PyTorchKR Papers

최신 논문

총 4,618편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:1905.12253

Instant Quantization of Neural Networks using Monte Carlo Methods

Low bit-width integer weights and activations are very important for efficient inference, especially with respect to lower power consumption. We propose Monte Carlo methods to quantize the weights and activations of pre-trained neural networks without any re-training. By performing importance sampling we obtain quantized low bit-width integer values from full-precision weights and activations. The precision, sparsity, and complexity are easily configurable by the amount of sampling performed. Our approach, called Monte Carlo Quantization (MCQ), is linear in both time and space, with the resulting quantized, sparse networks showing minimal accuracy loss when compared to the original full-precision networks. Our method either outperforms or achieves competitive results on multiple benchmarks compared to previous quantization methods that do require additional training.

arXiv:1905.12171

Brain-inspired reverse adversarial examples

A human does not have to see all elephants to recognize an animal as an elephant. On contrast, current state-of-the-art deep learning approaches heavily depend on the variety of training samples and the capacity of the network. In practice, the size of network is always limited and it is impossible to access all the data samples. Under this circumstance, deep learning models are extremely fragile to human-imperceivable adversarial examples, which impose threats to all safety critical systems. Inspired by the association and attention mechanisms of the human brain, we propose reverse adversarial examples method that can greatly improve models' robustness on unseen data. Experiments show that our reverse adversarial method can improve accuracy on average 19.02% on ResNet18, MobileNet, and VGG16 on unseen data transformation. Besides, the proposed method is also applicable to compressed models and shows potential to compensate the robustness drop brought by model quantization - an absolute 30.78% accuracy improvement.

arXiv:1905.12003

Texture CNN for Thermoelectric Metal Pipe Image Classification

In this paper, the concept of representation learning based on deep neural networks is applied as an alternative to the use of handcrafted features in a method for automatic visual inspection of corroded thermoelectric metallic pipes. A texture convolutional neural network (TCNN) replaces handcrafted features based on Local Phase Quantization (LPQ) and Haralick descriptors (HD) with the advantage of learning an appropriate textural representation and the decision boundaries into a single optimization process. Experimental results have shown that it is possible to reach the accuracy of 99.20% in the task of identifying different levels of corrosion in the internal surface of thermoelectric pipe walls, while using a compact network that requires much less effort in tuning parameters when compared to the handcrafted approach since the TCNN architecture is compact regarding the number of layers and connections. The observed results open up the possibility of using deep neural networks in real-time applications such as the automatic inspection of thermoelectric metal pipes.

arXiv:1905.12334

Mixed Precision Training With 8-bit Floating Point

Reduced precision computation for deep neural networks is one of the key areas addressing the widening compute gap driven by an exponential growth in model size. In recent years, deep learning training has largely migrated to 16-bit precision, with significant gains in performance and energy efficiency. However, attempts to train DNNs at 8-bit precision have met with significant challenges because of the higher precision and dynamic range requirements of back-propagation. In this paper, we propose a method to train deep neural networks using 8-bit floating point representation for weights, activations, errors, and gradients. In addition to reducing compute precision, we also reduced the precision requirements for the master copy of weights from 32-bit to 16-bit. We demonstrate state-of-the-art accuracy across multiple data sets (imagenet-1K, WMT16) and a broader set of workloads (Resnet-18/34/50, GNMT, Transformer) than previously reported. We propose an enhanced loss scaling method to augment the reduced subnormal range of 8-bit floating point for improved error propagation. We also examine the impact of quantization noise on generalization and propose a stochastic rounding technique to address gradient noise. As a result of applying all these techniques, we report slightly higher validation accuracy compared to full precision baseline.

arXiv:1905.11478

Learning In Practice: Reasoning About Quantization

There is a mismatch between the standard theoretical analyses of statistical machine learning and how learning is used in practice. The foundational assumption supporting the theory is that we can represent features and models using real-valued parameters. In practice, however, we do not use real numbers at any point during training or deployment. Instead, we rely on discrete and finite quantizations of the reals, typically floating points. In this paper, we propose a framework for reasoning about learning under arbitrary quantizations. Using this formalization, we prove the convergence of quantization-aware versions of the Perceptron and Frank-Wolfe algorithms. Finally, we report the results of an extensive empirical study of the impact of quantization using a broad spectrum of datasets.

