PyTorchKR Papers

최신 논문

총 4,526편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2207.08063

Subclass Knowledge Distillation with Known Subclass Labels

This work introduces a novel knowledge distillation framework for classification tasks where information on existing subclasses is available and taken into consideration. In classification tasks with a small number of classes or binary detection, the amount of information transferred from the teacher to the student is restricted, thus limiting the utility of knowledge distillation. Performance can be improved by leveraging information of possible subclasses within the classes. To that end, we propose the so-called Subclass Knowledge Distillation (SKD), a process of transferring the knowledge of predicted subclasses from a teacher to a smaller student. Meaningful information that is not in the teacher's class logits but exists in subclass logits (e.g., similarities within classes) will be conveyed to the student through the SKD, which will then boost the student's performance. Analytically, we measure how much extra information the teacher can provide the student via the SKD to demonstrate the efficacy of our work. The framework developed is evaluated in clinical application, namely colorectal polyp binary classification. It is a practical problem with two classes and a number of subclasses per class. In this application, clinician-provided annotations are used to define subclasses based on the annotation label's variability in a curriculum style of learning. A lightweight, low-complexity student trained with the SKD framework achieves an F1-score of 85.05%, an improvement of 1.47%, and a 2.10% gain over the student that is trained with and without conventional knowledge distillation, respectively. The 2.10% F1-score gap between students trained with and without the SKD can be explained by the extra subclass knowledge, i.e., the extra 0.4656 label bits per sample that the teacher can transfer in our experiment.

arXiv:2207.08003

SSMTL++: Revisiting Self-Supervised Multi-Task Learning for Video Anomaly Detection

A self-supervised multi-task learning (SSMTL) framework for video anomaly detection was recently introduced in literature. Due to its highly accurate results, the method attracted the attention of many researchers. In this work, we revisit the self-supervised multi-task learning framework, proposing several updates to the original method. First, we study various detection methods, e.g. based on detecting high-motion regions using optical flow or background subtraction, since we believe the currently used pre-trained YOLOv3 is suboptimal, e.g. objects in motion or objects from unknown classes are never detected. Second, we modernize the 3D convolutional backbone by introducing multi-head self-attention modules, inspired by the recent success of vision transformers. As such, we alternatively introduce both 2D and 3D convolutional vision transformer (CvT) blocks. Third, in our attempt to further improve the model, we study additional self-supervised learning tasks, such as predicting segmentation maps through knowledge distillation, solving jigsaw puzzles, estimating body pose through knowledge distillation, predicting masked regions (inpainting), and adversarial learning with pseudo-anomalies. We conduct experiments to assess the performance impact of the introduced changes. Upon finding more promising configurations of the framework, dubbed SSMTL++v1 and SSMTL++v2, we extend our preliminary experiments to more data sets, demonstrating that our performance gains are consistent across all data sets. In most cases, our results on Avenue, ShanghaiTech and UBnormal raise the state-of-the-art performance bar to a new level.

arXiv:2207.08605공식 코드 · Papers with Code

Class-incremental Novel Class Discovery

We study the new task of class-incremental Novel Class Discovery (class-iNCD), which refers to the problem of discovering novel categories in an unlabelled data set by leveraging a pre-trained model that has been trained on a labelled data set containing disjoint yet related categories. Apart from discovering novel classes, we also aim at preserving the ability of the model to recognize previously seen base categories. Inspired by rehearsal-based incremental learning methods, in this paper we propose a novel approach for class-iNCD which prevents forgetting of past information about the base classes by jointly exploiting base class feature prototypes and feature-level knowledge distillation. We also propose a self-training clustering strategy that simultaneously clusters novel categories and trains a joint classifier for both the base and novel classes. This makes our method able to operate in a class-incremental setting. Our experiments, conducted on three common benchmarks, demonstrate that our method significantly outperforms state-of-the-art approaches. Code is available at https://github.com/OatmealLiu/class-iNCD

arXiv:2207.08356

Learning Knowledge Representation with Meta Knowledge Distillation for Single Image Super-Resolution

