Accurate liver and lesion segmentation from computed tomography (CT) images are highly demanded in clinical practice for assisting the diagnosis and assessment of hepatic tumor disease. However, automatic liver and lesion segmentation from contrast-enhanced CT volumes is extremely challenging due to the diversity in contrast, resolution, and quality of images. Previous methods based on UNet for 2D slice-by-slice or 3D volume-by-volume segmentation either lack sufficient spatial contexts or suffer from high GPU computational cost, which limits the performance. To tackle these issues, we propose a novel context-aware PolyUNet for accurate liver and lesion segmentation. It jointly explores structural diversity and consecutive t-adjacent slices to enrich feature expressive power and spatial contextual information while avoiding the overload of GPU memory consumption. In addition, we utilize zoom out/in and two-stage refinement strategy to exclude the irrelevant contexts and focus on the specific region for the fine-grained segmentation. Our method achieved very competitive performance at the MICCAI 2017 Liver Tumor Segmentation (LiTS) Challenge among all tasks with a single model and ranked the $3^{rd}$, $12^{th}$, $2^{nd}$, and $5^{th}$ places in the liver segmentation, lesion segmentation, lesion detection, and tumor burden estimation, respectively.
In recent years, the size of pre-trained language models (PLMs) has grown by leaps and bounds. However, efficiency issues of these large-scale PLMs limit their utilization in real-world scenarios. We present a suite of cost-effective techniques for the use of PLMs to deal with the efficiency issues of pre-training, fine-tuning, and inference. (1) We introduce knowledge inheritance to accelerate the pre-training process by exploiting existing PLMs instead of training models from scratch. (2) We explore the best practice of prompt tuning with large-scale PLMs. Compared with conventional fine-tuning, prompt tuning significantly reduces the number of task-specific parameters. (3) We implement a new inference toolkit, namely InfMoE, for using large-scale PLMs with limited computational resources. Based on our cost-effective pipeline, we pre-train two models: an encoder-decoder bilingual model with 11 billion parameters (CPM-2) and its corresponding MoE version with 198 billion parameters. In our experiments, we compare CPM-2 with mT5 on downstream tasks. Experimental results show that CPM-2 has excellent general language intelligence. Moreover, we validate the efficiency of InfMoE when conducting inference of large-scale models having tens of billions of parameters on a single GPU. All source code and model parameters are available at https://github.com/TsinghuaAI/CPM.
In this technical report, we introduce the methods of HIKVISION_LiDAR_Det in the challenge of waymo open dataset real-time 3D detection. Our solution for the competition are built upon Centerpoint 3D detection framework. Several variants of CenterPoint are explored, including center attention head and feature pyramid network neck. In order to achieve real time detection, methods like batchnorm merge, half-precision floating point network and GPU-accelerated voxelization process are adopted. By using these methods, our team ranks 6th among all the methods on real-time 3D detection challenge in the waymo open dataset.
Deploying an end-to-end automatic speech recognition (ASR) model on mobile/embedded devices is a challenging task, since the device computational power and energy consumption requirements are dynamically changed in practice. To overcome the issue, we present a training and pruning method for ASR based on the connectionist temporal classification (CTC) which allows reduction of model depth at run-time without any extra fine-tuning. To achieve the goal, we adopt two regularization methods, intermediate CTC and stochastic depth, to train a model whose performance does not degrade much after pruning. We present an in-depth analysis of layer behaviors using singular vector canonical correlation analysis (SVCCA), and efficient strategies for finding layers which are safe to prune. Using the proposed method, we show that a Transformer-CTC model can be pruned in various depth on demand, improving real-time factor from 0.005 to 0.002 on GPU, while each pruned sub-model maintains the accuracy of individually trained model of the same depth.