arXiv:1905.11452공식 코드 · Papers with Code

Mixed Precision DNNs: All you need is a good parametrization

Efficient deep neural network (DNN) inference on mobile or embedded devices typically involves quantization of the network parameters and activations. In particular, mixed precision networks achieve better performance than networks with homogeneous bitwidth for the same size constraint. Since choosing the optimal bitwidths is not straight forward, training methods, which can learn them, are desirable. Differentiable quantization with straight-through gradients allows to learn the quantizer's parameters using gradient methods. We show that a suited parametrization of the quantizer is the key to achieve a stable training and a good final performance. Specifically, we propose to parametrize the quantizer with the step size and dynamic range. The bitwidth can then be inferred from them. Other parametrizations, which explicitly use the bitwidth, consistently perform worse. We confirm our findings with experiments on CIFAR-10 and ImageNet and we obtain mixed precision DNNs with learned quantization parameters, achieving state-of-the-art performance.

arXiv:1905.10452공식 코드 · Papers with Code

Additive Noise Annealing and Approximation Properties of Quantized Neural Networks

We present a theoretical and experimental investigation of the quantization problem for artificial neural networks. We provide a mathematical definition of quantized neural networks and analyze their approximation capabilities, showing in particular that any Lipschitz-continuous map defined on a hypercube can be uniformly approximated by a quantized neural network. We then focus on the regularization effect of additive noise on the arguments of multi-step functions inherent to the quantization of continuous variables. In particular, when the expectation operator is applied to a non-differentiable multi-step random function, and if the underlying probability density is differentiable (in either classical or weak sense), then a differentiable function is retrieved, with explicit bounds on its Lipschitz constant. Based on these results, we propose a novel gradient-based training algorithm for quantized neural networks that generalizes the straight-through estimator, acting on noise applied to the network's parameters. We evaluate our algorithm on the CIFAR-10 and ImageNet image classification benchmarks, showing state-of-the-art performance on AlexNet and MobileNetV2 for ternary networks.

arXiv:1905.10936공식 코드 · Papers with Code

Communication-Efficient Distributed Blockwise Momentum SGD with Error-Feedback

Communication overhead is a major bottleneck hampering the scalability of distributed machine learning systems. Recently, there has been a surge of interest in using gradient compression to improve the communication efficiency of distributed neural network training. Using 1-bit quantization, signSGD with majority vote achieves a 32x reduction on communication cost. However, its convergence is based on unrealistic assumptions and can diverge in practice. In this paper, we propose a general distributed compressed SGD with Nesterov's momentum. We consider two-way compression, which compresses the gradients both to and from workers. Convergence analysis on nonconvex problems for general gradient compressors is provided. By partitioning the gradient into blocks, a blockwise compressor is introduced such that each gradient block is compressed and transmitted in 1-bit format with a scaling factor, leading to a nearly 32x reduction on communication. Experimental results show that the proposed method converges as fast as full-precision distributed momentum SGD and achieves the same testing accuracy. In particular, on distributed ResNet training with 7 workers on the ImageNet, the proposed algorithm achieves the same testing accuracy as momentum SGD using full-precision gradients, but with $46\%$ less wall clock time.

arXiv:1905.10759

HadaNets: Flexible Quantization Strategies for Neural Networks

On-board processing elements on UAVs are currently inadequate for training and inference of Deep Neural Networks. This is largely due to the energy consumption of memory accesses in such a network. HadaNets introduce a flexible train-from-scratch tensor quantization scheme by pairing a full precision tensor to a binary tensor in the form of a Hadamard product. Unlike wider reduced precision neural network models, we preserve the train-time parameter count, thus out-performing XNOR-Nets without a train-time memory penalty. Such training routines could see great utility in semi-supervised online learning tasks. Our method also offers advantages in model compression, as we reduce the model size of ResNet-18 by 7.43 times with respect to a full precision model without utilizing any other compression techniques. We also demonstrate a 'Hadamard Binary Matrix Multiply' kernel, which delivers a 10-fold increase in performance over full precision matrix multiplication with a similarly optimized kernel.

