PyTorchKR Papers

최신 논문

총 5,576편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2212.05505

Focal-PETR: Embracing Foreground for Efficient Multi-Camera 3D Object Detection

The dominant multi-camera 3D detection paradigm is based on explicit 3D feature construction, which requires complicated indexing of local image-view features via 3D-to-2D projection. Other methods implicitly introduce geometric positional encoding and perform global attention (e.g., PETR) to build the relationship between image tokens and 3D objects. The 3D-to-2D perspective inconsistency and global attention lead to a weak correlation between foreground tokens and queries, resulting in slow convergence. We propose Focal-PETR with instance-guided supervision and spatial alignment module to adaptively focus object queries on discriminative foreground regions. Focal-PETR additionally introduces a down-sampling strategy to reduce the consumption of global attention. Due to the highly parallelized implementation and down-sampling strategy, our model, without depth supervision, achieves leading performance on the large-scale nuScenes benchmark and a superior speed of 30 FPS on a single RTX3090 GPU. Extensive experiments show that our method outperforms PETR while consuming 3x fewer training hours. The code will be made publicly available.

arXiv:2212.05598공식 코드 · Papers with Code

Recurrent Vision Transformers for Object Detection with Event Cameras

We present Recurrent Vision Transformers (RVTs), a novel backbone for object detection with event cameras. Event cameras provide visual information with sub-millisecond latency at a high-dynamic range and with strong robustness against motion blur. These unique properties offer great potential for low-latency object detection and tracking in time-critical scenarios. Prior work in event-based vision has achieved outstanding detection performance but at the cost of substantial inference time, typically beyond 40 milliseconds. By revisiting the high-level design of recurrent vision backbones, we reduce inference time by a factor of 6 while retaining similar performance. To achieve this, we explore a multi-stage design that utilizes three key concepts in each stage: First, a convolutional prior that can be regarded as a conditional positional embedding. Second, local and dilated global self-attention for spatial feature interaction. Third, recurrent temporal feature aggregation to minimize latency while retaining temporal information. RVTs can be trained from scratch to reach state-of-the-art performance on event-based object detection - achieving an mAP of 47.2% on the Gen1 automotive dataset. At the same time, RVTs offer fast inference (<12 ms on a T4 GPU) and favorable parameter efficiency (5 times fewer than prior art). Our study brings new insights into effective design choices that can be fruitful for research beyond event-based vision.

arXiv:2212.05339공식 코드 · Papers with Code

Elixir: Train a Large Language Model on a Small GPU Cluster

In recent years, large language models have achieved great success due to their unprecedented size. However, training these models poses a challenge for most researchers as it requires a substantial number of GPUs. To reduce GPU memory usage, memory partitioning, and memory offloading have been proposed. These approaches eliminate memory redundancies and offload memory usage to the CPU and NVMe memory, respectively, enabling training on small GPU clusters. However, directly deploying these solutions often leads to suboptimal efficiency. Only experienced experts can unleash the full potential of hardware by carefully tuning the distributed configuration. Thus, we present a novel solution, Elixir, which automates efficient large-model training based on pre-runtime model profiling. Elixir aims to identify the optimal combination of partitioning and offloading techniques to maximize training throughput. In our experiments, Elixir significantly outperforms the current state-of-the-art baseline. Our optimal configuration achieves up to a 3.4$\times$ speedup on GPT-2 models compared with SOTA solutions. We hope that our work will benefit individuals who lack computing resources and expertise, granting them access to large models. The beta version of Elixir is now available at https://github.com/hpcaitech/ColossalAI/tree/feature/elixir.

arXiv:2212.04540

TinyKG: Memory-Efficient Training Framework for Knowledge Graph Neural Recommender Systems

There has been an explosion of interest in designing various Knowledge Graph Neural Networks (KGNNs), which achieve state-of-the-art performance and provide great explainability for recommendation. The promising performance is mainly resulting from their capability of capturing high-order proximity messages over the knowledge graphs. However, training KGNNs at scale is challenging due to the high memory usage. In the forward pass, the automatic differentiation engines (\textsl{e.g.}, TensorFlow/PyTorch) generally need to cache all intermediate activation maps in order to compute gradients in the backward pass, which leads to a large GPU memory footprint. Existing work solves this problem by utilizing multi-GPU distributed frameworks. Nonetheless, this poses a practical challenge when seeking to deploy KGNNs in memory-constrained environments, especially for industry-scale graphs. Here we present TinyKG, a memory-efficient GPU-based training framework for KGNNs for the tasks of recommendation. Specifically, TinyKG uses exact activations in the forward pass while storing a quantized version of activations in the GPU buffers. During the backward pass, these low-precision activations are dequantized back to full-precision tensors, in order to compute gradients. To reduce the quantization errors, TinyKG applies a simple yet effective quantization algorithm to compress the activations, which ensures unbiasedness with low variance. As such, the training memory footprint of KGNNs is largely reduced with negligible accuracy loss. To evaluate the performance of our TinyKG, we conduct comprehensive experiments on real-world datasets. We found that our TinyKG with INT2 quantization aggressively reduces the memory footprint of activation maps with $7 \times$, only with $2\%$ loss in accuracy, allowing us to deploy KGNNs on memory-constrained devices.