Analyzing the worst-case performance of deep neural networks against input perturbations amounts to solving a large-scale non-convex optimization problem, for which several past works have proposed convex relaxations as a promising alternative. However, even for reasonably-sized neural networks, these relaxations are not tractable, and so must be replaced by even weaker relaxations in practice. In this work, we propose a novel operator splitting method that can directly solve a convex relaxation of the problem to high accuracy, by splitting it into smaller sub-problems that often have analytical solutions. The method is modular, scales to very large problem instances, and compromises operations that are amenable to fast parallelization with GPU acceleration. We demonstrate our method in bounding the worst-case performance of large convolutional networks in image classification and reinforcement learning settings, and in reachability analysis of neural network dynamical systems.
An important paradigm of natural language processing consists of large-scale pre-training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full fine-tuning, which retrains all model parameters, becomes less feasible. Using GPT-3 175B as an example -- deploying independent instances of fine-tuned models, each with 175B parameters, is prohibitively expensive. We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks. Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. LoRA performs on-par or better than fine-tuning in model quality on RoBERTa, DeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters, a higher training throughput, and, unlike adapters, no additional inference latency. We also provide an empirical investigation into rank-deficiency in language model adaptation, which sheds light on the efficacy of LoRA. We release a package that facilitates the integration of LoRA with PyTorch models and provide our implementations and model checkpoints for RoBERTa, DeBERTa, and GPT-2 at https://github.com/microsoft/LoRA.
We seek to investigate the scalability of neuromorphic computing for computer vision, with the objective of replicating non-neuromorphic performance on computer vision tasks while reducing power consumption. We convert the deep Artificial Neural Network (ANN) architecture U-Net to a Spiking Neural Network (SNN) architecture using the Nengo framework. Both rate-based and spike-based models are trained and optimized for benchmarking performance and power, using a modified version of the ISBI 2D EM Segmentation dataset consisting of microscope images of cells. We propose a partitioning method to optimize inter-chip communication to improve speed and energy efficiency when deploying multi-chip networks on the Loihi neuromorphic chip. We explore the advantages of regularizing firing rates of Loihi neurons for converting ANN to SNN with minimum accuracy loss and optimized energy consumption. We propose a percentile based regularization loss function to limit the spiking rate of the neuron between a desired range. The SNN is converted directly from the corresponding ANN, and demonstrates similar semantic segmentation as the ANN using the same number of neurons and weights. However, the neuromorphic implementation on the Intel Loihi neuromorphic chip is over 2x more energy-efficient than conventional hardware (CPU, GPU) when running online (one image at a time). These power improvements are achieved without sacrificing the task performance accuracy of the network, and when all weights (Loihi, CPU, and GPU networks) are quantized to 8 bits.
In this paper, we propose an implementation of both Large Deformation Diffeomorphic Metric Mapping (LDDMM) and Metamorphosis image registration using a semi-Lagrangian scheme for geodesic shooting. We propose to solve both problems as an inexact matching providing a single and unifying cost function. We demonstrate that for image registration the use of a semi-Lagrangian scheme is more stable than a standard Eulerian scheme. Our GPU implementation is based on PyTorch, which greatly simplifies and accelerates the computations thanks to its powerful automatic differentiation engine. It will be freely available at https://github.com/antonfrancois/Demeter_metamorphosis.
In an autonomous driving system, it is essential to recognize vehicles, pedestrians and cyclists from images. Besides the high accuracy of the prediction, the requirement of real-time running brings new challenges for convolutional network models. In this report, we introduce a real-time method to detect the 2D objects from images. We aggregate several popular one-stage object detectors and train the models of variety input strategies independently, to yield better performance for accurate multi-scale detection of each category, especially for small objects. For model acceleration, we leverage TensorRT to optimize the inference time of our detection pipeline. As shown in the leaderboard, our proposed detection framework ranks the 2nd place with 75.00% L1 mAP and 69.72% L2 mAP in the real-time 2D detection track of the Waymo Open Dataset Challenges, while our framework achieves the latency of 45.8ms/frame on an Nvidia Tesla V100 GPU.
