PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2201.02624

Microdosing: Knowledge Distillation for GAN based Compression

Recently, significant progress has been made in learned image and video compression. In particular the usage of Generative Adversarial Networks has lead to impressive results in the low bit rate regime. However, the model size remains an important issue in current state-of-the-art proposals and existing solutions require significant computation effort on the decoding side. This limits their usage in realistic scenarios and the extension to video compression. In this paper, we demonstrate how to leverage knowledge distillation to obtain equally capable image decoders at a fraction of the original number of parameters. We investigate several aspects of our solution including sequence specialization with side information for image coding. Finally, we also show how to transfer the obtained benefits into the setting of video compression. Overall, this allows us to reduce the model size by a factor of 20 and to achieve 50% reduction in decoding time.

arXiv:2201.03019공식 코드 · Papers with Code

Robust and Resource-Efficient Data-Free Knowledge Distillation by Generative Pseudo Replay

Data-Free Knowledge Distillation (KD) allows knowledge transfer from a trained neural network (teacher) to a more compact one (student) in the absence of original training data. Existing works use a validation set to monitor the accuracy of the student over real data and report the highest performance throughout the entire process. However, validation data may not be available at distillation time either, making it infeasible to record the student snapshot that achieved the peak accuracy. Therefore, a practical data-free KD method should be robust and ideally provide monotonically increasing student accuracy during distillation. This is challenging because the student experiences knowledge degradation due to the distribution shift of the synthetic data. A straightforward approach to overcome this issue is to store and rehearse the generated samples periodically, which increases the memory footprint and creates privacy concerns. We propose to model the distribution of the previously observed synthetic samples with a generative network. In particular, we design a Variational Autoencoder (VAE) with a training objective that is customized to learn the synthetic data representations optimally. The student is rehearsed by the generative pseudo replay technique, with samples produced by the VAE. Hence knowledge degradation can be prevented without storing any samples. Experiments on image classification benchmarks show that our method optimizes the expected value of the distilled model accuracy while eliminating the large memory overhead incurred by the sample-storing methods.

arXiv:2112.12612공식 코드 · Papers with Code

Towards Disturbance-Free Visual Mobile Manipulation

Deep reinforcement learning has shown promising results on an abundance of robotic tasks in simulation, including visual navigation and manipulation. Prior work generally aims to build embodied agents that solve their assigned tasks as quickly as possible, while largely ignoring the problems caused by collision with objects during interaction. This lack of prioritization is understandable: there is no inherent cost in breaking virtual objects. As a result, "well-trained" agents frequently collide with objects before achieving their primary goals, a behavior that would be catastrophic in the real world. In this paper, we study the problem of training agents to complete the task of visual mobile manipulation in the ManipulaTHOR environment while avoiding unnecessary collision (disturbance) with objects. We formulate disturbance avoidance as a penalty term in the reward function, but find that directly training with such penalized rewards often results in agents being unable to escape poor local optima. Instead, we propose a two-stage training curriculum where an agent is first allowed to freely explore and build basic competencies without penalization, after which a disturbance penalty is introduced to refine the agent's behavior. Results on testing scenes show that our curriculum not only avoids these poor local optima, but also leads to 10% absolute gains in success rate without disturbance, compared to our state-of-the-art baselines. Moreover, our curriculum is significantly more performant than a safe RL algorithm that casts collision avoidance as a constraint. Finally, we propose a novel disturbance-prediction auxiliary task that accelerates learning.

arXiv:2112.08670

Amortized Noisy Channel Neural Machine Translation

Noisy channel models have been especially effective in neural machine translation (NMT). However, recent approaches like "beam search and rerank" (BSR) incur significant computation overhead during inference, making real-world application infeasible. We aim to study if it is possible to build an amortized noisy channel NMT model such that when we do greedy decoding during inference, the translation accuracy matches that of BSR in terms of reward (based on the source-to-target log probability and the target-to-source log probability) and quality (based on BLEU and BLEURT). We attempt three approaches to train the new model: knowledge distillation, one-step-deviation imitation learning, and Q learning. The first approach obtains the noisy channel signal from a pseudo-corpus, and the latter two approaches aim to optimize toward a noisy-channel MT reward directly. For all three approaches, the generated translations fail to achieve rewards comparable to BSR, but the translation quality approximated by BLEU and BLEURT is similar to the quality of BSR-produced translations. Additionally, all three approaches speed up inference by 1-2 orders of magnitude.

