PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2011.14554

A Selective Survey on Versatile Knowledge Distillation Paradigm for Neural Network Models

This paper aims to provide a selective survey about knowledge distillation(KD) framework for researchers and practitioners to take advantage of it for developing new optimized models in the deep neural network field. To this end, we give a brief overview of knowledge distillation and some related works including learning using privileged information(LUPI) and generalized distillation(GD). Even though knowledge distillation based on the teacher-student architecture was initially devised as a model compression technique, it has found versatile applications over various frameworks. In this paper, we review the characteristics of knowledge distillation from the hypothesis that the three important ingredients of knowledge distillation are distilled knowledge and loss,teacher-student paradigm, and the distillation process. In addition, we survey the versatility of the knowledge distillation by studying its direct applications and its usage in combination with other deep learning paradigms. Finally we present some future works in knowledge distillation including explainable knowledge distillation where the analytical analysis of the performance gain is studied and the self-supervised learning which is a hot research topic in deep learning community.

arXiv:2011.13317

Adaptive Multiplane Image Generation from a Single Internet Picture

In the last few years, several works have tackled the problem of novel view synthesis from stereo images or even from a single picture. However, previous methods are computationally expensive, specially for high-resolution images. In this paper, we address the problem of generating a multiplane image (MPI) from a single high-resolution picture. We present the adaptive-MPI representation, which allows rendering novel views with low computational requirements. To this end, we propose an adaptive slicing algorithm that produces an MPI with a variable number of image planes. We present a new lightweight CNN for depth estimation, which is learned by knowledge distillation from a larger network. Occluded regions in the adaptive-MPI are inpainted also by a lightweight CNN. We show that our method is capable of producing high-quality predictions with one order of magnitude less parameters compared to previous approaches. The robustness of our method is evidenced on challenging pictures from the Internet.

arXiv:2011.13256공식 코드 · Papers with Code

Channel-wise Knowledge Distillation for Dense Prediction

Knowledge distillation (KD) has been proven to be a simple and effective tool for training compact models. Almost all KD variants for dense prediction tasks align the student and teacher networks' feature maps in the spatial domain, typically by minimizing point-wise and/or pair-wise discrepancy. Observing that in semantic segmentation, some layers' feature activations of each channel tend to encode saliency of scene categories (analogue to class activation mapping), we propose to align features channel-wise between the student and teacher networks. To this end, we first transform the feature map of each channel into a probabilty map using softmax normalization, and then minimize the Kullback-Leibler (KL) divergence of the corresponding channels of the two networks. By doing so, our method focuses on mimicking the soft distributions of channels between networks. In particular, the KL divergence enables learning to pay more attention to the most salient regions of the channel-wise maps, presumably corresponding to the most useful signals for semantic segmentation. Experiments demonstrate that our channel-wise distillation outperforms almost all existing spatial distillation methods for semantic segmentation considerably, and requires less computational cost during training. We consistently achieve superior performance on three benchmarks with various network structures. Code is available at: https://git.io/Distiller

arXiv:2011.12913공식 코드 · Papers with Code

torchdistill: A Modular, Configuration-Driven Framework for Knowledge Distillation

While knowledge distillation (transfer) has been attracting attentions from the research community, the recent development in the fields has heightened the need for reproducible studies and highly generalized frameworks to lower barriers to such high-quality, reproducible deep learning research. Several researchers voluntarily published frameworks used in their knowledge distillation studies to help other interested researchers reproduce their original work. Such frameworks, however, are usually neither well generalized nor maintained, thus researchers are still required to write a lot of code to refactor/build on the frameworks for introducing new methods, models, datasets and designing experiments. In this paper, we present our developed open-source framework built on PyTorch and dedicated for knowledge distillation studies. The framework is designed to enable users to design experiments by declarative PyYAML configuration files, and helps researchers complete the recently proposed ML Code Completeness Checklist. Using the developed framework, we demonstrate its various efficient training strategies, and implement a variety of knowledge distillation methods. We also reproduce some of their original experimental results on the ImageNet and COCO datasets presented at major machine learning conferences such as ICLR, NeurIPS, CVPR and ECCV, including recent state-of-the-art methods. All the source code, configurations, log files and trained model weights are publicly available at https://github.com/yoshitomo-matsubara/torchdistill .

