PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2103.14891

KnowRU: Knowledge Reusing via Knowledge Distillation in Multi-agent Reinforcement Learning

Recently, deep Reinforcement Learning (RL) algorithms have achieved dramatically progress in the multi-agent area. However, training the increasingly complex tasks would be time-consuming and resources-exhausting. To alleviate this problem, efficient leveraging the historical experience is essential, which is under-explored in previous studies as most of the exiting methods may fail to achieve this goal in a continuously variational system due to their complicated design and environmental dynamics. In this paper, we propose a method, named "KnowRU" for knowledge reusing which can be easily deployed in the majority of the multi-agent reinforcement learning algorithms without complicated hand-coded design. We employ the knowledge distillation paradigm to transfer the knowledge among agents with the goal to accelerate the training phase for new tasks, while improving the asymptotic performance of agents. To empirically demonstrate the robustness and effectiveness of KnowRU, we perform extensive experiments on state-of-the-art multi-agent reinforcement learning (MARL) algorithms on collaborative and competitive scenarios. The results show that KnowRU can outperform the recently reported methods, which emphasizes the importance of the proposed knowledge reusing for MARL.

arXiv:2103.14496

Weakly-Supervised Domain Adaptation of Deep Regression Trackers via Reinforced Knowledge Distillation

Deep regression trackers are among the fastest tracking algorithms available, and therefore suitable for real-time robotic applications. However, their accuracy is inadequate in many domains due to distribution shift and overfitting. In this paper we overcome such limitations by presenting the first methodology for domain adaption of such a class of trackers. To reduce the labeling effort we propose a weakly-supervised adaptation strategy, in which reinforcement learning is used to express weak supervision as a scalar application-dependent and temporally-delayed feedback. At the same time, knowledge distillation is employed to guarantee learning stability and to compress and transfer knowledge from more powerful but slower trackers. Extensive experiments on five different robotic vision domains demonstrate the relevance of our methodology. Real-time speed is achieved on embedded devices and on machines without GPUs, while accuracy reaches significant results.

arXiv:2103.14475공식 코드 · Papers with Code

Distilling Object Detectors via Decoupled Features

Knowledge distillation is a widely used paradigm for inheriting information from a complicated teacher network to a compact student network and maintaining the strong performance. Different from image classification, object detectors are much more sophisticated with multiple loss functions in which features that semantic information rely on are tangled. In this paper, we point out that the information of features derived from regions excluding objects are also essential for distilling the student detector, which is usually ignored in existing approaches. In addition, we elucidate that features from different regions should be assigned with different importance during distillation. To this end, we present a novel distillation algorithm via decoupled features (DeFeat) for learning a better student detector. Specifically, two levels of decoupled features will be processed for embedding useful information into the student, i.e., decoupled features from neck and decoupled proposals from classification head. Extensive experiments on various detectors with different backbones show that the proposed DeFeat is able to surpass the state-of-the-art distillation methods for object detection. For example, DeFeat improves ResNet50 based Faster R-CNN from 37.4% to 40.9% mAP, and improves ResNet50 based RetinaNet from 36.5% to 39.7% mAP on COCO benchmark. Our implementation is available at https://github.com/ggjy/DeFeat.pytorch.

arXiv:2103.14473공식 코드 · Papers with Code

Distilling a Powerful Student Model via Online Knowledge Distillation

Existing online knowledge distillation approaches either adopt the student with the best performance or construct an ensemble model for better holistic performance. However, the former strategy ignores other students' information, while the latter increases the computational complexity during deployment. In this paper, we propose a novel method for online knowledge distillation, termed FFSD, which comprises two key components: Feature Fusion and Self-Distillation, towards solving the above problems in a unified framework. Different from previous works, where all students are treated equally, the proposed FFSD splits them into a leader student and a common student set. Then, the feature fusion module converts the concatenation of feature maps from all common students into a fused feature map. The fused representation is used to assist the learning of the leader student. To enable the leader student to absorb more diverse information, we design an enhancement strategy to increase the diversity among students. Besides, a self-distillation module is adopted to convert the feature map of deeper layers into a shallower one. Then, the shallower layers are encouraged to mimic the transformed feature maps of the deeper layers, which helps the students to generalize better. After training, we simply adopt the leader student, which achieves superior performance, over the common students, without increasing the storage or inference cost. Extensive experiments on CIFAR-100 and ImageNet demonstrate the superiority of our FFSD over existing works. The code is available at https://github.com/SJLeo/FFSD.