arXiv:1905.10988

Natural Compression for Distributed Deep Learning

Modern deep learning models are often trained in parallel over a collection of distributed machines to reduce training time. In such settings, communication of model updates among machines becomes a significant performance bottleneck and various lossy update compression techniques have been proposed to alleviate this problem. In this work, we introduce a new, simple yet theoretically and practically effective compression technique: natural compression (NC). Our technique is applied individually to all entries of the to-be-compressed update vector and works by randomized rounding to the nearest (negative or positive) power of two, which can be computed in a "natural" way by ignoring the mantissa. We show that compared to no compression, NC increases the second moment of the compressed vector by not more than the tiny factor $\frac{9}{8}$, which means that the effect of NC on the convergence speed of popular training algorithms, such as distributed SGD, is negligible. However, the communications savings enabled by NC are substantial, leading to $3$-$4\times$ improvement in overall theoretical running time. For applications requiring more aggressive compression, we generalize NC to natural dithering, which we prove is exponentially better than the common random dithering technique. Our compression operators can be used on their own or in combination with existing operators for a more aggressive combined effect and offer new state-of-the-art both in theory and practice.

arXiv:1905.11261

A Unified Theory of SGD: Variance Reduction, Sampling, Quantization and Coordinate Descent

In this paper we introduce a unified analysis of a large family of variants of proximal stochastic gradient descent ({\tt SGD}) which so far have required different intuitions, convergence analyses, have different applications, and which have been developed separately in various communities. We show that our framework includes methods with and without the following tricks, and their combinations: variance reduction, importance sampling, mini-batch sampling, quantization, and coordinate sub-sampling. As a by-product, we obtain the first unified theory of {\tt SGD} and randomized coordinate descent ({\tt RCD}) methods, the first unified theory of variance reduced and non-variance-reduced {\tt SGD} methods, and the first unified theory of quantized and non-quantized methods. A key to our approach is a parametric assumption on the iterates and stochastic gradients. In a single theorem we establish a linear convergence result under this assumption and strong-quasi convexity of the loss function. Whenever we recover an existing method as a special case, our theorem gives the best known complexity result. Our approach can be used to motivate the development of new useful methods, and offers pre-proved convergence guarantees. To illustrate the strength of our approach, we develop five new variants of {\tt SGD}, and through numerical experiments demonstrate some of their properties.

arXiv:1905.11062공식 코드 · Papers with Code

Quantization-Based Regularization for Autoencoders

Autoencoders and their variations provide unsupervised models for learning low-dimensional representations for downstream tasks. Without proper regularization, autoencoder models are susceptible to the overfitting problem and the so-called posterior collapse phenomenon. In this paper, we introduce a quantization-based regularizer in the bottleneck stage of autoencoder models to learn meaningful latent representations. We combine both perspectives of Vector Quantized-Variational AutoEncoders (VQ-VAE) and classical denoising regularization methods of neural networks. We interpret quantizers as regularizers that constrain latent representations while fostering a similarity-preserving mapping at the encoder. Before quantization, we impose noise on the latent codes and use a Bayesian estimator to optimize the quantizer-based representation. The introduced bottleneck Bayesian estimator outputs the posterior mean of the centroids to the decoder, and thus, is performing soft quantization of the noisy latent codes. We show that our proposed regularization method results in improved latent representations for both supervised learning and clustering downstream tasks when compared to autoencoders using other bottleneck structures.

arXiv:1905.10138

Structured Compression by Weight Encryption for Unstructured Pruning and Quantization

Model compression techniques, such as pruning and quantization, are becoming increasingly important to reduce the memory footprints and the amount of computations. Despite model size reduction, achieving performance enhancement on devices is, however, still challenging mainly due to the irregular representations of sparse matrix formats. This paper proposes a new weight representation scheme for Sparse Quantized Neural Networks, specifically achieved by fine-grained and unstructured pruning method. The representation is encrypted in a structured regular format, which can be efficiently decoded through XOR-gate network during inference in a parallel manner. We demonstrate various deep learning models that can be compressed and represented by our proposed format with fixed and high compression ratio. For example, for fully-connected layers of AlexNet on ImageNet dataset, we can represent the sparse weights by only 0.28 bits/weight for 1-bit quantization and 91% pruning rate with a fixed decoding rate and full memory bandwidth usage. Decoding through XOR-gate network can be performed without any model accuracy degradation with additional patch data associated with small overhead.