arXiv:2201.00558

Which Student is Best? A Comprehensive Knowledge Distillation Exam for Task-Specific BERT Models

We perform knowledge distillation (KD) benchmark from task-specific BERT-base teacher models to various student models: BiLSTM, CNN, BERT-Tiny, BERT-Mini, and BERT-Small. Our experiment involves 12 datasets grouped in two tasks: text classification and sequence labeling in the Indonesian language. We also compare various aspects of distillations including the usage of word embeddings and unlabeled data augmentation. Our experiments show that, despite the rising popularity of Transformer-based models, using BiLSTM and CNN student models provide the best trade-off between performance and computational resource (CPU, RAM, and storage) compared to pruned BERT models. We further propose some quick wins on performing KD to produce small NLP models via efficient KD training mechanisms involving simple choices of loss functions, word embeddings, and unlabeled data preparation.

arXiv:2201.00766공식 코드 · Papers with Code

Class-Incremental Continual Learning into the eXtended DER-verse

The staple of human intelligence is the capability of acquiring knowledge in a continuous fashion. In stark contrast, Deep Networks forget catastrophically and, for this reason, the sub-field of Class-Incremental Continual Learning fosters methods that learn a sequence of tasks incrementally, blending sequentially-gained knowledge into a comprehensive prediction. This work aims at assessing and overcoming the pitfalls of our previous proposal Dark Experience Replay (DER), a simple and effective approach that combines rehearsal and Knowledge Distillation. Inspired by the way our minds constantly rewrite past recollections and set expectations for the future, we endow our model with the abilities to i) revise its replay memory to welcome novel information regarding past data ii) pave the way for learning yet unseen classes. We show that the application of these strategies leads to remarkable improvements; indeed, the resulting method - termed eXtended-DER (X-DER) - outperforms the state of the art on both standard benchmarks (such as CIFAR-100 and miniImagenet) and a novel one here introduced. To gain a better understanding, we further provide extensive ablation studies that corroborate and extend the findings of our previous research (e.g. the value of Knowledge Distillation and flatter minima in continual learning setups).

arXiv:2201.00111

Role of Data Augmentation Strategies in Knowledge Distillation for Wearable Sensor Data

Deep neural networks are parametrized by several thousands or millions of parameters, and have shown tremendous success in many classification problems. However, the large number of parameters makes it difficult to integrate these models into edge devices such as smartphones and wearable devices. To address this problem, knowledge distillation (KD) has been widely employed, that uses a pre-trained high capacity network to train a much smaller network, suitable for edge devices. In this paper, for the first time, we study the applicability and challenges of using KD for time-series data for wearable devices. Successful application of KD requires specific choices of data augmentation methods during training. However, it is not yet known if there exists a coherent strategy for choosing an augmentation approach during KD. In this paper, we report the results of a detailed study that compares and contrasts various common choices and some hybrid data augmentation strategies in KD based human activity analysis. Research in this area is often limited as there are not many comprehensive databases available in the public domain from wearable devices. Our study considers databases from small scale publicly available to one derived from a large scale interventional study into human activity and sedentary behavior. We find that the choice of data augmentation techniques during KD have a variable level of impact on end performance, and find that the optimal network choice as well as data augmentation strategies are specific to a dataset at hand. However, we also conclude with a general set of recommendations that can provide a strong baseline performance across databases.

arXiv:2112.15358

Conditional Generative Data-free Knowledge Distillation

Knowledge distillation has made remarkable achievements in model compression. However, most existing methods require the original training data, which is usually unavailable due to privacy and security issues. In this paper, we propose a conditional generative data-free knowledge distillation (CGDD) framework for training lightweight networks without any training data. This method realizes efficient knowledge distillation based on conditional image generation. Specifically, we treat the preset labels as ground truth to train a conditional generator in a semi-supervised manner. The trained generator can produce specified classes of training images. For training the student network, we force it to extract the knowledge hidden in teacher feature maps, which provide crucial cues for the learning process. Moreover, an adversarial training framework for promoting distillation performance is constructed by designing several loss functions. This framework helps the student model to explore larger data space. To demonstrate the effectiveness of the proposed method, we conduct extensive experiments on different datasets. Compared with other data-free works, our work obtains state-of-the-art results on CIFAR100, Caltech101, and different versions of ImageNet datasets. The codes will be released.