arXiv:2103.14431공식 코드 · Papers with Code

Multimodal Knowledge Expansion

The popularity of multimodal sensors and the accessibility of the Internet have brought us a massive amount of unlabeled multimodal data. Since existing datasets and well-trained models are primarily unimodal, the modality gap between a unimodal network and unlabeled multimodal data poses an interesting problem: how to transfer a pre-trained unimodal network to perform the same task on unlabeled multimodal data? In this work, we propose multimodal knowledge expansion (MKE), a knowledge distillation-based framework to effectively utilize multimodal data without requiring labels. Opposite to traditional knowledge distillation, where the student is designed to be lightweight and inferior to the teacher, we observe that a multimodal student model consistently denoises pseudo labels and generalizes better than its teacher. Extensive experiments on four tasks and different modalities verify this finding. Furthermore, we connect the mechanism of MKE to semi-supervised learning and offer both empirical and theoretical explanations to understand the denoising capability of a multimodal student.

arXiv:2103.14337

Hands-on Guidance for Distilling Object Detectors

Knowledge distillation can lead to deploy-friendly networks against the plagued computational complexity problem, but previous methods neglect the feature hierarchy in detectors. Motivated by this, we propose a general framework for detection distillation. Our method, called Hands-on Guidance Distillation, distills the latent knowledge of all stage features for imposing more comprehensive supervision, and focuses on the essence simultaneously for promoting more intense knowledge absorption. Specifically, a series of novel mechanisms are designed elaborately, including correspondence establishment for consistency, hands-on imitation loss measure and re-weighted optimization from both micro and macro perspectives. We conduct extensive evaluations with different distillation configurations over VOC and COCO datasets, which show better performance on accuracy and speed trade-offs. Meanwhile, feasibility experiments on different structural networks further prove the robustness of our HGD.

arXiv:2103.14210공식 코드 · Papers with Code

Leaning Compact and Representative Features for Cross-Modality Person Re-Identification

This paper pays close attention to the cross-modality visible-infrared person re-identification (VI Re-ID) task, which aims to match pedestrian samples between visible and infrared modes. In order to reduce the modality-discrepancy between samples from different cameras, most existing works usually use constraints based on Euclidean metric. Because of the Euclidean based distance metric strategy cannot effectively measure the internal angles between the embedded vectors, the existing solutions cannot learn the angularly discriminative feature embedding. Since the most important factor affecting the classification task based on embedding vector is whether there is an angularly discriminative feature space, in this paper, we present a new loss function called Enumerate Angular Triplet (EAT) loss. Also, motivated by the knowledge distillation, to narrow down the features between different modalities before feature embedding, we further present a novel Cross-Modality Knowledge Distillation (CMKD) loss. Benefit from the above two considerations, the embedded features are discriminative enough in a way to tackle modality-discrepancy problem. The experimental results on RegDB and SYSU-MM01 datasets have demonstrated that the proposed method is superior to the other most advanced methods in terms of impressive performance. Code is available at https://github.com/IVIPLab/LCCRF.

arXiv:2103.14636

A Practical Survey on Faster and Lighter Transformers

Recurrent neural networks are effective models to process sequences. However, they are unable to learn long-term dependencies because of their inherent sequential nature. As a solution, Vaswani et al. introduced the Transformer, a model solely based on the attention mechanism that is able to relate any two positions of the input sequence, hence modelling arbitrary long dependencies. The Transformer has improved the state-of-the-art across numerous sequence modelling tasks. However, its effectiveness comes at the expense of a quadratic computational and memory complexity with respect to the sequence length, hindering its adoption. Fortunately, the deep learning community has always been interested in improving the models' efficiency, leading to a plethora of solutions such as parameter sharing, pruning, mixed-precision, and knowledge distillation. Recently, researchers have directly addressed the Transformer's limitation by designing lower-complexity alternatives such as the Longformer, Reformer, Linformer, and Performer. However, due to the wide range of solutions, it has become challenging for researchers and practitioners to determine which methods to apply in practice in order to meet the desired trade-off between capacity, computation, and memory. This survey addresses this issue by investigating popular approaches to make Transformers faster and lighter and by providing a comprehensive explanation of the methods' strengths, limitations, and underlying assumptions.