arXiv:2011.11472

Generative Adversarial Simulator

Knowledge distillation between machine learning models has opened many new avenues for parameter count reduction, performance improvements, or amortizing training time when changing architectures between the teacher and student network. In the case of reinforcement learning, this technique has also been applied to distill teacher policies to students. Until now, policy distillation required access to a simulator or real world trajectories. In this paper we introduce a simulator-free approach to knowledge distillation in the context of reinforcement learning. A key challenge is having the student learn the multiplicity of cases that correspond to a given action. While prior work has shown that data-free knowledge distillation is possible with supervised learning models by generating synthetic examples, these approaches to are vulnerable to only producing a single prototype example for each class. We propose an extension to explicitly handle multiple observations per output class that seeks to find as many exemplars as possible for a given output class by reinitializing our data generator and making use of an adversarial loss. To the best of our knowledge, this is the first demonstration of simulator-free knowledge distillation between a teacher and a student policy. This new approach improves over the state of the art on data-free learning of student networks on benchmark datasets (MNIST, Fashion-MNIST, CIFAR-10), and we also demonstrate that it specifically tackles issues with multiple input modes. We also identify open problems when distilling agents trained in high dimensional environments such as Pong, Breakout, or Seaquest.

arXiv:2011.11108공식 코드 · Papers with Code

Multiresolution Knowledge Distillation for Anomaly Detection

Unsupervised representation learning has proved to be a critical component of anomaly detection/localization in images. The challenges to learn such a representation are two-fold. Firstly, the sample size is not often large enough to learn a rich generalizable representation through conventional techniques. Secondly, while only normal samples are available at training, the learned features should be discriminative of normal and anomalous samples. Here, we propose to use the "distillation" of features at various layers of an expert network, pre-trained on ImageNet, into a simpler cloner network to tackle both issues. We detect and localize anomalies using the discrepancy between the expert and cloner networks' intermediate activation values given the input data. We show that considering multiple intermediate hints in distillation leads to better exploiting the expert's knowledge and more distinctive discrepancy compared to solely utilizing the last layer activation values. Notably, previous methods either fail in precise anomaly localization or need expensive region-based training. In contrast, with no need for any special or intensive training procedure, we incorporate interpretability algorithms in our novel framework for the localization of anomalous regions. Despite the striking contrast between some test datasets and ImageNet, we achieve competitive or significantly superior results compared to the SOTA methods on MNIST, F-MNIST, CIFAR-10, MVTecAD, Retinal-OCT, and two Medical datasets on both anomaly detection and localization.

arXiv:2011.10904공식 코드 · Papers with Code

Evolving Search Space for Neural Architecture Search

The automation of neural architecture design has been a coveted alternative to human experts. Recent works have small search space, which is easier to optimize but has a limited upper bound of the optimal solution. Extra human design is needed for those methods to propose a more suitable space with respect to the specific task and algorithm capacity. To further enhance the degree of automation for neural architecture search, we present a Neural Search-space Evolution (NSE) scheme that iteratively amplifies the results from the previous effort by maintaining an optimized search space subset. This design minimizes the necessity of a well-designed search space. We further extend the flexibility of obtainable architectures by introducing a learnable multi-branch setting. By employing the proposed method, a consistent performance gain is achieved during a progressive search over upcoming search spaces. We achieve 77.3% top-1 retrain accuracy on ImageNet with 333M FLOPs, which yielded a state-of-the-art performance among previous auto-generated architectures that do not involve knowledge distillation or weight pruning. When the latency constraint is adopted, our result also performs better than the previous best-performing mobile models with a 77.9% Top-1 retrain accuracy.

arXiv:2011.09899

MixMix: All You Need for Data-Free Compression Are Feature and Data Mixing

User data confidentiality protection is becoming a rising challenge in the present deep learning research. Without access to data, conventional data-driven model compression faces a higher risk of performance degradation. Recently, some works propose to generate images from a specific pretrained model to serve as training data. However, the inversion process only utilizes biased feature statistics stored in one model and is from low-dimension to high-dimension. As a consequence, it inevitably encounters the difficulties of generalizability and inexact inversion, which leads to unsatisfactory performance. To address these problems, we propose MixMix based on two simple yet effective techniques: (1) Feature Mixing: utilizes various models to construct a universal feature space for generalized inversion; (2) Data Mixing: mixes the synthesized images and labels to generate exact label information. We prove the effectiveness of MixMix from both theoretical and empirical perspectives. Extensive experiments show that MixMix outperforms existing methods on the mainstream compression tasks, including quantization, knowledge distillation, and pruning. Specifically, MixMix achieves up to 4% and 20% accuracy uplift on quantization and pruning, respectively, compared to existing data-free compression work.