Knowledge distillation (KD), which can efficiently transfer knowledge from a cumbersome network (teacher) to a compact network (student), has demonstrated its advantages in some computer vision applications. The representation of knowledge is vital for knowledge transferring and student learning, which is generally defined in hand-crafted manners or uses the intermediate features directly. In this paper, we propose a model-agnostic meta knowledge distillation method under the teacher-student architecture for the single image super-resolution task. It provides a more flexible and accurate way to help the teachers transmit knowledge in accordance with the abilities of students via knowledge representation networks (KRNets) with learnable parameters. In order to improve the perception ability of knowledge representation to students' requirements, we propose to solve the transformation process from intermediate outputs to transferred knowledge by employing the student features and the correlation between teacher and student in the KRNets. Specifically, the texture-aware dynamic kernels are generated and then extract texture features to be improved and the corresponding teacher guidance so as to decompose the distillation problem into texture-wise supervision for further promoting the recovery quality of high-frequency details. In addition, the KRNets are optimized in a meta-learning manner to ensure the knowledge transferring and the student learning are beneficial to improving the reconstructed quality of the student. Experiments conducted on various single image super-resolution datasets demonstrate that our proposed method outperforms existing defined knowledge representation related distillation methods, and can help super-resolution algorithms achieve better reconstruction quality without introducing any inference complexity.

arXiv:2207.08739공식 코드 · Papers with Code

Rethinking Data Augmentation for Robust Visual Question Answering

Data Augmentation (DA) -- generating extra training samples beyond original training set -- has been widely-used in today's unbiased VQA models to mitigate the language biases. Current mainstream DA strategies are synthetic-based methods, which synthesize new samples by either editing some visual regions/words, or re-generating them from scratch. However, these synthetic samples are always unnatural and error-prone. To avoid this issue, a recent DA work composes new augmented samples by randomly pairing pristine images and other human-written questions. Unfortunately, to guarantee augmented samples have reasonable ground-truth answers, they manually design a set of heuristic rules for several question types, which extremely limits its generalization abilities. To this end, we propose a new Knowledge Distillation based Data Augmentation for VQA, dubbed KDDAug. Specifically, we first relax the requirements of reasonable image-question pairs, which can be easily applied to any question types. Then, we design a knowledge distillation (KD) based answer assignment to generate pseudo answers for all composed image-question pairs, which are robust to both in-domain and out-of-distribution settings. Since KDDAug is a model-agnostic DA strategy, it can be seamlessly incorporated into any VQA architectures. Extensive ablation studies on multiple backbones and benchmarks have demonstrated the effectiveness and generalization abilities of KDDAug.

arXiv:2207.07564

Rethinking Attention Mechanism in Time Series Classification

Attention-based models have been widely used in many areas, such as computer vision and natural language processing. However, relevant applications in time series classification (TSC) have not been explored deeply yet, causing a significant number of TSC algorithms still suffer from general problems of attention mechanism, like quadratic complexity. In this paper, we promote the efficiency and performance of the attention mechanism by proposing our flexible multi-head linear attention (FMLA), which enhances locality awareness by layer-wise interactions with deformable convolutional blocks and online knowledge distillation. What's more, we propose a simple but effective mask mechanism that helps reduce the noise influence in time series and decrease the redundancy of the proposed FMLA by masking some positions of each given series proportionally. To stabilize this mechanism, samples are forwarded through the model with random mask layers several times and their outputs are aggregated to teach the same model with regular mask layers. We conduct extensive experiments on 85 UCR2018 datasets to compare our algorithm with 11 well-known ones and the results show that our algorithm has comparable performance in terms of top-1 accuracy. We also compare our model with three Transformer-based models with respect to the floating-point operations per second and number of parameters and find that our algorithm achieves significantly better efficiency with lower complexity.

arXiv:2207.06867

Deep versus Wide: An Analysis of Student Architectures for Task-Agnostic Knowledge Distillation of Self-Supervised Speech Models

Self-supervised learning (SSL) is seen as a very promising approach with high performance for several speech downstream tasks. Since the parameters of SSL models are generally so large that training and inference require a lot of memory and computational cost, it is desirable to produce compact SSL models without a significant performance degradation by applying compression methods such as knowledge distillation (KD). Although the KD approach is able to shrink the depth and/or width of SSL model structures, there has been little research on how varying the depth and width impacts the internal representation of the small-footprint model. This paper provides an empirical study that addresses the question. We investigate the performance on SUPERB while varying the structure and KD methods so as to keep the number of parameters constant; this allows us to analyze the contribution of the representation introduced by varying the model architecture. Experiments demonstrate that a certain depth is essential for solving content-oriented tasks (e.g. automatic speech recognition) accurately, whereas a certain width is necessary for achieving high performance on several speaker-oriented tasks (e.g. speaker identification). Based on these observations, we identify, for SUPERB, a more compressed model with better performance than previous studies.