arXiv:2103.13811

Student Network Learning via Evolutionary Knowledge Distillation

Knowledge distillation provides an effective way to transfer knowledge via teacher-student learning, where most existing distillation approaches apply a fixed pre-trained model as teacher to supervise the learning of student network. This manner usually brings in a big capability gap between teacher and student networks during learning. Recent researches have observed that a small teacher-student capability gap can facilitate knowledge transfer. Inspired by that, we propose an evolutionary knowledge distillation approach to improve the transfer effectiveness of teacher knowledge. Instead of a fixed pre-trained teacher, an evolutionary teacher is learned online and consistently transfers intermediate knowledge to supervise student network learning on-the-fly. To enhance intermediate knowledge representation and mimicking, several simple guided modules are introduced between corresponding teacher-student blocks. In this way, the student can simultaneously obtain rich internal knowledge and capture its growth process, leading to effective student network learning. Extensive experiments clearly demonstrate the effectiveness of our approach as well as good adaptability in the low-resolution and few-sample visual recognition scenarios.

arXiv:2103.13733

Spirit Distillation: Precise Real-time Semantic Segmentation of Road Scenes with Insufficient Data

Semantic segmentation of road scenes is one of the key technologies for realizing autonomous driving scene perception, and the effectiveness of deep Convolutional Neural Networks(CNNs) for this task has been demonstrated. State-of-art CNNs for semantic segmentation suffer from excessive computations as well as large-scale training data requirement. Inspired by the ideas of Fine-tuning-based Transfer Learning (FTT) and feature-based knowledge distillation, we propose a new knowledge distillation method for cross-domain knowledge transference and efficient data-insufficient network training, named Spirit Distillation(SD), which allow the student network to mimic the teacher network to extract general features, so that a compact and accurate student network can be trained for real-time semantic segmentation of road scenes. Then, in order to further alleviate the trouble of insufficient data and improve the robustness of the student, an Enhanced Spirit Distillation (ESD) method is proposed, which commits to exploit a more comprehensive general features extraction capability by considering images from both the target and the proximity domains as input. To our knowledge, this paper is a pioneering work on the application of knowledge distillation to few-shot learning. Persuasive experiments conducted on Cityscapes semantic segmentation with the prior knowledge transferred from COCO2017 and KITTI demonstrate that our methods can train a better student network (mIOU and high-precision accuracy boost by 1.4% and 8.2% respectively, with 78.2% segmentation variance) with only 41.8% FLOPs (see Fig. 1).

arXiv:2103.13678공식 코드 · Papers with Code

Pruning-then-Expanding Model for Domain Adaptation of Neural Machine Translation

Domain Adaptation is widely used in practical applications of neural machine translation, which aims to achieve good performance on both the general-domain and in-domain. However, the existing methods for domain adaptation usually suffer from catastrophic forgetting, domain divergence, and model explosion. To address these three problems, we propose a method of "divide and conquer" which is based on the importance of neurons or parameters in the translation model. In our method, we first prune the model and only keep the important neurons or parameters, making them responsible for both general-domain and in-domain translation. Then we further train the pruned model supervised by the original unpruned model with the knowledge distillation method. Last we expand the model to the original size and fine-tune the added parameters for the in-domain translation. We conduct experiments on different languages and domains and the results show that our method can achieve significant improvements compared with several strong baselines.

arXiv:2103.12532

Balanced softmax cross-entropy for incremental learning with and without memory

When incrementally trained on new classes, deep neural networks are subject to catastrophic forgetting which leads to an extreme deterioration of their performance on the old classes while learning the new ones. Using a small memory containing few samples from past classes has shown to be an effective method to mitigate catastrophic forgetting. However, due to the limited size of the replay memory, there is a large imbalance between the number of samples for the new and the old classes in the training dataset resulting in bias in the final model. To address this issue, we propose to use the Balanced Softmax Cross-Entropy and show that it can be seamlessly combined with state-of-the-art approaches for class-incremental learning in order to improve their accuracy while also potentially decreasing the computational cost of the training procedure. We further extend this approach to the more demanding class-incremental learning without memory setting and achieve competitive results with memory-based approaches. Experiments on the challenging ImageNet, ImageNet-Subset and CIFAR100 benchmarks with various settings demonstrate the benefits of our approach.