Recent work has shown that Neural Ordinary Differential Equations (ODEs) can serve as generative models of images using the perspective of Continuous Normalizing Flows (CNFs). Such models offer exact likelihood calculation, and invertible generation/density estimation. In this work we introduce a Multi-Resolution variant of such models (MRCNF), by characterizing the conditional distribution over the additional information required to generate a fine image that is consistent with the coarse image. We introduce a transformation between resolutions that allows for no change in the log likelihood. We show that this approach yields comparable likelihood values for various image datasets, with improved performance at higher resolutions, with fewer parameters, using only 1 GPU. Further, we examine the out-of-distribution properties of (Multi-Resolution) Continuous Normalizing Flows, and find that they are similar to those of other likelihood-based generative models.
The cost of DNA sequencing has dropped exponentially over the past decade, making genomic data accessible to a growing number of scientists. In bioinformatics, localization of short DNA sequences (reads) within large genomic sequences is commonly facilitated by constructing index data structures which allow for efficient querying of substrings. Recent metagenomic classification pipelines annotate reads with taxonomic labels by analyzing their $k$-mer histograms with respect to a reference genome database. CPU-based index construction is often performed in a preprocessing phase due to the relatively high cost of building irregular data structures such as hash maps. However, the rapidly growing amount of available reference genomes establishes the need for index construction and querying at interactive speeds. In this paper, we introduce MetaCache-GPU -- an ultra-fast metagenomic short read classifier specifically tailored to fit the characteristics of CUDA-enabled accelerators. Our approach employs a novel hash table variant featuring efficient minhash fingerprinting of reads for locality-sensitive hashing and their rapid insertion using warp-aggregated operations. Our performance evaluation shows that MetaCache-GPU is able to build large reference databases in a matter of seconds, enabling instantaneous operability, while popular CPU-based tools such as Kraken2 require over an hour for index construction on the same data. In the context of an ever-growing number of reference genomes, MetaCache-GPU is the first metagenomic classifier that makes analysis pipelines with on-demand composition of large-scale reference genome sets practical. The source code is publicly available at https://github.com/muellan/metacache .
We present a novel method for real-time pose and shape reconstruction of two strongly interacting hands. Our approach is the first two-hand tracking solution that combines an extensive list of favorable properties, namely it is marker-less, uses a single consumer-level depth camera, runs in real time, handles inter- and intra-hand collisions, and automatically adjusts to the user's hand shape. In order to achieve this, we embed a recent parametric hand pose and shape model and a dense correspondence predictor based on a deep neural network into a suitable energy minimization framework. For training the correspondence prediction network, we synthesize a two-hand dataset based on physical simulations that includes both hand pose and shape annotations while at the same time avoiding inter-hand penetrations. To achieve real-time rates, we phrase the model fitting in terms of a nonlinear least-squares problem so that the energy can be optimized based on a highly efficient GPU-based Gauss-Newton optimizer. We show state-of-the-art results in scenes that exceed the complexity level demonstrated by previous work, including tight two-hand grasps, significant inter-hand occlusions, and gesture interaction.
With the emergence of low-cost robotic systems, such as unmanned aerial vehicle, the importance of embedded high-performance image processing has increased. For a long time, FPGAs were the only processing hardware that were capable of high-performance computing, while at the same time preserving a low power consumption, essential for embedded systems. However, the recently increasing availability of embedded GPU-based systems, such as the NVIDIA Jetson series, comprised of an ARM CPU and a NVIDIA Tegra GPU, allows for massively parallel embedded computing on graphics hardware. With this in mind, we propose an approach for real-time embedded stereo processing on ARM and CUDA-enabled devices, which is based on the popular and widely used Semi-Global Matching algorithm. In this, we propose an optimization of the algorithm for embedded CUDA GPUs, by using massively parallel computing, as well as using the NEON intrinsics to optimize the algorithm for vectorized SIMD processing on embedded ARM CPUs. We have evaluated our approach with different configurations on two public stereo benchmark datasets to demonstrate that they can reach an error rate as low as 3.3%. Furthermore, our experiments show that the fastest configuration of our approach reaches up to 46 FPS on VGA image resolution. Finally, in a use-case specific qualitative evaluation, we have evaluated the power consumption of our approach and deployed it on the DJI Manifold 2-G attached to a DJI Matrix 210v2 RTK unmanned aerial vehicle (UAV), demonstrating its suitability for real-time stereo processing onboard a UAV.