arXiv:2212.05031

Towards a learning-based performance modeling for accelerating Deep Neural Networks

Emerging applications such as Deep Learning are often data-driven, thus traditional approaches based on auto-tuners are not performance effective across the wide range of inputs used in practice. In the present paper, we start an investigation of predictive models based on machine learning techniques in order to optimize Convolution Neural Networks (CNNs). As a use-case, we focus on the ARM Compute Library which provides three different implementations of the convolution operator at different numeric precision. Starting from a collation of benchmarks, we build and validate models learned by Decision Tree and naive Bayesian classifier. Preliminary experiments on Midgard-based ARM Mali GPU show that our predictive model outperforms all the convolution operators manually selected by the library.

arXiv:2212.03942

An Efficient Evolutionary Deep Learning Framework Based on Multi-source Transfer Learning to Evolve Deep Convolutional Neural Networks

Convolutional neural networks (CNNs) have constantly achieved better performance over years by introducing more complex topology, and enlarging the capacity towards deeper and wider CNNs. This makes the manual design of CNNs extremely difficult, so the automated design of CNNs has come into the research spotlight, which has obtained CNNs that outperform manually-designed CNNs. However, the computational cost is still the bottleneck of automatically designing CNNs. In this paper, inspired by transfer learning, a new evolutionary computation based framework is proposed to efficiently evolve CNNs without compromising the classification accuracy. The proposed framework leverages multi-source domains, which are smaller datasets than the target domain datasets, to evolve a generalised CNN block only once. And then, a new stacking method is proposed to both widen and deepen the evolved block, and a grid search method is proposed to find optimal stacking solutions. The experimental results show the proposed method acquires good CNNs faster than 15 peer competitors within less than 40 GPU-hours. Regarding the classification accuracy, the proposed method gains its strong competitiveness against the peer competitors, which achieves the best error rates of 3.46%, 18.36% and 1.76% for the CIFAR-10, CIFAR-100 and SVHN datasets, respectively.

arXiv:2212.03854

BiPMAP: A Toolbox for Predictions of Perceived Motion Artifacts on Modern Displays

Presenting dynamic scenes without incurring motion artifacts visible to observers requires sustained effort from the display industry. A tool that predicts motion artifacts and simulates artifact elimination through optimizing the display configuration is highly desired to guide the design and manufacture of modern displays. Despite the popular demands, there is no such tool available in the market. In this study, we deliver an interactive toolkit, Binocular Perceived Motion Artifact Predictor (BiPMAP), as an executable file with GPU acceleration. BiPMAP accounts for an extensive collection of user-defined parameters and directly visualizes a variety of motion artifacts by presenting the perceived continuous and sampled moving stimuli side-by-side. For accurate artifact predictions, BiPMAP utilizes a novel model of the human contrast sensitivity function to effectively imitate the frequency modulation of the human visual system. In addition, BiPMAP is capable of deriving various in-plane motion artifacts for 2D displays and depth distortion in 3D stereoscopic displays.

arXiv:2212.03659공식 코드 · Papers with Code

Multi-Objective Linear Ensembles for Robust and Sparse Training of Few-Bit Neural Networks