arXiv:2011.09757공식 코드 · Papers with Code

KD3A: Unsupervised Multi-Source Decentralized Domain Adaptation via Knowledge Distillation

Conventional unsupervised multi-source domain adaptation (UMDA) methods assume all source domains can be accessed directly. This neglects the privacy-preserving policy, that is, all the data and computations must be kept decentralized. There exists three problems in this scenario: (1) Minimizing the domain distance requires the pairwise calculation of the data from source and target domains, which is not accessible. (2) The communication cost and privacy security limit the application of UMDA methods (e.g., the domain adversarial training). (3) Since users have no authority to check the data quality, the irrelevant or malicious source domains are more likely to appear, which causes negative transfer. In this study, we propose a privacy-preserving UMDA paradigm named Knowledge Distillation based Decentralized Domain Adaptation (KD3A), which performs domain adaptation through the knowledge distillation on models from different source domains. KD3A solves the above problems with three components: (1) A multi-source knowledge distillation method named Knowledge Vote to learn high-quality domain consensus knowledge. (2) A dynamic weighting strategy named Consensus Focus to identify both the malicious and irrelevant domains. (3) A decentralized optimization strategy for domain distance named BatchNorm MMD. The extensive experiments on DomainNet demonstrate that KD3A is robust to the negative transfer and brings a 100x reduction of communication cost compared with other decentralized UMDA methods. Moreover, our KD3A significantly outperforms state-of-the-art UMDA approaches.

arXiv:2011.09361공식 코드 · Papers with Code

A Knowledge Distillation Ensemble Framework for Predicting Short and Long-term Hospitalisation Outcomes from Electronic Health Records Data

The ability to perform accurate prognosis of patients is crucial for proactive clinical decision making, informed resource management and personalised care. Existing outcome prediction models suffer from a low recall of infrequent positive outcomes. We present a highly-scalable and robust machine learning framework to automatically predict adversity represented by mortality and ICU admission from time-series vital signs and laboratory results obtained within the first 24 hours of hospital admission. The stacked platform comprises two components: a) an unsupervised LSTM Autoencoder that learns an optimal representation of the time-series, using it to differentiate the less frequent patterns which conclude with an adverse event from the majority patterns that do not, and b) a gradient boosting model, which relies on the constructed representation to refine prediction, incorporating static features of demographics, admission details and clinical summaries. The model is used to assess a patient's risk of adversity over time and provides visual justifications of its prediction based on the patient's static features and dynamic signals. Results of three case studies for predicting mortality and ICU admission show that the model outperforms all existing outcome prediction models, achieving PR-AUC of 0.891 (95$%$ CI: 0.878 - 0.969) in predicting mortality in ICU and general ward settings and 0.908 (95$%$ CI: 0.870-0.935) in predicting ICU admission.

arXiv:2011.09158

Privileged Knowledge Distillation for Online Action Detection

Online Action Detection (OAD) in videos is proposed as a per-frame labeling task to address the real-time prediction tasks that can only obtain the previous and current video frames. This paper presents a novel learning-with-privileged based framework for online action detection where the future frames only observable at the training stages are considered as a form of privileged information. Knowledge distillation is employed to transfer the privileged information from the offline teacher to the online student. We note that this setting is different from conventional KD because the difference between the teacher and student models mostly lies in input data rather than the network architecture. We propose Privileged Knowledge Distillation (PKD) which (i) schedules a curriculum learning procedure and (ii) inserts auxiliary nodes to the student model, both for shrinking the information gap and improving learning performance. Compared to other OAD methods that explicitly predict future frames, our approach avoids learning unpredictable unnecessary yet inconsistent visual contents and achieves state-of-the-art accuracy on two popular OAD benchmarks, TVSeries and THUMOS14.

arXiv:2011.09113

Effectiveness of Arbitrary Transfer Sets for Data-free Knowledge Distillation

Knowledge Distillation is an effective method to transfer the learning across deep neural networks. Typically, the dataset originally used for training the Teacher model is chosen as the "Transfer Set" to conduct the knowledge transfer to the Student. However, this original training data may not always be freely available due to privacy or sensitivity concerns. In such scenarios, existing approaches either iteratively compose a synthetic set representative of the original training dataset, one sample at a time or learn a generative model to compose such a transfer set. However, both these approaches involve complex optimization (GAN training or several backpropagation steps to synthesize one sample) and are often computationally expensive. In this paper, as a simple alternative, we investigate the effectiveness of "arbitrary transfer sets" such as random noise, publicly available synthetic, and natural datasets, all of which are completely unrelated to the original training dataset in terms of their visual or semantic contents. Through extensive experiments on multiple benchmark datasets such as MNIST, FMNIST, CIFAR-10 and CIFAR-100, we discover and validate surprising effectiveness of using arbitrary data to conduct knowledge distillation when this dataset is "target-class balanced". We believe that this important observation can potentially lead to designing baselines for the data-free knowledge distillation task.