arXiv:2103.11083

Compacting Deep Neural Networks for Internet of Things: Methods and Applications

Deep Neural Networks (DNNs) have shown great success in completing complex tasks. However, DNNs inevitably bring high computational cost and storage consumption due to the complexity of hierarchical structures, thereby hindering their wide deployment in Internet-of-Things (IoT) devices, which have limited computational capability and storage capacity. Therefore, it is a necessity to investigate the technologies to compact DNNs. Despite tremendous advances in compacting DNNs, few surveys summarize compacting-DNNs technologies, especially for IoT applications. Hence, this paper presents a comprehensive study on compacting-DNNs technologies. We categorize compacting-DNNs technologies into three major types: 1) network model compression, 2) Knowledge Distillation (KD), 3) modification of network structures. We also elaborate on the diversity of these approaches and make side-by-side comparisons. Moreover, we discuss the applications of compacted DNNs in various IoT applications and outline future directions.

arXiv:2103.11367공식 코드 · Papers with Code

ROSITA: Refined BERT cOmpreSsion with InTegrAted techniques

Pre-trained language models of the BERT family have defined the state-of-the-arts in a wide range of NLP tasks. However, the performance of BERT-based models is mainly driven by the enormous amount of parameters, which hinders their application to resource-limited scenarios. Faced with this problem, recent studies have been attempting to compress BERT into a small-scale model. However, most previous work primarily focuses on a single kind of compression technique, and few attention has been paid to the combination of different methods. When BERT is compressed with integrated techniques, a critical question is how to design the entire compression framework to obtain the optimal performance. In response to this question, we integrate three kinds of compression methods (weight pruning, low-rank factorization and knowledge distillation (KD)) and explore a range of designs concerning model architecture, KD strategy, pruning frequency and learning rate schedule. We find that a careful choice of the designs is crucial to the performance of the compressed model. Based on the empirical findings, our best compressed model, dubbed Refined BERT cOmpreSsion with InTegrAted techniques (ROSITA), is $7.5 \times$ smaller than BERT while maintains $98.5\%$ of the performance on five tasks of the GLUE benchmark, outperforming the previous BERT compression methods with similar parameter budget. The code is available at https://github.com/llyx97/Rosita.

arXiv:2103.10741

Online Lifelong Generalized Zero-Shot Learning

Methods proposed in the literature for zero-shot learning (ZSL) are typically suitable for offline learning and cannot continually learn from sequential streaming data. The sequential data comes in the form of tasks during training. Recently, a few attempts have been made to handle this issue and develop continual ZSL (CZSL) methods. However, these CZSL methods require clear task-boundary information between the tasks during training, which is not practically possible. This paper proposes a task-free (i.e., task-agnostic) CZSL method, which does not require any task information during continual learning. The proposed task-free CZSL method employs a variational autoencoder (VAE) for performing ZSL. To develop the CZSL method, we combine the concept of experience replay with knowledge distillation and regularization. Here, knowledge distillation is performed using the training sample's dark knowledge, which essentially helps overcome the catastrophic forgetting issue. Further, it is enabled for task-free learning using short-term memory. Finally, a classifier is trained on the synthetic features generated at the latent space of the VAE. Moreover, the experiments are conducted in a challenging and practical ZSL setup, i.e., generalized ZSL (GZSL). These experiments are conducted for two kinds of single-head continual learning settings: (i) mild setting-: task-boundary is known only during training but not during testing; (ii) strict setting-: task-boundary is not known at training, as well as testing. Experimental results on five benchmark datasets exhibit the validity of the approach for CZSL.

arXiv:2103.10673

Cost-effective Deployment of BERT Models in Serverless Environment

In this study we demonstrate the viability of deploying BERT-style models to serverless environments in a production setting. Since the freely available pre-trained models are too large to be deployed in this way, we utilize knowledge distillation and fine-tune the models on proprietary datasets for two real-world tasks: sentiment analysis and semantic textual similarity. As a result, we obtain models that are tuned for a specific domain and deployable in serverless environments. The subsequent performance analysis shows that this solution results in latency levels acceptable for production use and that it is also a cost-effective approach for small-to-medium size deployments of BERT models, all without any infrastructure overhead.