Training neural networks (NNs) using combinatorial optimization solvers has gained attention in recent years. In low-data settings, state-of-the-art mixed integer linear programming solvers can train exactly a NN, avoiding intensive GPU-based training and hyper-parameter tuning and simultaneously training and sparsifying the network. We study the case of few-bit discrete-valued neural networks, both Binarized Neural Networks (BNNs), whose values are restricted to +-1, and Integer Neural Networks (INNs), whose values lie in a range {-P, ..., P}. Few-bit NNs receive increasing recognition due to their lightweight architecture and ability to run on low-power devices. This paper proposes new methods to improve the training of BNNs and INNs. Our contribution is a multi-objective ensemble approach based on training a single NN for each possible pair of classes and applying a majority voting scheme to predict the final output. Our approach results in training robust sparsified networks whose output is not affected by small perturbations on the input and whose number of active weights is as small as possible. We compare this BeMi approach to the current state-of-the-art in solver-based NN training and gradient-based training, focusing on BNN learning in few-shot contexts. We compare the benefits and drawbacks of INNs versus BNNs, bringing new light to the distribution of weights over the {-P, ..., P} interval. Finally, we compare multi-objective versus single-objective training of INNs, showing that robustness and network simplicity can be acquired simultaneously, thus obtaining better test performances. While the previous state-of-the-art approaches achieve an average accuracy of 51.1% on the MNIST dataset, the BeMi ensemble approach achieves an average accuracy of 68.4% when trained with 10 images per class and 81.8% when trained with 40 images per class, having up to 75.3% NN links removed.

arXiv:2212.03300공식 코드 · Papers with Code

Supervised Tractogram Filtering using Geometric Deep Learning

A tractogram is a virtual representation of the brain white matter. It is composed of millions of virtual fibers, encoded as 3D polylines, which approximate the white matter axonal pathways. To date, tractograms are the most accurate white matter representation and thus are used for tasks like presurgical planning and investigations of neuroplasticity, brain disorders, or brain networks. However, it is a well-known issue that a large portion of tractogram fibers is not anatomically plausible and can be considered artifacts of the tracking procedure. With Verifyber, we tackle the problem of filtering out such non-plausible fibers using a novel fully-supervised learning approach. Differently from other approaches based on signal reconstruction and/or brain topology regularization, we guide our method with the existing anatomical knowledge of the white matter. Using tractograms annotated according to anatomical principles, we train our model, Verifyber, to classify fibers as either anatomically plausible or non-plausible. The proposed Verifyber model is an original Geometric Deep Learning method that can deal with variable size fibers, while being invariant to fiber orientation. Our model considers each fiber as a graph of points, and by learning features of the edges between consecutive points via the proposed sequence Edge Convolution, it can capture the underlying anatomical properties. The output filtering results highly accurate and robust across an extensive set of experiments, and fast; with a 12GB GPU, filtering a tractogram of 1M fibers requires less than a minute. Verifyber implementation and trained models are available at https://github.com/FBK-NILab/verifyber.

arXiv:2212.02791

Event-based Monocular Dense Depth Estimation with Recurrent Transformers

Event cameras, offering high temporal resolutions and high dynamic ranges, have brought a new perspective to address common challenges (e.g., motion blur and low light) in monocular depth estimation. However, how to effectively exploit the sparse spatial information and rich temporal cues from asynchronous events remains a challenging endeavor. To this end, we propose a novel event-based monocular depth estimator with recurrent transformers, namely EReFormer, which is the first pure transformer with a recursive mechanism to process continuous event streams. Technically, for spatial modeling, a novel transformer-based encoder-decoder with a spatial transformer fusion module is presented, having better global context information modeling capabilities than CNN-based methods. For temporal modeling, we design a gate recurrent vision transformer unit that introduces a recursive mechanism into transformers, improving temporal modeling capabilities while alleviating the expensive GPU memory cost. The experimental results show that our EReFormer outperforms state-of-the-art methods by a margin on both synthetic and real-world datasets. We hope that our work will attract further research to develop stunning transformers in the event-based vision community. Our open-source code can be found in the supplemental material.

arXiv:2212.02687

Vision Transformer Computation and Resilience for Dynamic Inference

State-of-the-art deep learning models for computer vision tasks are based on the transformer architecture and often deployed in real-time applications. In this scenario, the resources available for every inference can vary, so it is useful to be able to dynamically adapt execution to trade accuracy for efficiency. To create dynamic models, we leverage the resilience of vision transformers to pruning and switch between different scaled versions of a model. Surprisingly, we find that most FLOPs are generated by convolutions, not attention. These relative FLOP counts are not a good predictor of GPU performance since GPUs have special optimizations for convolutions. Some models are fairly resilient and their model execution can be adapted without retraining, while all models achieve better accuracy with retraining alternative execution paths. These insights mean that we can leverage CNN accelerators and these alternative execution paths to enable efficient and dynamic vision transformer inference. Our analysis shows that leveraging this type of dynamic execution can lead to saving 28\% of energy with a 1.4\% accuracy drop for SegFormer (63 GFLOPs), with no additional training, and 53\% of energy for ResNet-50 (4 GFLOPs) with a 3.3\% accuracy drop by switching between pretrained Once-For-All models.

