Typically, the deployment of face recognition models in the wild needs to identify low-resolution faces with extremely low computational cost. To address this problem, a feasible solution is compressing a complex face model to achieve higher speed and lower memory at the cost of minimal performance drop. Inspired by that, this paper proposes a learning approach to recognize low-resolution faces via selective knowledge distillation. In this approach, a two-stream convolutional neural network (CNN) is first initialized to recognize high-resolution faces and resolution-degraded faces with a teacher stream and a student stream, respectively. The teacher stream is represented by a complex CNN for high-accuracy recognition, and the student stream is represented by a much simpler CNN for low-complexity recognition. To avoid significant performance drop at the student stream, we then selectively distil the most informative facial features from the teacher stream by solving a sparse graph optimization problem, which are then used to regularize the fine-tuning process of the student stream. In this way, the student stream is actually trained by simultaneously handling two tasks with limited computational resources: approximating the most informative facial cues via feature regression, and recovering the missing facial cues via low-resolution face classification. Experimental results show that the student stream performs impressively in recognizing low-resolution faces and costs only 0.15MB memory and runs at 418 faces per second on CPU and 9,433 faces per second on GPU.
Accurate localization and segmentation of intervertebral disc (IVD) is crucial for the assessment of spine disease diagnosis. Despite the technological advances in medical imaging, IVD localization and segmentation are still manually performed, which is time-consuming and prone to errors. If, in addition, multi-modal imaging is considered, the burden imposed on disease assessments increases substantially. In this paper, we propose an architecture for IVD localization and segmentation in multi-modal MRI, which extends the well-known UNet. Compared to single images, multi-modal data brings complementary information, contributing to better data representation and discriminative power. Our contributions are three-fold. First, how to effectively integrate and fully leverage multi-modal data remains almost unexplored. In this work, each MRI modality is processed in a different path to better exploit their unique information. Second, inspired by HyperDenseNet, the network is densely-connected both within each path and across different paths, granting the model the freedom to learn where and how the different modalities should be processed and combined. Third, we improved standard U-Net modules by extending inception modules with two dilated convolutions blocks of different scale, which helps handling multi-scale context. We report experiments over the data set of the public MICCAI 2018 Challenge on Automatic Intervertebral Disc Localization and Segmentation, with 13 multi-modal MRI images used for training and 3 for validation. We trained IVD-Net on an NVidia TITAN XP GPU with 16 GBs RAM, using ADAM as optimizer and a learning rate of 10e-5 during 200 epochs. Training took about 5 hours, and segmentation of a whole volume about 2-3 seconds, on average. Several baselines, with different multi-modal fusion strategies, were used to demonstrate the effectiveness of the proposed architecture.
Convolutional neural networks (CNN) have achieved impressive performance on the wide variety of tasks (classification, detection, etc.) across multiple domains at the cost of high computational and memory requirements. Thus, leveraging CNNs for real-time applications necessitates model compression approaches that not only reduce the total number of parameters but reduce the overall computation as well. In this work, we present a stability-based approach for filter-level pruning of CNNs. We evaluate our proposed approach on different architectures (LeNet, VGG-16, ResNet, and Faster RCNN) and datasets and demonstrate its generalizability through extensive experiments. Moreover, our compressed models can be used at run-time without requiring any special libraries or hardware. Our model compression method reduces the number of FLOPS by an impressive factor of 6.03X and GPU memory footprint by more than 17X, significantly outperforming other state-of-the-art filter pruning methods.
Although CNNs have gained the ability to transfer learned knowledge from source task to target task by virtue of large annotated datasets but consume huge processing time to fine-tune without GPU. In this paper, we propose a new computationally efficient transfer learning approach using classification layer features of pre-trained CNNs by appending layer after existing classification layer. We demonstrate that fine-tuning of the appended layer with existing classification layer for new task converges much faster than baseline and in average outperforms baseline classification accuracy. Furthermore, we execute thorough experiments to examine the influence of quantity, similarity, and dissimilarity of training sets in our classification outcomes to demonstrate transferability of classification layer features.