arXiv:2207.06694공식 코드 · Papers with Code

Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting

End-to-end text spotting has attached great attention recently due to its benefits on global optimization and high maintainability for real applications. However, the input scale has always been a tough trade-off since recognizing a small text instance usually requires enlarging the whole image, which brings high computational costs. In this paper, to address this problem, we propose a novel cost-efficient Dynamic Low-resolution Distillation (DLD) text spotting framework, which aims to infer images in different small but recognizable resolutions and achieve a better balance between accuracy and efficiency. Concretely, we adopt a resolution selector to dynamically decide the input resolutions for different images, which is constraint by both inference accuracy and computational cost. Another sequential knowledge distillation strategy is conducted on the text recognition branch, making the low-res input obtains comparable performance to a high-res image. The proposed method can be optimized end-to-end and adopted in any current text spotting framework to improve the practicability. Extensive experiments on several text spotting benchmarks show that the proposed method vastly improves the usability of low-res models. The code is available at https://github.com/hikopensource/DAVAR-Lab-OCR/.

arXiv:2207.06667공식 코드 · Papers with Code

Large-scale Knowledge Distillation with Elastic Heterogeneous Computing Resources

Although more layers and more parameters generally improve the accuracy of the models, such big models generally have high computational complexity and require big memory, which exceed the capacity of small devices for inference and incurs long training time. In addition, it is difficult to afford long training time and inference time of big models even in high performance servers, as well. As an efficient approach to compress a large deep model (a teacher model) to a compact model (a student model), knowledge distillation emerges as a promising approach to deal with the big models. Existing knowledge distillation methods cannot exploit the elastic available computing resources and correspond to low efficiency. In this paper, we propose an Elastic Deep Learning framework for knowledge Distillation, i.e., EDL-Dist. The advantages of EDL-Dist are three-fold. First, the inference and the training process is separated. Second, elastic available computing resources can be utilized to improve the efficiency. Third, fault-tolerance of the training and inference processes is supported. We take extensive experimentation to show that the throughput of EDL-Dist is up to 3.125 times faster than the baseline method (online knowledge distillation) while the accuracy is similar or higher.

arXiv:2207.06936공식 코드 · Papers with Code

Multi-Level Branched Regularization for Federated Learning

A critical challenge of federated learning is data heterogeneity and imbalance across clients, which leads to inconsistency between local networks and unstable convergence of global models. To alleviate the limitations, we propose a novel architectural regularization technique that constructs multiple auxiliary branches in each local model by grafting local and global subnetworks at several different levels and that learns the representations of the main pathway in the local model congruent to the auxiliary hybrid pathways via online knowledge distillation. The proposed technique is effective to robustify the global model even in the non-iid setting and is applicable to various federated learning frameworks conveniently without incurring extra communication costs. We perform comprehensive empirical studies and demonstrate remarkable performance gains in terms of accuracy and efficiency compared to existing methods. The source code is available at our project page.

arXiv:2207.05784공식 코드 · Papers with Code

Distilled Non-Semantic Speech Embeddings with Binary Neural Networks for Low-Resource Devices

This work introduces BRILLsson, a novel binary neural network-based representation learning model for a broad range of non-semantic speech tasks. We train the model with knowledge distillation from a large and real-valued TRILLsson model with only a fraction of the dataset used to train TRILLsson. The resulting BRILLsson models are only 2MB in size with a latency less than 8ms, making them suitable for deployment in low-resource devices such as wearables. We evaluate BRILLsson on eight benchmark tasks (including but not limited to spoken language identification, emotion recognition, health condition diagnosis, and keyword spotting), and demonstrate that our proposed ultra-light and low-latency models perform as well as large-scale models.