State-of-the-art methods for scalable Gaussian processes use iterative algorithms, requiring fast matrix vector multiplies (MVMs) with the covariance kernel. The Structured Kernel Interpolation (SKI) framework accelerates these MVMs by performing efficient MVMs on a grid and interpolating back to the original space. In this work, we develop a connection between SKI and the permutohedral lattice used for high-dimensional fast bilateral filtering. Using a sparse simplicial grid instead of a dense rectangular one, we can perform GP inference exponentially faster in the dimension than SKI. Our approach, Simplex-GP, enables scaling SKI to high dimensions, while maintaining strong predictive performance. We additionally provide a CUDA implementation of Simplex-GP, which enables significant GPU acceleration of MVM based inference.
Bird's Eye View (BEV) is a popular representation for processing 3D point clouds, and by its nature is fundamentally sparse. Motivated by the computational limitations of mobile robot platforms, we create a fast, high-performance BEV 3D object detector that maintains and exploits this input sparsity to decrease runtimes over non-sparse baselines and avoids the tradeoff between pseudoimage area and runtime. We present results on KITTI, a canonical 3D detection dataset, and Matterport-Chair, a novel Matterport3D-derived chair detection dataset from scenes in real furnished homes. We evaluate runtime characteristics using a desktop GPU, an embedded ML accelerator, and a robot CPU, demonstrating that our method results in significant detection speedups (2X or more) for embedded systems with only a modest decrease in detection quality. Our work represents a new approach for practitioners to optimize models for embedded systems by maintaining and exploiting input sparsity throughout their entire pipeline to reduce runtime and resource usage while preserving detection performance.
We formalize and analyze a fundamental component of differentiable neural architecture search (NAS): local "operation scoring" at each operation choice. We view existing operation scoring functions as inexact proxies for accuracy, and we find that they perform poorly when analyzed empirically on NAS benchmarks. From this perspective, we introduce a novel \textit{perturbation-based zero-cost operation scoring} (Zero-Cost-PT) approach, which utilizes zero-cost proxies that were recently studied in multi-trial NAS but degrade significantly on larger search spaces, typical for differentiable NAS. We conduct a thorough empirical evaluation on a number of NAS benchmarks and large search spaces, from NAS-Bench-201, NAS-Bench-1Shot1, NAS-Bench-Macro, to DARTS-like and MobileNet-like spaces, showing significant improvements in both search time and accuracy. On the ImageNet classification task on the DARTS search space, our approach improved accuracy compared to the best current training-free methods (TE-NAS) while being over 10$\times$ faster (total searching time 25 minutes on a single GPU), and observed significantly better transferability on architectures searched on the CIFAR-10 dataset with an accuracy increase of 1.8 pp. Our code is available at: https://github.com/zerocostptnas/zerocost_operation_score.
Deep graph neural networks (GNNs) have achieved excellent results on various tasks on increasingly large graph datasets with millions of nodes and edges. However, memory complexity has become a major obstacle when training deep GNNs for practical applications due to the immense number of nodes, edges, and intermediate activations. To improve the scalability of GNNs, prior works propose smart graph sampling or partitioning strategies to train GNNs with a smaller set of nodes or sub-graphs. In this work, we study reversible connections, group convolutions, weight tying, and equilibrium models to advance the memory and parameter efficiency of GNNs. We find that reversible connections in combination with deep network architectures enable the training of overparameterized GNNs that significantly outperform existing methods on multiple datasets. Our models RevGNN-Deep (1001 layers with 80 channels each) and RevGNN-Wide (448 layers with 224 channels each) were both trained on a single commodity GPU and achieve an ROC-AUC of $87.74 \pm 0.13$ and $88.24 \pm 0.15$ on the ogbn-proteins dataset. To the best of our knowledge, RevGNN-Deep is the deepest GNN in the literature by one order of magnitude. Please visit our project website https://www.deepgcns.org/arch/gnn1000 for more information.