arXiv:2112.15278

Data-Free Knowledge Transfer: A Survey

In the last decade, many deep learning models have been well trained and made a great success in various fields of machine intelligence, especially for computer vision and natural language processing. To better leverage the potential of these well-trained models in intra-domain or cross-domain transfer learning situations, knowledge distillation (KD) and domain adaptation (DA) are proposed and become research highlights. They both aim to transfer useful information from a well-trained model with original training data. However, the original data is not always available in many cases due to privacy, copyright or confidentiality. Recently, the data-free knowledge transfer paradigm has attracted appealing attention as it deals with distilling valuable knowledge from well-trained models without requiring to access to the training data. In particular, it mainly consists of the data-free knowledge distillation (DFKD) and source data-free domain adaptation (SFDA). On the one hand, DFKD aims to transfer the intra-domain knowledge of original data from a cumbersome teacher network to a compact student network for model compression and efficient inference. On the other hand, the goal of SFDA is to reuse the cross-domain knowledge stored in a well-trained source model and adapt it to a target domain. In this paper, we provide a comprehensive survey on data-free knowledge transfer from the perspectives of knowledge distillation and unsupervised domain adaptation, to help readers have a better understanding of the current research status and ideas. Applications and challenges of the two areas are briefly reviewed, respectively. Furthermore, we provide some insights to the subject of future research.

arXiv:2201.00187

Image Restoration using Feature-guidance

Image restoration is the task of recovering a clean image from a degraded version. In most cases, the degradation is spatially varying, and it requires the restoration network to both localize and restore the affected regions. In this paper, we present a new approach suitable for handling the image-specific and spatially-varying nature of degradation in images affected by practically occurring artifacts such as blur, rain-streaks. We decompose the restoration task into two stages of degradation localization and degraded region-guided restoration, unlike existing methods which directly learn a mapping between the degraded and clean images. Our premise is to use the auxiliary task of degradation mask prediction to guide the restoration process. We demonstrate that the model trained for this auxiliary task contains vital region knowledge, which can be exploited to guide the restoration network's training using attentive knowledge distillation technique. Further, we propose mask-guided convolution and global context aggregation module that focuses solely on restoring the degraded regions. The proposed approach's effectiveness is demonstrated by achieving significant improvement over strong baselines.

arXiv:2112.10063공식 코드 · Papers with Code

Deep Graph-level Anomaly Detection by Glocal Knowledge Distillation

Graph-level anomaly detection (GAD) describes the problem of detecting graphs that are abnormal in their structure and/or the features of their nodes, as compared to other graphs. One of the challenges in GAD is to devise graph representations that enable the detection of both locally- and globally-anomalous graphs, i.e., graphs that are abnormal in their fine-grained (node-level) or holistic (graph-level) properties, respectively. To tackle this challenge we introduce a novel deep anomaly detection approach for GAD that learns rich global and local normal pattern information by joint random distillation of graph and node representations. The random distillation is achieved by training one GNN to predict another GNN with randomly initialized network weights. Extensive experiments on 16 real-world graph datasets from diverse domains show that our model significantly outperforms seven state-of-the-art models. Code and datasets are available at https://git.io/GLocalKD.

arXiv:2201.00011

An Efficient Federated Distillation Learning System for Multi-task Time Series Classification

This paper proposes an efficient federated distillation learning system (EFDLS) for multi-task time series classification (TSC). EFDLS consists of a central server and multiple mobile users, where different users may run different TSC tasks. EFDLS has two novel components, namely a feature-based student-teacher (FBST) framework and a distance-based weights matching (DBWM) scheme. Within each user, the FBST framework transfers knowledge from its teacher's hidden layers to its student's hidden layers via knowledge distillation, with the teacher and student having identical network structure. For each connected user, its student model's hidden layers' weights are uploaded to the EFDLS server periodically. The DBWM scheme is deployed on the server, with the least square distance used to measure the similarity between the weights of two given models. This scheme finds a partner for each connected user such that the user's and its partner's weights are the closest among all the weights uploaded. The server exchanges and sends back the user's and its partner's weights to these two users which then load the received weights to their teachers' hidden layers. Experimental results show that the proposed EFDLS achieves excellent performance on a set of selected UCR2018 datasets regarding top-1 accuracy.