Stacking-based deep neural network (S-DNN) is aggregated with pluralities of basic learning modules, one after another, to synthesize a deep neural network (DNN) alternative for pattern classification. Contrary to the DNNs trained end to end by backpropagation (BP), each S-DNN layer, i.e., a self-learnable module, is to be trained decisively and independently without BP intervention. In this paper, a ridge regression-based S-DNN, dubbed deep analytic network (DAN), along with its kernelization (K-DAN), are devised for multilayer feature re-learning from the pre-extracted baseline features and the structured features. Our theoretical formulation demonstrates that DAN/K-DAN re-learn by perturbing the intra/inter-class variations, apart from diminishing the prediction errors. We scrutinize the DAN/K-DAN performance for pattern classification on datasets of varying domains - faces, handwritten digits, generic objects, to name a few. Unlike the typical BP-optimized DNNs to be trained from gigantic datasets by GPU, we disclose that DAN/K-DAN are trainable using only CPU even for small-scale training sets. Our experimental results disclose that DAN/K-DAN outperform the present S-DNNs and also the BP-trained DNNs, including multiplayer perceptron, deep belief network, etc., without data augmentation applied.
In this paper, we present a novel neural network architecture for retinal vessel segmentation that improves over the state of the art on two benchmark datasets, is the first to run in real time on high resolution images, and its small memory and processing requirements make it deployable in mobile and embedded systems. The M2U-Net has a new encoder-decoder architecture that is inspired by the U-Net. It adds pretrained components of MobileNetV2 in the encoder part and novel contractive bottleneck blocks in the decoder part that, combined with bilinear upsampling, drastically reduce the parameter count to 0.55M compared to 31.03M in the original U-Net. We have evaluated its performance against a wide body of previously published results on three public datasets. On two of them, the M2U-Net achieves new state-of-the-art performance by a considerable margin. When implemented on a GPU, our method is the first to achieve real-time inference speeds on high-resolution fundus images. We also implemented our proposed network on an ARM-based embedded system where it segments images in between 0.6 and 15 sec, depending on the resolution. Thus, the M2U-Net enables a number of applications of retinal vessel structure extraction, such as early diagnosis of eye diseases, retinal biometric authentication systems, and robot assisted microsurgery.
This paper focuses on YOLO-LITE, a real-time object detection model developed to run on portable devices such as a laptop or cellphone lacking a Graphics Processing Unit (GPU). The model was first trained on the PASCAL VOC dataset then on the COCO dataset, achieving a mAP of 33.81% and 12.26% respectively. YOLO-LITE runs at about 21 FPS on a non-GPU computer and 10 FPS after implemented onto a website with only 7 layers and 482 million FLOPS. This speed is 3.8x faster than the fastest state of art model, SSD MobilenetvI. Based on the original object detection algorithm YOLOV2, YOLO- LITE was designed to create a smaller, faster, and more efficient model increasing the accessibility of real-time object detection to a variety of devices.
Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We address the overhead of the gradient synchronization with 2D-Torus all-reduce. Specifically, 2D-Torus all-reduce arranges GPUs in a logical 2D grid and performs a series of collective operation in different orientations. These two techniques are implemented with Neural Network Libraries (NNL). We have successfully trained ImageNet/ResNet-50 in 122 seconds without significant accuracy loss on ABCI cluster.
End-to-end approaches have drawn much attention recently for significantly simplifying the construction of an automatic speech recognition (ASR) system. RNN transducer (RNN-T) is one of the popular end-to-end methods. Previous studies have shown that RNN-T is difficult to train and a very complex training process is needed for a reasonable performance. In this paper, we explore RNN-T for a Chinese large vocabulary continuous speech recognition (LVCSR) task and aim to simplify the training process while maintaining performance. First, a new strategy of learning rate decay is proposed to accelerate the model convergence. Second, we find that adding convolutional layers at the beginning of the network and using ordered data can discard the pre-training process of the encoder without loss of performance. Besides, we design experiments to find a balance among the usage of GPU memory, training circle and model performance. Finally, we achieve 16.9% character error rate (CER) on our test set which is 2% absolute improvement from a strong BLSTM CE system with language model trained on the same text corpus.