We propose a benchmark to study surrogate model accuracy for protein-ligand docking. We share a dataset consisting of 200 million 3D complex structures and 2D structure scores across a consistent set of 13 million "in-stock" molecules over 15 receptors, or binding sites, across the SARS-CoV-2 proteome. Our work shows surrogate docking models have six orders of magnitude more throughput than standard docking protocols on the same supercomputer node types. We demonstrate the power of high-speed surrogate models by running each target against 1 billion molecules in under a day (50k predictions per GPU seconds). We showcase a workflow for docking utilizing surrogate ML models as a pre-filter. Our workflow is ten times faster at screening a library of compounds than the standard technique, with an error rate less than 0.01\% of detecting the underlying best scoring 0.1\% of compounds. Our analysis of the speedup explains that to screen more molecules under a docking paradigm, another order of magnitude speedup must come from model accuracy rather than computing speed (which, if increased, will not anymore alter our throughput to screen molecules). We believe this is strong evidence for the community to begin focusing on improving the accuracy of surrogate models to improve the ability to screen massive compound libraries 100x or even 1000x faster than current techniques.
Neural networks have demonstrated their outstanding performance in a wide range of tasks. Specifically recurrent architectures based on long-short term memory (LSTM) cells have manifested excellent capability to model time dependencies in real-world data. However, standard recurrent architectures cannot estimate their uncertainty which is essential for safety-critical applications such as in medicine. In contrast, Bayesian recurrent neural networks (RNNs) are able to provide uncertainty estimation with improved accuracy. Nonetheless, Bayesian RNNs are computationally and memory demanding, which limits their practicality despite their advantages. To address this issue, we propose an FPGA-based hardware design to accelerate Bayesian LSTM-based RNNs. To further improve the overall algorithmic-hardware performance, a co-design framework is proposed to explore the most fitting algorithmic-hardware configurations for Bayesian RNNs. We conduct extensive experiments on healthcare applications to demonstrate the improvement of our design and the effectiveness of our framework. Compared with GPU implementation, our FPGA-based design can achieve up to 10 times speedup with nearly 106 times higher energy efficiency. To the best of our knowledge, this is the first work targeting acceleration of Bayesian RNNs on FPGAs.
Graph neural networks (GNNs) are powerful tools for learning from graph data and are widely used in various applications such as social network recommendation, fraud detection, and graph search. The graphs in these applications are typically large, usually containing hundreds of millions of nodes. Training GNN models on such large graphs efficiently remains a big challenge. Despite a number of sampling-based methods have been proposed to enable mini-batch training on large graphs, these methods have not been proved to work on truly industry-scale graphs, which require GPUs or mixed-CPU-GPU training. The state-of-the-art sampling-based methods are usually not optimized for these real-world hardware setups, in which data movement between CPUs and GPUs is a bottleneck. To address this issue, we propose Global Neighborhood Sampling that aims at training GNNs on giant graphs specifically for mixed-CPU-GPU training. The algorithm samples a global cache of nodes periodically for all mini-batches and stores them in GPUs. This global cache allows in-GPU importance sampling of mini-batches, which drastically reduces the number of nodes in a mini-batch, especially in the input layer, to reduce data copy between CPU and GPU and mini-batch computation without compromising the training convergence rate or model accuracy. We provide a highly efficient implementation of this method and show that our implementation outperforms an efficient node-wise neighbor sampling baseline by a factor of 2X-4X on giant graphs. It outperforms an efficient implementation of LADIES with small layers by a factor of 2X-14X while achieving much higher accuracy than LADIES.We also theoretically analyze the proposed algorithm and show that with cached node data of a proper size, it enjoys a comparable convergence rate as the underlying node-wise sampling method.