arXiv:2011.08960

Deep Serial Number: Computational Watermarking for DNN Intellectual Property Protection

In this paper, we present DSN (Deep Serial Number), a simple yet effective watermarking algorithm designed specifically for deep neural networks (DNNs). Unlike traditional methods that incorporate identification signals into DNNs, our approach explores a novel Intellectual Property (IP) protection mechanism for DNNs, effectively thwarting adversaries from using stolen networks. Inspired by the success of serial numbers in safeguarding conventional software IP, we propose the first implementation of serial number embedding within DNNs. To achieve this, DSN is integrated into a knowledge distillation framework, in which a private teacher DNN is initially trained. Subsequently, its knowledge is distilled and imparted to a series of customized student DNNs. Each customer DNN functions correctly only upon input of a valid serial number. Experimental results across various applications demonstrate DSN's efficacy in preventing unauthorized usage without compromising the original DNN performance. The experiments further show that DSN is resistant to different categories of watermark attacks.

arXiv:2011.08508

Generalized Continual Zero-Shot Learning

Recently, zero-shot learning (ZSL) emerged as an exciting topic and attracted a lot of attention. ZSL aims to classify unseen classes by transferring the knowledge from seen classes to unseen classes based on the class description. Despite showing promising performance, ZSL approaches assume that the training samples from all seen classes are available during the training, which is practically not feasible. To address this issue, we propose a more generalized and practical setup for ZSL, i.e., continual ZSL (CZSL), where classes arrive sequentially in the form of a task and it actively learns from the changing environment by leveraging the past experience. Further, to enhance the reliability, we develop CZSL for a single head continual learning setting where task identity is revealed during the training process but not during the testing. To avoid catastrophic forgetting and intransigence, we use knowledge distillation and storing and replay the few samples from previous tasks using a small episodic memory. We develop baselines and evaluate generalized CZSL on five ZSL benchmark datasets for two different settings of continual learning: with and without class incremental. Moreover, CZSL is developed for two types of variational autoencoders, which generates two types of features for classification: (i) generated features at output space and (ii) generated discriminative features at the latent space. The experimental results clearly indicate the single head CZSL is more generalizable and suitable for practical applications.

arXiv:2011.08505

Digging Deeper into CRNN Model in Chinese Text Images Recognition

Automatic text image recognition is a prevalent application in computer vision field. One efficient way is use Convolutional Recurrent Neural Network(CRNN) to accomplish task in an end-to-end(End2End) fashion. However, CRNN notoriously fails to detect multi-row images and excel-like images. In this paper, we present one alternative to first recognize single-row images, then extend the same architecture to recognize multi-row images with proposed multiple methods. To recognize excel-like images containing box lines, we propose Line-Deep Denoising Convolutional AutoEncoder(Line-DDeCAE) to recover box lines. Finally, we present one Knowledge Distillation(KD) method to compress original CRNN model without loss of generality. To carry out experiments, we first generate artificial samples from one Chinese novel book, then conduct various experiments to verify our methods.

arXiv:2011.08007

Domain Adaptive Knowledge Distillation for Driving Scene Semantic Segmentation

Practical autonomous driving systems face two crucial challenges: memory constraints and domain gap issues. In this paper, we present a novel approach to learn domain adaptive knowledge in models with limited memory, thus bestowing the model with the ability to deal with these issues in a comprehensive manner. We term this as "Domain Adaptive Knowledge Distillation" and address the same in the context of unsupervised domain-adaptive semantic segmentation by proposing a multi-level distillation strategy to effectively distil knowledge at different levels. Further, we introduce a novel cross entropy loss that leverages pseudo labels from the teacher. These pseudo teacher labels play a multifaceted role towards: (i) knowledge distillation from the teacher network to the student network & (ii) serving as a proxy for the ground truth for target domain images, where the problem is completely unsupervised. We introduce four paradigms for distilling domain adaptive knowledge and carry out extensive experiments and ablation studies on real-to-real as well as synthetic-to-real scenarios. Our experiments demonstrate the profound success of our proposed method.