This paper describes a method based on a sequence-to-sequence learning (Seq2Seq) with attention and context preservation mechanism for voice conversion (VC) tasks. Seq2Seq has been outstanding at numerous tasks involving sequence modeling such as speech synthesis and recognition, machine translation, and image captioning. In contrast to current VC techniques, our method 1) stabilizes and accelerates the training procedure by considering guided attention and proposed context preservation losses, 2) allows not only spectral envelopes but also fundamental frequency contours and durations of speech to be converted, 3) requires no context information such as phoneme labels, and 4) requires no time-aligned source and target speech data in advance. In our experiment, the proposed VC framework can be trained in only one day, using only one GPU of an NVIDIA Tesla K80, while the quality of the synthesized speech is higher than that of speech converted by Gaussian mixture model-based VC and is comparable to that of speech generated by recurrent neural network-based text-to-speech synthesis, which can be regarded as an upper limit on VC performance.
Attention-based encoder decoder network uses a left-to-right beam search algorithm in the inference step. The current beam search expands hypotheses and traverses the expanded hypotheses at the next time step. This traversal is implemented using a for-loop program in general, and it leads to speed down of the recognition process. In this paper, we propose a parallelism technique for beam search, which accelerates the search process by vectorizing multiple hypotheses to eliminate the for-loop program. We also propose a technique to batch multiple speech utterances for off-line recognition use, which reduces the for-loop program with regard to the traverse of multiple utterances. This extension is not trivial during beam search unlike during training due to several pruning and thresholding techniques for efficient decoding. In addition, our method can combine scores of external modules, RNNLM and CTC, in a batch as shallow fusion. We achieved 3.7 x speedup compared with the original beam search algorithm by vectoring hypotheses, and achieved 10.5 x speedup by further changing processing unit to GPU.
Distributed training of deep nets is an important technique to address some of the present day computing challenges like memory consumption and computational demands. Classical distributed approaches, synchronous or asynchronous, are based on the parameter server architecture, i.e., worker nodes compute gradients which are communicated to the parameter server while updated parameters are returned. Recently, distributed training with AllReduce operations gained popularity as well. While many of those operations seem appealing, little is reported about wall-clock training time improvements. In this paper, we carefully analyze the AllReduce based setup, propose timing models which include network latency, bandwidth, cluster size and compute time, and demonstrate that a pipelined training with a width of two combines the best of both synchronous and asynchronous training. Specifically, for a setup consisting of a four-node GPU cluster we show wall-clock time training improvements of up to 5.4x compared to conventional approaches.
There exists a plethora of techniques for inducing structured sparsity in parametric models during the optimization process, with the final goal of resource-efficient inference. However, few methods target a specific number of floating-point operations (FLOPs) as part of the optimization objective, despite many reporting FLOPs as part of the results. Furthermore, a one-size-fits-all approach ignores realistic system constraints, which differ significantly between, say, a GPU and a mobile phone -- FLOPs on the former incur less latency than on the latter; thus, it is important for practitioners to be able to specify a target number of FLOPs during model compression. In this work, we extend a state-of-the-art technique to directly incorporate FLOPs as part of the optimization objective and show that, given a desired FLOPs requirement, different neural networks can be successfully trained for image classification.
Image smoothing represents a fundamental component of many disparate computer vision and graphics applications. In this paper, we present a unified unsupervised (label-free) learning framework that facilitates generating flexible and high-quality smoothing effects by directly learning from data using deep convolutional neural networks (CNNs). The heart of the design is the training signal as a novel energy function that includes an edge-preserving regularizer which helps maintain important yet potentially vulnerable image structures, and a spatially-adaptive Lp flattening criterion which imposes different forms of regularization onto different image regions for better smoothing quality. We implement a diverse set of image smoothing solutions employing the unified framework targeting various applications such as, image abstraction, pencil sketching, detail enhancement, texture removal and content-aware image manipulation, and obtain results comparable with or better than previous methods. Moreover, our method is extremely fast with a modern GPU (e.g, 200 fps for 1280x720 images). Our codes and model are released in https://github.com/fqnchina/ImageSmoothing.
We describe a simple, low-level approach for embedding probabilistic programming in a deep learning ecosystem. In particular, we distill probabilistic programming down to a single abstraction---the random variable. Our lightweight implementation in TensorFlow enables numerous applications: a model-parallel variational auto-encoder (VAE) with 2nd-generation tensor processing units (TPUv2s); a data-parallel autoregressive model (Image Transformer) with TPUv2s; and multi-GPU No-U-Turn Sampler (NUTS). For both a state-of-the-art VAE on 64x64 ImageNet and Image Transformer on 256x256 CelebA-HQ, our approach achieves an optimal linear speedup from 1 to 256 TPUv2 chips. With NUTS, we see a 100x speedup on GPUs over Stan and 37x over PyMC3.