arXiv:1905.09063

NTP : A Neural Network Topology Profiler

Performance of end-to-end neural networks on a given hardware platform is a function of its compute and memory signature, which in-turn, is governed by a wide range of parameters such as topology size, primitives used, framework used, batching strategy, latency requirements, precision etc. Current benchmarking tools suffer from limitations such as a) being either too granular like DeepBench [1] (or) b) mandate a working implementation that is either framework specific or hardware-architecture specific or both (or) c) provide only high level benchmark metrics. In this paper, we present NTP (Neural Net Topology Profiler), a sophisticated benchmarking framework, to effectively identify memory and compute signature of an end-to-end topology on multiple hardware architectures, without the need for an actual implementation. NTP is tightly integrated with hardware specific benchmarking tools to enable exhaustive data collection and analysis. Using NTP, a deep learning researcher can quickly establish baselines needed to understand performance of an end-to-end neural network topology and make high level architectural decisions. Further, integration of NTP with frameworks like Tensorflow, Pytorch, Intel OpenVINO etc. allows for performance comparison along several vectors like a) Comparison of different frameworks on a given hardware b) Comparison of different hardware using a given framework c) Comparison across different heterogeneous hardware configurations for given framework etc. These capabilities empower a researcher to effortlessly make architectural decisions needed for achieving optimized performance on any hardware platform. The paper documents the architectural approach of NTP and demonstrates the capabilities of the tool by benchmarking Mozilla DeepSpeech, a popular Speech Recognition topology.

arXiv:1905.08318

DeepCABAC: Context-adaptive binary arithmetic coding for deep neural network compression

We present DeepCABAC, a novel context-adaptive binary arithmetic coder for compressing deep neural networks. It quantizes each weight parameter by minimizing a weighted rate-distortion function, which implicitly takes the impact of quantization on to the accuracy of the network into account. Subsequently, it compresses the quantized values into a bitstream representation with minimal redundancies. We show that DeepCABAC is able to reach very high compression ratios across a wide set of different network architectures and datasets. For instance, we are able to compress by x63.6 the VGG16 ImageNet model with no loss of accuracy, thus being able to represent the entire network with merely 8.7MB.

arXiv:1905.06474

Optimizing MRF-ASL Scan Design for Precise Quantification of Brain Hemodynamics using Neural Network Regression

Purpose: Arterial Spin Labeling (ASL) is a quantitative, non-invasive alternative to perfusion imaging with contrast agents. Fixing values of certain model parameters in traditional ASL, which actually vary from region to region, may introduce bias in perfusion estimates. Adopting Magnetic Resonance Fingerprinting (MRF) for ASL is an alternative where these parameters are estimated alongside perfusion, but multiparametric estimation can degrade precision. We aim to improve the sensitivity of ASL-MRF signals to underlying parameters to counter this problem, and provide precise estimates. We also propose a regression based estimation framework for MRF-ASL. Methods: To improve the sensitivity of MRF-ASL signals to underlying parameters, we optimize ASL labeling durations using the Cramer-Rao Lower Bound (CRLB). This paper also proposes a neural network regression based estimation framework trained using noisy synthetic signals generated from our ASL signal model. Results: We test our methods in silico and in vivo, and compare with multiple post labeling delay (multi-PLD) ASL and unoptimized MRF-ASL. We present comparisons of estimated maps for six parameters accounted for in our signal model. Conclusions: The scan design process facilitates precise estimates of multiple hemodynamic parameters and tissue properties from a single scan, in regions of gray and white matter, as well as regions with anomalous perfusion activity in the brain. The regression based estimation approach provides perfusion estimates rapidly, and bypasses problems with quantization error. Keywords: Arterial Spin Labeling, Magnetic Resonance Fingerprinting, Optimization, Cramer-Rao Bound, Scan Design, Regression, Neural Networks, Deep Learning, Precision, Estimation, Brain Hemodynamics.