arXiv:2207.06389공식 코드 · Papers with Code

ProDiff: Progressive Fast Diffusion Model For High-Quality Text-to-Speech

Denoising diffusion probabilistic models (DDPMs) have recently achieved leading performances in many generative tasks. However, the inherited iterative sampling process costs hinder their applications to text-to-speech deployment. Through the preliminary study on diffusion model parameterization, we find that previous gradient-based TTS models require hundreds or thousands of iterations to guarantee high sample quality, which poses a challenge for accelerating sampling. In this work, we propose ProDiff, on progressive fast diffusion model for high-quality text-to-speech. Unlike previous work estimating the gradient for data density, ProDiff parameterizes the denoising model by directly predicting clean data to avoid distinct quality degradation in accelerating sampling. To tackle the model convergence challenge with decreased diffusion iterations, ProDiff reduces the data variance in the target site via knowledge distillation. Specifically, the denoising model uses the generated mel-spectrogram from an N-step DDIM teacher as the training target and distills the behavior into a new model with N/2 steps. As such, it allows the TTS model to make sharp predictions and further reduces the sampling time by orders of magnitude. Our evaluation demonstrates that ProDiff needs only 2 iterations to synthesize high-fidelity mel-spectrograms, while it maintains sample quality and diversity competitive with state-of-the-art models using hundreds of steps. ProDiff enables a sampling speed of 24x faster than real-time on a single NVIDIA 2080Ti GPU, making diffusion models practically applicable to text-to-speech synthesis deployment for the first time. Our extensive ablation studies demonstrate that each design in ProDiff is effective, and we further show that ProDiff can be easily extended to the multi-speaker setting. Audio samples are available at \url{https://ProDiff.github.io/.}

arXiv:2207.06242

SlimSeg: Slimmable Semantic Segmentation with Boundary Supervision

Accurate semantic segmentation models typically require significant computational resources, inhibiting their use in practical applications. Recent works rely on well-crafted lightweight models to achieve fast inference. However, these models cannot flexibly adapt to varying accuracy and efficiency requirements. In this paper, we propose a simple but effective slimmable semantic segmentation (SlimSeg) method, which can be executed at different capacities during inference depending on the desired accuracy-efficiency tradeoff. More specifically, we employ parametrized channel slimming by stepwise downward knowledge distillation during training. Motivated by the observation that the differences between segmentation results of each submodel are mainly near the semantic borders, we introduce an additional boundary guided semantic segmentation loss to further improve the performance of each submodel. We show that our proposed SlimSeg with various mainstream networks can produce flexible models that provide dynamic adjustment of computational cost and better performance than independent models. Extensive experiments on semantic segmentation benchmarks, Cityscapes and CamVid, demonstrate the generalization ability of our framework.

arXiv:2207.06075

DSPNet: Towards Slimmable Pretrained Networks based on Discriminative Self-supervised Learning

Self-supervised learning (SSL) has achieved promising downstream performance. However, when facing various resource budgets in real-world applications, it costs a huge computation burden to pretrain multiple networks of various sizes one by one. In this paper, we propose Discriminative-SSL-based Slimmable Pretrained Networks (DSPNet), which can be trained at once and then slimmed to multiple sub-networks of various sizes, each of which faithfully learns good representation and can serve as good initialization for downstream tasks with various resource budgets. Specifically, we extend the idea of slimmable networks to a discriminative SSL paradigm, by integrating SSL and knowledge distillation gracefully. We show comparable or improved performance of DSPNet on ImageNet to the networks individually pretrained one by one under the linear evaluation and semi-supervised evaluation protocols, while reducing large training cost. The pretrained models also generalize well on downstream detection and segmentation tasks. Code will be made public.

arXiv:2207.06300공식 코드 · Papers with Code

Re2G: Retrieve, Rerank, Generate

As demonstrated by GPT-3 and T5, transformers grow in capability as parameter spaces become larger and larger. However, for tasks that require a large amount of knowledge, non-parametric memory allows models to grow dramatically with a sub-linear increase in computational cost and GPU memory requirements. Recent models such as RAG and REALM have introduced retrieval into conditional generation. These models incorporate neural initial retrieval from a corpus of passages. We build on this line of research, proposing Re2G, which combines both neural initial retrieval and reranking into a BART-based sequence-to-sequence generation. Our reranking approach also permits merging retrieval results from sources with incomparable scores, enabling an ensemble of BM25 and neural initial retrieval. To train our system end-to-end, we introduce a novel variation of knowledge distillation to train the initial retrieval, reranker, and generation using only ground truth on the target sequence output. We find large gains in four diverse tasks: zero-shot slot filling, question answering, fact-checking, and dialog, with relative gains of 9% to 34% over the previous state-of-the-art on the KILT leaderboard. We make our code available as open source at https://github.com/IBM/kgi-slot-filling/tree/re2g.

arXiv:2207.05730

1st Place Solution to the EPIC-Kitchens Action Anticipation Challenge 2022

In this report, we describe the technical details of our submission to the EPIC-Kitchens Action Anticipation Challenge 2022. In this competition, we develop the following two approaches. 1) Anticipation Time Knowledge Distillation using the soft labels learned by the teacher model as knowledge to guide the student network to learn the information of anticipation time; 2) Verb-Noun Relation Module for building the relationship between verbs and nouns. Our method achieves state-of-the-art results on the testing set of EPIC-Kitchens Action Anticipation Challenge 2022.