We propose an efficient Stereographic Projection Neural Network (SPNet) for learning representations of 3D objects. We first transform a 3D input volume into a 2D planar image using stereographic projection. We then present a shallow 2D convolutional neural network (CNN) to estimate the object category followed by view ensemble, which combines the responses from multiple views of the object to further enhance the predictions. Specifically, the proposed approach consists of four stages: (1) Stereographic projection of a 3D object, (2) view-specific feature learning, (3) view selection and (4) view ensemble. The proposed approach performs comparably to the state-of-the-art methods while having substantially lower GPU memory as well as network parameters. Despite its lightness, the experiments on 3D object classification and shape retrievals demonstrate the high performance of the proposed method.
In this paper we propose WaveGlow: a flow-based network capable of generating high quality speech from mel-spectrograms. WaveGlow combines insights from Glow and WaveNet in order to provide fast, efficient and high-quality audio synthesis, without the need for auto-regression. WaveGlow is implemented using only a single network, trained using only a single cost function: maximizing the likelihood of the training data, which makes the training procedure simple and stable. Our PyTorch implementation produces audio samples at a rate of more than 500 kHz on an NVIDIA V100 GPU. Mean Opinion Scores show that it delivers audio quality as good as the best publicly available WaveNet implementation. All code will be made publicly available online.
In trained deep neural networks, unstructured pruning can reduce redundant weights to lower storage cost. However, it requires the customization of hardwares to speed up practical inference. Another trend accelerates sparse model inference on general-purpose hardwares by adopting coarse-grained sparsity to prune or regularize consecutive weights for efficient computation. But this method often sacrifices model accuracy. In this paper, we propose a novel fine-grained sparsity approach, balanced sparsity, to achieve high model accuracy with commercial hardwares efficiently. Our approach adapts to high parallelism property of GPU, showing incredible potential for sparsity in the widely deployment of deep learning services. Experiment results show that balanced sparsity achieves up to 3.1x practical speedup for model inference on GPU, while retains the same high model accuracy as fine-grained sparsity.
The human brain consists of a large number of interconnected neurons communicating via exchange of electrical spikes. Simulations play an important role in better understanding electrical activity in the brain and offers a way to to compare measured data to simulated data such that experimental data can be interpreted better. A key component in such simulations is an efficient solver for the Hines matrices used in computing inter-neuron signal propagation. In order to achieve high performance simulations, it is crucial to have an efficient solver algorithm. In this report we explain a new parallel GPU solver for these matrices which offers fine grained parallelization and allows for work balancing during the simulation setup.
Big data powered Deep Learning (DL) and its applications have blossomed in recent years, fueled by three technological trends: a large amount of digitized data openly accessible, a growing number of DL software frameworks in open source and commercial markets, and a selection of affordable parallel computing hardware devices. However, no single DL framework, to date, dominates in terms of performance and accuracy even for baseline classification tasks on standard datasets, making the selection of a DL framework an overwhelming task. This paper takes a holistic approach to conduct empirical comparison and analysis of four representative DL frameworks with three unique contributions. First, given a selection of CPU-GPU configurations, we show that for a specific DL framework, different configurations of its hyper-parameters may have a significant impact on both performance and accuracy of DL applications. Second, to the best of our knowledge, this study is the first to identify the opportunities for improving the training time performance and the accuracy of DL frameworks by configuring parallel computing libraries and tuning individual and multiple hyper-parameters. Third, we also conduct a comparative measurement study on the resource consumption patterns of four DL frameworks and their performance and accuracy implications, including CPU and memory usage, and their correlations to varying settings of hyper-parameters under different configuration combinations of hardware, parallel computing libraries. We argue that this measurement study provides in-depth empirical comparison and analysis of four representative DL frameworks, and offers practical guidance for service providers to deploying and delivering DL as a Service (DLaaS) and for application developers and DLaaS consumers to select the right DL frameworks for the right DL workloads.