arXiv:2103.10825

Variational Knowledge Distillation for Disease Classification in Chest X-Rays

Disease classification relying solely on imaging data attracts great interest in medical image analysis. Current models could be further improved, however, by also employing Electronic Health Records (EHRs), which contain rich information on patients and findings from clinicians. It is challenging to incorporate this information into disease classification due to the high reliance on clinician input in EHRs, limiting the possibility for automated diagnosis. In this paper, we propose \textit{variational knowledge distillation} (VKD), which is a new probabilistic inference framework for disease classification based on X-rays that leverages knowledge from EHRs. Specifically, we introduce a conditional latent variable model, where we infer the latent representation of the X-ray image with the variational posterior conditioning on the associated EHR text. By doing so, the model acquires the ability to extract the visual features relevant to the disease during learning and can therefore perform more accurate classification for unseen patients at inference based solely on their X-ray scans. We demonstrate the effectiveness of our method on three public benchmark datasets with paired X-ray images and EHRs. The results show that the proposed variational knowledge distillation can consistently improve the performance of medical image classification and significantly surpasses current methods.

arXiv:2103.10081공식 코드 · Papers with Code

Self-Supervised Adaptation for Video Super-Resolution

Recent single-image super-resolution (SISR) networks, which can adapt their network parameters to specific input images, have shown promising results by exploiting the information available within the input data as well as large external datasets. However, the extension of these self-supervised SISR approaches to video handling has yet to be studied. Thus, we present a new learning algorithm that allows conventional video super-resolution (VSR) networks to adapt their parameters to test video frames without using the ground-truth datasets. By utilizing many self-similar patches across space and time, we improve the performance of fully pre-trained VSR networks and produce temporally consistent video frames. Moreover, we present a test-time knowledge distillation technique that accelerates the adaptation speed with less hardware resources. In our experiments, we demonstrate that our novel learning algorithm can fine-tune state-of-the-art VSR networks and substantially elevate performance on numerous benchmark datasets.

arXiv:2103.10047

Similarity Transfer for Knowledge Distillation

Knowledge distillation is a popular paradigm for learning portable neural networks by transferring the knowledge from a large model into a smaller one. Most existing approaches enhance the student model by utilizing the similarity information between the categories of instance level provided by the teacher model. However, these works ignore the similarity correlation between different instances that plays an important role in confidence prediction. To tackle this issue, we propose a novel method in this paper, called similarity transfer for knowledge distillation (STKD), which aims to fully utilize the similarities between categories of multiple samples. Furthermore, we propose to better capture the similarity correlation between different instances by the mixup technique, which creates virtual samples by a weighted linear interpolation. Note that, our distillation loss can fully utilize the incorrect classes similarities by the mixed labels. The proposed approach promotes the performance of student model as the virtual sample created by multiple images produces a similar probability distribution in the teacher and student networks. Experiments and ablation studies on several public classification datasets including CIFAR-10,CIFAR-100,CINIC-10 and Tiny-ImageNet verify that this light-weight method can effectively boost the performance of the compact student model. It shows that STKD substantially has outperformed the vanilla knowledge distillation and has achieved superior accuracy over the state-of-the-art knowledge distillation methods.

arXiv:2103.09903

Transformer-based ASR Incorporating Time-reduction Layer and Fine-tuning with Self-Knowledge Distillation

End-to-end automatic speech recognition (ASR), unlike conventional ASR, does not have modules to learn the semantic representation from speech encoder. Moreover, the higher frame-rate of speech representation prevents the model to learn the semantic representation properly. Therefore, the models that are constructed by the lower frame-rate of speech encoder lead to better performance. For Transformer-based ASR, the lower frame-rate is not only important for learning better semantic representation but also for reducing the computational complexity due to the self-attention mechanism which has O(n^2) order of complexity in both training and inference. In this paper, we propose a Transformer-based ASR model with the time reduction layer, in which we incorporate time reduction layer inside transformer encoder layers in addition to traditional sub-sampling methods to input features that further reduce the frame-rate. This can help in reducing the computational cost of the self-attention process for training and inference with performance improvement. Moreover, we introduce a fine-tuning approach for pre-trained ASR models using self-knowledge distillation (S-KD) which further improves the performance of our ASR model. Experiments on LibriSpeech datasets show that our proposed methods outperform all other Transformer-based ASR systems. Furthermore, with language model (LM) fusion, we achieve new state-of-the-art word error rate (WER) results for Transformer-based ASR models with just 30 million parameters trained without any external data.