arXiv:2207.05409공식 코드 · Papers with Code

Knowledge Condensation Distillation

Knowledge Distillation (KD) transfers the knowledge from a high-capacity teacher network to strengthen a smaller student. Existing methods focus on excavating the knowledge hints and transferring the whole knowledge to the student. However, the knowledge redundancy arises since the knowledge shows different values to the student at different learning stages. In this paper, we propose Knowledge Condensation Distillation (KCD). Specifically, the knowledge value on each sample is dynamically estimated, based on which an Expectation-Maximization (EM) framework is forged to iteratively condense a compact knowledge set from the teacher to guide the student learning. Our approach is easy to build on top of the off-the-shelf KD methods, with no extra training parameters and negligible computation overhead. Thus, it presents one new perspective for KD, in which the student that actively identifies teacher's knowledge in line with its aptitude can learn to learn more effectively and efficiently. Experiments on standard benchmarks manifest that the proposed KCD can well boost the performance of student model with even higher distillation efficiency. Code is available at https://github.com/dzy3/KCD.

arXiv:2207.05345공식 코드 · Papers with Code

HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors

Conventional knowledge distillation (KD) methods for object detection mainly concentrate on homogeneous teacher-student detectors. However, the design of a lightweight detector for deployment is often significantly different from a high-capacity detector. Thus, we investigate KD among heterogeneous teacher-student pairs for a wide application. We observe that the core difficulty for heterogeneous KD (hetero-KD) is the significant semantic gap between the backbone features of heterogeneous detectors due to the different optimization manners. Conventional homogeneous KD (homo-KD) methods suffer from such a gap and are hard to directly obtain satisfactory performance for hetero-KD. In this paper, we propose the HEtero-Assists Distillation (HEAD) framework, leveraging heterogeneous detection heads as assistants to guide the optimization of the student detector to reduce this gap. In HEAD, the assistant is an additional detection head with the architecture homogeneous to the teacher head attached to the student backbone. Thus, a hetero-KD is transformed into a homo-KD, allowing efficient knowledge transfer from the teacher to the student. Moreover, we extend HEAD into a Teacher-Free HEAD (TF-HEAD) framework when a well-trained teacher detector is unavailable. Our method has achieved significant improvement compared to current detection KD methods. For example, on the MS-COCO dataset, TF-HEAD helps R18 RetinaNet achieve 33.9 mAP (+2.2), while HEAD further pushes the limit to 36.2 mAP (+4.5).

arXiv:2207.05306공식 코드 · Papers with Code

Contrastive Deep Supervision

The success of deep learning is usually accompanied by the growth in neural network depth. However, the traditional training method only supervises the neural network at its last layer and propagates the supervision layer-by-layer, which leads to hardship in optimizing the intermediate layers. Recently, deep supervision has been proposed to add auxiliary classifiers to the intermediate layers of deep neural networks. By optimizing these auxiliary classifiers with the supervised task loss, the supervision can be applied to the shallow layers directly. However, deep supervision conflicts with the well-known observation that the shallow layers learn low-level features instead of task-biased high-level semantic features. To address this issue, this paper proposes a novel training framework named Contrastive Deep Supervision, which supervises the intermediate layers with augmentation-based contrastive learning. Experimental results on nine popular datasets with eleven models demonstrate its effects on general image classification, fine-grained image classification and object detection in supervised learning, semi-supervised learning and knowledge distillation. Codes have been released in Github.

arXiv:2207.05273

Cross-Architecture Knowledge Distillation

Transformer attracts much attention because of its ability to learn global relations and superior performance. In order to achieve higher performance, it is natural to distill complementary knowledge from Transformer to convolutional neural network (CNN). However, most existing knowledge distillation methods only consider homologous-architecture distillation, such as distilling knowledge from CNN to CNN. They may not be suitable when applying to cross-architecture scenarios, such as from Transformer to CNN. To deal with this problem, a novel cross-architecture knowledge distillation method is proposed. Specifically, instead of directly mimicking output/intermediate features of the teacher, partially cross attention projector and group-wise linear projector are introduced to align the student features with the teacher's in two projected feature spaces. And a multi-view robust training scheme is further presented to improve the robustness and stability of the framework. Extensive experiments show that the proposed method outperforms 14 state-of-the-arts on both small-scale and large-scale datasets.