arXiv:2212.02567

cs-net: structural approach to time-series forecasting for high-dimensional feature space data with limited observations

In recent years, deep-learning-based approaches have been introduced to solving time-series forecasting-related problems. These novel methods have demonstrated impressive performance in univariate and low-dimensional multivariate time-series forecasting tasks. However, when these novel methods are used to handle high-dimensional multivariate forecasting problems, their performance is highly restricted by a practical training time and a reasonable GPU memory configuration. In this paper, inspired by a change of basis in the Hilbert space, we propose a flexible data feature extraction technique that excels in high-dimensional multivariate forecasting tasks. Our approach was originally developed for the National Science Foundation (NSF) Algorithms for Threat Detection (ATD) 2022 Challenge. Implemented using the attention mechanism and Convolutional Neural Networks (CNN) architecture, our method demonstrates great performance and compatibility. Our models trained on the GDELT Dataset finished 1st and 2nd places in the ATD sprint series and hold promise for other datasets for time series forecasting.

arXiv:2212.01543

The RoyalFlush System for the WMT 2022 Efficiency Task

This paper describes the submission of the RoyalFlush neural machine translation system for the WMT 2022 translation efficiency task. Unlike the commonly used autoregressive translation system, we adopted a two-stage translation paradigm called Hybrid Regression Translation (HRT) to combine the advantages of autoregressive and non-autoregressive translation. Specifically, HRT first autoregressively generates a discontinuous sequence (e.g., make a prediction every $k$ tokens, $k>1$) and then fills in all previously skipped tokens at once in a non-autoregressive manner. Thus, we can easily trade off the translation quality and speed by adjusting $k$. In addition, by integrating other modeling techniques (e.g., sequence-level knowledge distillation and deep-encoder-shallow-decoder layer allocation strategy) and a mass of engineering efforts, HRT improves 80\% inference speed and achieves equivalent translation performance with the same-capacity AT counterpart. Our fastest system reaches 6k+ words/second on the GPU latency setting, estimated to be about 3.1x faster than the last year's winner.

arXiv:2212.02224

Bi-Level Optimization Augmented with Conditional Variational Autoencoder for Autonomous Driving in Dense Traffic

Autonomous driving has a natural bi-level structure. The goal of the upper behavioural layer is to provide appropriate lane change, speeding up, and braking decisions to optimize a given driving task. However, this layer can only indirectly influence the driving efficiency through the lower-level trajectory planner, which takes in the behavioural inputs to produce motion commands. Existing sampling-based approaches do not fully exploit the strong coupling between the behavioural and planning layer. On the other hand, end-to-end Reinforcement Learning (RL) can learn a behavioural layer while incorporating feedback from the lower-level planner. However, purely data-driven approaches often fail in safety metrics in unseen environments. This paper presents a novel alternative; a parameterized bi-level optimization that jointly computes the optimal behavioural decisions and the resulting downstream trajectory. Our approach runs in real-time using a custom GPU-accelerated batch optimizer, and a Conditional Variational Autoencoder learnt warm-start strategy. Extensive simulations show that our approach outperforms state-of-the-art model predictive control and RL approaches in terms of collision rate while being competitive in driving efficiency.

arXiv:2212.01567

AdaCM: Adaptive ColorMLP for Real-Time Universal Photo-realistic Style Transfer

Photo-realistic style transfer aims at migrating the artistic style from an exemplar style image to a content image, producing a result image without spatial distortions or unrealistic artifacts. Impressive results have been achieved by recent deep models. However, deep neural network based methods are too expensive to run in real-time. Meanwhile, bilateral grid based methods are much faster but still contain artifacts like overexposure. In this work, we propose the \textbf{Adaptive ColorMLP (AdaCM)}, an effective and efficient framework for universal photo-realistic style transfer. First, we find the complex non-linear color mapping between input and target domain can be efficiently modeled by a small multi-layer perceptron (ColorMLP) model. Then, in \textbf{AdaCM}, we adopt a CNN encoder to adaptively predict all parameters for the ColorMLP conditioned on each input content and style image pair. Experimental results demonstrate that AdaCM can generate vivid and high-quality stylization results. Meanwhile, our AdaCM is ultrafast and can process a 4K resolution image in 6ms on one V100 GPU.

arXiv:2212.02135공식 코드 · Papers with Code

SoftCTC -- Semi-Supervised Learning for Text Recognition using Soft Pseudo-Labels

