surpasses the outcomes of fine-tuning strong models with whole datasets. This compelling evidence underscores the significant potential of weak-to-strong generalization, showcasing its capability to substantially elevate the performance of vision foundation models. The code is available at https://github.com/ggjy/vision_weak_to_strong.

arXiv:2402.03898공식 코드 · Papers with Code

DistiLLM: Towards Streamlined Distillation for Large Language Models

Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3$\times$ speedup compared to recent KD methods.

arXiv:2402.02446공식 코드 · Papers with Code

LQER: Low-Rank Quantization Error Reconstruction for LLMs

Post-training quantization of Large Language Models (LLMs) is challenging. In this work, we introduce Low-rank Quantization Error Reduction (LQER), which combines quantization and low-rank approximation to recover the model capability. LQER leverages an activation-induced scale matrix to drive the singular value distribution of quantization error towards a desirable distribution, which enables nearly-lossless W4A8 quantization on various LLMs and downstream tasks without the need for knowledge distillation, grid search, or gradient-base iterative optimization. Unlike existing methods, the computation pattern of LQER eliminates the need for specialized Scatter and Gather processes to collect high-precision weights from irregular memory locations. Our W4A8 LLMs achieve near-lossless performance on six popular downstream tasks, while using 1.36$\times$ fewer hardware resources than the leading state-of-the-art method. We open-source our framework at https://github.com/ChengZhang-98/lqer