arXiv:2201.00007공식 코드 · Papers with Code

Confidence-Aware Multi-Teacher Knowledge Distillation

Knowledge distillation is initially introduced to utilize additional supervision from a single teacher model for the student model training. To boost the student performance, some recent variants attempt to exploit diverse knowledge sources from multiple teachers. However, existing studies mainly integrate knowledge from diverse sources by averaging over multiple teacher predictions or combining them using other various label-free strategies, which may mislead student in the presence of low-quality teacher predictions. To tackle this problem, we propose Confidence-Aware Multi-teacher Knowledge Distillation (CA-MKD), which adaptively assigns sample-wise reliability for each teacher prediction with the help of ground-truth labels, with those teacher predictions close to one-hot labels assigned large weights. Besides, CA-MKD incorporates intermediate layers to stable the knowledge transfer process. Extensive experiments show that our CA-MKD consistently outperforms all compared state-of-the-art methods across various teacher-student architectures.

arXiv:2112.14938

Automatic Mixed-Precision Quantization Search of BERT

Pre-trained language models such as BERT have shown remarkable effectiveness in various natural language processing tasks. However, these models usually contain millions of parameters, which prevents them from practical deployment on resource-constrained devices. Knowledge distillation, Weight pruning, and Quantization are known to be the main directions in model compression. However, compact models obtained through knowledge distillation may suffer from significant accuracy drop even for a relatively small compression ratio. On the other hand, there are only a few quantization attempts that are specifically designed for natural language processing tasks. They suffer from a small compression ratio or a large error rate since manual setting on hyper-parameters is required and fine-grained subgroup-wise quantization is not supported. In this paper, we proposed an automatic mixed-precision quantization framework designed for BERT that can simultaneously conduct quantization and pruning in a subgroup-wise level. Specifically, our proposed method leverages Differentiable Neural Architecture Search to assign scale and precision for parameters in each sub-group automatically, and at the same time pruning out redundant groups of parameters. Extensive evaluations on BERT downstream tasks reveal that our proposed method outperforms baselines by providing the same performance with much smaller model size. We also show the feasibility of obtaining the extremely light-weight model by combining our solution with orthogonal methods such as DistilBERT.

arXiv:2112.13966공식 코드 · Papers with Code

Online Adversarial Knowledge Distillation for Graph Neural Networks

Knowledge distillation, a technique recently gaining popularity for enhancing model generalization in Convolutional Neural Networks (CNNs), operates under the assumption that both teacher and student models are trained on identical data distributions. However, its effect on Graph Neural Networks (GNNs) is less than satisfactory since the graph topology and node attributes are prone to evolve, thereby leading to the issue of distribution shift. In this paper, we tackle this challenge by simultaneously training a group of graph neural networks in an online distillation fashion, where the group knowledge plays a role as a dynamic virtual teacher and the structure changes in graph neural networks are effectively captured. To improve the distillation performance, two types of knowledge are transferred among the students to enhance each other: local knowledge reflecting information in the graph topology and node attributes, and global knowledge reflecting the prediction over classes. We transfer the global knowledge with KL-divergence as the vanilla knowledge distillation does, while exploiting the complicated structure of the local knowledge with an efficient adversarial cyclic learning framework. Extensive experiments verified the effectiveness of our proposed online adversarial distillation approach. The code is published at https://github.com/wangz3066/OnlineDistillGCN.

arXiv:2112.13642

Extracting knowledge from features with multilevel abstraction

Knowledge distillation aims at transferring the knowledge from a large teacher model to a small student model with great improvements of the performance of the student model. Therefore, the student network can replace the teacher network to deploy on low-resource devices since the higher performance, lower number of parameters and shorter inference time. Self-knowledge distillation (SKD) attracts a great attention recently that a student model itself is a teacher model distilling knowledge from. To the best of our knowledge, self knowledge distillation can be divided into two main streams: data augmentation and refined knowledge auxiliary. In this paper, we purpose a novel SKD method in a different way from the main stream methods. Our method distills knowledge from multilevel abstraction features. Experiments and ablation studies show its great effectiveness and generalization on various kinds of tasks with various kinds of model structures. Our codes have been released on GitHub.