arXiv:1905.06900

Derived Codebooks for High-Accuracy Nearest Neighbor Search

High-dimensional Nearest Neighbor (NN) search is central in multimedia search systems. Product Quantization (PQ) is a widespread NN search technique which has a high performance and good scalability. PQ compresses high-dimensional vectors into compact codes thanks to a combination of quantizers. Large databases can, therefore, be stored entirely in RAM, enabling fast responses to NN queries. In almost all cases, PQ uses 8-bit quantizers as they offer low response times. In this paper, we advocate the use of 16-bit quantizers. Compared to 8-bit quantizers, 16-bit quantizers boost accuracy but they increase response time by a factor of 3 to 10. We propose a novel approach that allows 16-bit quantizers to offer the same response time as 8-bit quantizers, while still providing a boost of accuracy. Our approach builds on two key ideas: (i) the construction of derived codebooks that allow a fast and approximate distance evaluation, and (ii) a two-pass NN search procedure which builds a candidate set using the derived codebooks, and then refines it using 16-bit quantizers. On 1 billion SIFT vectors, with an inverted index, our approach offers a Recall@100 of 0.85 in 5.2 ms. By contrast, 16-bit quantizers alone offer a Recall@100 of 0.85 in 39 ms, and 8-bit quantizers a Recall@100 of 0.82 in 3.8 ms.

arXiv:1905.06118

Learning to Groove with Inverse Sequence Transformations

We explore models for translating abstract musical ideas (scores, rhythms) into expressive performances using Seq2Seq and recurrent Variational Information Bottleneck (VIB) models. Though Seq2Seq models usually require painstakingly aligned corpora, we show that it is possible to adapt an approach from the Generative Adversarial Network (GAN) literature (e.g. Pix2Pix (Isola et al., 2017) and Vid2Vid (Wang et al. 2018a)) to sequences, creating large volumes of paired data by performing simple transformations and training generative models to plausibly invert these transformations. Music, and drumming in particular, provides a strong test case for this approach because many common transformations (quantization, removing voices) have clear semantics, and models for learning to invert them have real-world applications. Focusing on the case of drum set players, we create and release a new dataset for this purpose, containing over 13 hours of recordings by professional drummers aligned with fine-grained timing and dynamics information. We also explore some of the creative potential of these models, including demonstrating improvements on state-of-the-art methods for Humanization (instantiating a performance from a musical score).

arXiv:1905.04129

Two-layer Near-lossless HDR Coding with Backward Compatibility to JPEG

We propose an efficient two-layer near-lossless coding method using an extended histogram packing technique with backward compatibility to the legacy JPEG standard. The JPEG XT, which is the international standard to compress HDR images, adopts a two-layer coding method for backward compatibility to the legacy JPEG standard. However, there are two problems with this two-layer coding method. One is that it does not exhibit better near-lossless performance than other methods for HDR image compression with single-layer structure. The other problem is that the determining the appropriate values of the coding parameters may be required for each input image to achieve good compression performance of near-lossless compression with the two-layer coding method of the JPEG XT. To solve these problems, we focus on a histogram-packing technique that takes into account the histogram sparseness of HDR images. We used zero-skip quantization, which is an extension of the histogram-packing technique proposed for lossless coding, for implementing the proposed near-lossless coding method. The experimental results indicate that the proposed method exhibits not only a better near-lossless compression performance than that of the two-layer coding method of the JPEG XT, but also there are no issue regarding the combination of parameter values without losing backward compatibility to the JPEG standard.

arXiv:1905.04079

Compressing Weight-updates for Image Artifacts Removal Neural Networks

In this paper, we present a novel approach for fine-tuning a decoder-side neural network in the context of image compression, such that the weight-updates are better compressible. At encoder side, we fine-tune a pre-trained artifact removal network on target data by using a compression objective applied on the weight-update. In particular, the compression objective encourages weight-updates which are sparse and closer to quantized values. This way, the final weight-update can be compressed more efficiently by pruning and quantization, and can be included into the encoded bitstream together with the image bitstream of a traditional codec. We show that this approach achieves reconstruction quality which is on-par or slightly superior to a traditional codec, at comparable bitrates. To our knowledge, this is the first attempt to combine image compression and neural network's weight update compression.