arXiv:2011.07491

Anomaly Detection in Video via Self-Supervised and Multi-Task Learning

Anomaly detection in video is a challenging computer vision problem. Due to the lack of anomalous events at training time, anomaly detection requires the design of learning methods without full supervision. In this paper, we approach anomalous event detection in video through self-supervised and multi-task learning at the object level. We first utilize a pre-trained detector to detect objects. Then, we train a 3D convolutional neural network to produce discriminative anomaly-specific information by jointly learning multiple proxy tasks: three self-supervised and one based on knowledge distillation. The self-supervised tasks are: (i) discrimination of forward/backward moving objects (arrow of time), (ii) discrimination of objects in consecutive/intermittent frames (motion irregularity) and (iii) reconstruction of object-specific appearance information. The knowledge distillation task takes into account both classification and detection information, generating large prediction discrepancies between teacher and student models when anomalies occur. To the best of our knowledge, we are the first to approach anomalous event detection in video as a multi-task learning problem, integrating multiple self-supervised and knowledge distillation proxy tasks in a single architecture. Our lightweight architecture outperforms the state-of-the-art methods on three benchmarks: Avenue, ShanghaiTech and UCSD Ped2. Additionally, we perform an ablation study demonstrating the importance of integrating self-supervised learning and normality-specific distillation in a multi-task learning setting.

arXiv:2011.07449

Online Ensemble Model Compression using Knowledge Distillation

This paper presents a novel knowledge distillation based model compression framework consisting of a student ensemble. It enables distillation of simultaneously learnt ensemble knowledge onto each of the compressed student models. Each model learns unique representations from the data distribution due to its distinct architecture. This helps the ensemble generalize better by combining every model's knowledge. The distilled students and ensemble teacher are trained simultaneously without requiring any pretrained weights. Moreover, our proposed method can deliver multi-compressed students with single training, which is efficient and flexible for different scenarios. We provide comprehensive experiments using state-of-the-art classification models to validate our framework's effectiveness. Notably, using our framework a 97% compressed ResNet110 student model managed to produce a 10.64% relative accuracy gain over its individual baseline training on CIFAR100 dataset. Similarly a 95% compressed DenseNet-BC(k=12) model managed a 8.17% relative accuracy gain.

arXiv:2011.05961공식 코드 · Papers with Code

Real-Time Decentralized knowledge Transfer at the Edge

The proliferation of edge networks creates islands of learning agents working on local streams of data. Transferring knowledge between these agents in real-time without exposing private data allows for collaboration to decrease learning time and increase model confidence. Incorporating knowledge from data that a local model did not see creates an ability to debias a local model or add to classification abilities on data never before seen. Transferring knowledge in a selective decentralized approach enables models to retain their local insights, allowing for local flavors of a machine learning model. This approach suits the decentralized architecture of edge networks, as a local edge node will serve a community of learning agents that will likely encounter similar data. We propose a method based on knowledge distillation for pairwise knowledge transfer pipelines from models trained on non-i.i.d. data and compare it to other popular knowledge transfer methods. Additionally, we test different scenarios of knowledge transfer network construction and show the practicality of our approach. Our experiments show knowledge transfer using our model outperforms standard methods in a real-time transfer scenario.

arXiv:2011.05705공식 코드 · Papers with Code

EGAD: Evolving Graph Representation Learning with Self-Attention and Knowledge Distillation for Live Video Streaming Events

In this study, we present a dynamic graph representation learning model on weighted graphs to accurately predict the network capacity of connections between viewers in a live video streaming event. We propose EGAD, a neural network architecture to capture the graph evolution by introducing a self-attention mechanism on the weights between consecutive graph convolutional networks. In addition, we account for the fact that neural architectures require a huge amount of parameters to train, thus increasing the online inference latency and negatively influencing the user experience in a live video streaming event. To address the problem of the high online inference of a vast number of parameters, we propose a knowledge distillation strategy. In particular, we design a distillation loss function, aiming to first pretrain a teacher model on offline data, and then transfer the knowledge from the teacher to a smaller student model with less parameters. We evaluate our proposed model on the link prediction task on three real-world datasets, generated by live video streaming events. The events lasted 80 minutes and each viewer exploited the distribution solution provided by the company Hive Streaming AB. The experiments demonstrate the effectiveness of the proposed model in terms of link prediction accuracy and number of required parameters, when evaluated against state-of-the-art approaches. In addition, we study the distillation performance of the proposed model in terms of compression ratio for different distillation strategies, where we show that the proposed model can achieve a compression ratio up to 15:100, preserving high link prediction accuracy. For reproduction purposes, our evaluation datasets and implementation are publicly available at https://stefanosantaris.github.io/EGAD.