arXiv:2112.11389

Supervised Graph Contrastive Pretraining for Text Classification

Contrastive pretraining techniques for text classification has been largely studied in an unsupervised setting. However, oftentimes labeled data from related tasks which share label semantics with current task is available. We hypothesize that using this labeled data effectively can lead to better generalization on current task. In this paper, we propose a novel way to effectively utilize labeled data from related tasks with a graph based supervised contrastive learning approach. We formulate a token-graph by extrapolating the supervised information from examples to tokens. Our formulation results in an embedding space where tokens with high/low probability of belonging to same class are near/further-away from one another. We also develop detailed theoretical insights which serve as a motivation for our method. In our experiments with $13$ datasets, we show our method outperforms pretraining schemes by $2.5\%$ and also example-level contrastive learning based formulation by $1.8\%$ on average. In addition, we show cross-domain effectiveness of our method in a zero-shot setting by $3.91\%$ on average. Lastly, we also demonstrate our method can be used as a noisy teacher in a knowledge distillation setting to significantly improve performance of transformer based models in low labeled data regime by $4.57\%$ on average.

arXiv:2112.10047

Controlling the Quality of Distillation in Response-Based Network Compression

The performance of a distillation-based compressed network is governed by the quality of distillation. The reason for the suboptimal distillation of a large network (teacher) to a smaller network (student) is largely attributed to the gap in the learning capacities of given teacher-student pair. While it is hard to distill all the knowledge of a teacher, the quality of distillation can be controlled to a large extent to achieve better performance. Our experiments show that the quality of distillation is largely governed by the quality of teacher's response, which in turn is heavily affected by the presence of similarity information in its response. A well-trained large capacity teacher loses similarity information between classes in the process of learning fine-grained discriminative properties for classification. The absence of similarity information causes the distillation process to be reduced from one example-many class learning to one example-one class learning, thereby throttling the flow of diverse knowledge from the teacher. With the implicit assumption that only the instilled knowledge can be distilled, instead of focusing only on the knowledge distilling process, we scrutinize the knowledge inculcation process. We argue that for a given teacher-student pair, the quality of distillation can be improved by finding the sweet spot between batch size and number of epochs while training the teacher. We discuss the steps to find this sweet spot for better distillation. We also propose the distillation hypothesis to differentiate the behavior of the distillation process between knowledge distillation and regularization effect. We conduct all our experiments on three different datasets.

arXiv:2112.12650공식 코드 · Papers with Code

Distilling the Knowledge of Romanian BERTs Using Multiple Teachers

Running large-scale pre-trained language models in computationally constrained environments remains a challenging problem yet to be addressed, while transfer learning from these models has become prevalent in Natural Language Processing tasks. Several solutions, including knowledge distillation, network quantization, or network pruning have been previously proposed; however, these approaches focus mostly on the English language, thus widening the gap when considering low-resource languages. In this work, we introduce three light and fast versions of distilled BERT models for the Romanian language: Distil-BERT-base-ro, Distil-RoBERT-base, and DistilMulti-BERT-base-ro. The first two models resulted from the individual distillation of knowledge from two base versions of Romanian BERTs available in literature, while the last one was obtained by distilling their ensemble. To our knowledge, this is the first attempt to create publicly available Romanian distilled BERT models, which were thoroughly evaluated on five tasks: part-of-speech tagging, named entity recognition, sentiment analysis, semantic textual similarity, and dialect identification. Our experimental results argue that the three distilled models offer performance comparable to their teachers, while being twice as fast on a GPU and ~35% smaller. In addition, we further test the similarity between the predictions of our students versus their teachers by measuring their label and probability loyalty, together with regression loyalty - a new metric introduced in this work.

arXiv:2112.11640

Self-Distillation Mixup Training for Non-autoregressive Neural Machine Translation

Recently, non-autoregressive (NAT) models predict outputs in parallel, achieving substantial improvements in generation speed compared to autoregressive (AT) models. While performing worse on raw data, most NAT models are trained as student models on distilled data generated by AT teacher models, which is known as sequence-level Knowledge Distillation. An effective training strategy to improve the performance of AT models is Self-Distillation Mixup (SDM) Training, which pre-trains a model on raw data, generates distilled data by the pre-trained model itself and finally re-trains a model on the combination of raw data and distilled data. In this work, we aim to view SDM for NAT models, but find directly adopting SDM to NAT models gains no improvements in terms of translation quality. Through careful analysis, we observe the invalidation is correlated to Modeling Diversity and Confirmation Bias between the AT teacher model and the NAT student models. Based on these findings, we propose an enhanced strategy named SDMRT by adding two stages to classic SDM: one is Pre-Rerank on self-distilled data, the other is Fine-Tune on Filtered teacher-distilled data. Our results outperform baselines by 0.6 to 1.2 BLEU on multiple NAT models. As another bonus, for Iterative Refinement NAT models, our methods can outperform baselines within half iteration number, which means 2X acceleration.