This paper explores semi-supervised training for sequence tasks, such as Optical Character Recognition or Automatic Speech Recognition. We propose a novel loss function $\unicode{x2013}$ SoftCTC $\unicode{x2013}$ which is an extension of CTC allowing to consider multiple transcription variants at the same time. This allows to omit the confidence based filtering step which is otherwise a crucial component of pseudo-labeling approaches to semi-supervised learning. We demonstrate the effectiveness of our method on a challenging handwriting recognition task and conclude that SoftCTC matches the performance of a finely-tuned filtering based pipeline. We also evaluated SoftCTC in terms of computational efficiency, concluding that it is significantly more efficient than a na\"ive CTC-based approach for training on multiple transcription variants, and we make our GPU implementation public.

arXiv:2212.02126

Accelerating Interactive Human-like Manipulation Learning with GPU-based Simulation and High-quality Demonstrations

Dexterous manipulation with anthropomorphic robot hands remains a challenging problem in robotics because of the high-dimensional state and action spaces and complex contacts. Nevertheless, skillful closed-loop manipulation is required to enable humanoid robots to operate in unstructured real-world environments. Reinforcement learning (RL) has traditionally imposed enormous interaction data requirements for optimizing such complex control problems. We introduce a new framework that leverages recent advances in GPU-based simulation along with the strength of imitation learning in guiding policy search towards promising behaviors to make RL training feasible in these domains. To this end, we present an immersive virtual reality teleoperation interface designed for interactive human-like manipulation on contact rich tasks and a suite of manipulation environments inspired by tasks of daily living. Finally, we demonstrate the complementary strengths of massively parallel RL and imitation learning, yielding robust and natural behaviors. Videos of trained policies, our source code, and the collected demonstration datasets are available at https://maltemosbach.github.io/interactive_ human_like_manipulation/.

arXiv:2212.01185공식 코드 · Papers with Code

A Learned Pixel-by-Pixel Lossless Image Compression Method with 59K Parameters and Parallel Decoding

This paper considers lossless image compression and presents a learned compression system that can achieve state-of-the-art lossless compression performance but uses only 59K parameters, which is more than 30x less than other learned systems proposed recently in the literature. The explored system is based on a learned pixel-by-pixel lossless image compression method, where each pixel's probability distribution parameters are obtained by processing the pixel's causal neighborhood (i.e. previously encoded/decoded pixels) with a simple neural network comprising 59K parameters. This causality causes the decoder to operate sequentially, i.e. the neural network has to be evaluated for each pixel sequentially, which increases decoding time significantly with common GPU software and hardware. To reduce the decoding time, parallel decoding algorithms are proposed and implemented. The obtained lossless image compression system is compared to traditional and learned systems in the literature in terms of compression performance, encoding-decoding times and computational complexity.

arXiv:2212.00827

Architectural Implications of Embedding Dimension during GCN on CPU and GPU

Graph Neural Networks (GNNs) are a class of neural networks designed to extract information from the graphical structure of data. Graph Convolutional Networks (GCNs) are a widely used type of GNN for transductive graph learning problems which apply convolution to learn information from graphs. GCN is a challenging algorithm from an architecture perspective due to inherent sparsity, low data reuse, and massive memory capacity requirements. Traditional neural algorithms exploit the high compute capacity of GPUs to achieve high performance for both inference and training. The architectural decision to use a GPU for GCN inference is a question explored in this work. GCN on both CPU and GPU was characterized in order to better understand the implications of graph size, embedding dimension, and sampling on performance.

arXiv:2212.01232공식 코드 · Papers with Code

Loss shaping enhances exact gradient learning with Eventprop in spiking neural networks

Event-based machine learning promises more energy-efficient AI on future neuromorphic hardware. Here, we investigate how the recently discovered Eventprop algorithm for gradient descent on exact gradients in spiking neural networks can be scaled up to challenging keyword recognition benchmarks. We implemented Eventprop in the GPU-enhanced Neural Networks framework and used it for training recurrent spiking neural networks on the Spiking Heidelberg Digits and Spiking Speech Commands datasets. We found that learning depended strongly on the loss function and extended Eventprop to a wider class of loss functions to enable effective training. We then tested a large number of data augmentations and regularisations as well as exploring different network structures; and heterogeneous and trainable timescales. We found that when combined with two specific augmentations, the right regularisation and a delay line input, Eventprop networks with one recurrent layer achieved state-of-the-art performance on Spiking Heidelberg Digits and good accuracy on Spiking Speech Commands. In comparison to a leading surrogate-gradient-based SNN training method, our GeNN Eventprop implementation is 3X faster and uses 4X less memory. This work is a significant step towards a low-power neuromorphic alternative to current machine learning paradigms.