Currently, the neural network architecture design is mostly guided by the indirect metric of computation complexity, i.e., FLOPs. However, the direct metric, e.g., speed, also depends on the other factors such as memory access cost and platform characterics. Thus, this work proposes to evaluate the direct metric on the target platform, beyond only considering FLOPs. Based on a series of controlled experiments, this work derives several practical guidelines for efficient network design. Accordingly, a new architecture is presented, called ShuffleNet V2. Comprehensive ablation experiments verify that our model is the state-of-the-art in terms of speed and accuracy tradeoff.

/pdf/shufflenet-v2-practical-guidelines-for-efficient-cnn-36xrqao62p.pdf

ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Deep Learning has enabled remarkable progress over the last years on a variety of tasks, such as image recognition, speech recognition, and machine translation. One crucial aspect for this progress are novel neural architectures. Currently employed architectures have mostly been developed manually by human experts, which is a time-consuming and error-prone process. Because of this, there is growing interest in automated neural architecture search methods. We provide an overview of existing work in this field of research and categorize them according to three dimensions: search space, search strategy, and performance estimation strategy.

Neural Architecture Search: A Survey

Network pruning is widely used for reducing the heavy inference cost of deep models in low-resource settings. A typical pruning algorithm is a three-stage pipeline, i.e., training (a large model), pruning and fine-tuning. During pruning, according to a certain criterion, redundant weights are pruned and important weights are kept to best preserve the accuracy. In this work, we make several surprising observations which contradict common beliefs. For all state-of-the-art structured pruning algorithms we examined, fine-tuning a pruned model only gives comparable or worse performance than training that model with randomly initialized weights. For pruning algorithms which assume a predefined target network architecture, one can get rid of the full pipeline and directly train the target network from scratch. Our observations are consistent for multiple network architectures, datasets, and tasks, which imply that: 1) training a large, over-parameterized model is often not necessary to obtain an efficient final model, 2) learned "important" weights of the large model are typically not useful for the small pruned model, 3) the pruned architecture itself, rather than a set of inherited "important" weights, is more crucial to the efficiency in the final model, which suggests that in some cases pruning can be useful as an architecture search paradigm. Our results suggest the need for more careful baseline evaluations in future research on structured pruning methods. We also compare with the "Lottery Ticket Hypothesis" (Frankle & Carbin 2019), and find that with optimal learning rate, the "winning ticket" initialization as used in Frankle & Carbin (2019) does not bring improvement over random initialization.

Rethinking the Value of Network Pruning

Deep neural networks (DNNs) have recently achieved great success in many visual recognition tasks. However, existing deep neural network models are computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, a natural thought is to perform model compression and acceleration in deep networks without significantly decreasing the model performance. During the past five years, tremendous progress has been made in this area. In this paper, we review the recent techniques for compacting and accelerating DNN models. In general, these techniques are divided into four categories: parameter pruning and quantization, low-rank factorization, transferred/compact convolutional filters, and knowledge distillation. Methods of parameter pruning and quantization are described first, after that the other techniques are introduced. For each category, we also provide insightful analysis about the performance, related applications, advantages, and drawbacks. Then we go through some very recent successful methods, for example, dynamic capacity networks and stochastic depths networks. After that, we survey the evaluation matrices, the main datasets used for evaluating the model performance, and recent benchmark efforts. Finally, we conclude this paper, discuss remaining the challenges and possible directions for future work.

A Survey of Model Compression and Acceleration for Deep Neural Networks

This paper presents a method for adding multiple tasks to a single deep neural network while avoiding catastrophic forgetting. Inspired by network pruning techniques, we exploit redundancies in large deep networks to free up parameters that can then be employed to learn new tasks. By performing iterative pruning and network re-training, we are able to sequentially "pack" multiple tasks into a single network while ensuring minimal drop in performance and minimal storage overhead. Unlike prior work that uses proxy losses to maintain accuracy on older tasks, we always optimize for the task at hand. We perform extensive experiments on a variety of network architectures and large-scale datasets, and observe much better robustness against catastrophic forgetting than prior work. In particular, we are able to add three fine-grained classification tasks to a single ImageNet-trained VGG-16 network and achieve accuracies close to those of separately trained networks for each task.

/pdf/packnet-adding-multiple-tasks-to-a-single-network-by-1lm11zdaq4.pdf

PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning

The deployment of deep convolutional neural networks (CNNs) in many real world applications is largely hindered by their high computational cost. In this paper, we propose a novel learning scheme for CNNs to simultaneously 1) reduce the model size; 2) decrease the run-time memory footprint; and 3) lower the number of computing operations, without compromising accuracy. This is achieved by enforcing channel-level sparsity in the network in a simple but effective way. Different from many existing approaches, the proposed method directly applies to modern CNN architectures, introduces minimum overhead to the training process, and requires no special software/hardware accelerators for the resulting models. We call our approach network slimming, which takes wide and large networks as input models, but during training insignificant channels are automatically identified and pruned afterwards, yielding thin and compact models with comparable accuracy. We empirically demonstrate the effectiveness of our approach with several state-of-the-art CNN models, including VGGNet, ResNet and DenseNet, on various image classification datasets. For VGGNet, a multi-pass version of network slimming gives a 20× reduction in model size and a 5× reduction in computing operations.

Learning Efficient Convolutional Networks through Network Slimming

The deployment of deep convolutional neural networks (CNNs) in many real world applications is largely hindered by their high computational cost. In this paper, we propose a novel learning scheme for CNNs to simultaneously 1) reduce the model size; 2) decrease the run-time memory footprint; and 3) lower the number of computing operations, without compromising accuracy. This is achieved by enforcing channel-level sparsity in the network in a simple but effective way. Different from many existing approaches, the proposed method directly applies to modern CNN architectures, introduces minimum overhead to the training process, and requires no special software/hardware accelerators for the resulting models. We call our approach network slimming, which takes wide and large networks as input models, but during training insignificant channels are automatically identified and pruned afterwards, yielding thin and compact models with comparable accuracy. We empirically demonstrate the effectiveness of our approach with several state-of-the-art CNN models, including VGGNet, ResNet and DenseNet, on various image classification datasets. For VGGNet, a multi-pass version of network slimming gives a 20x reduction in model size and a 5x reduction in computing operations.

Embodiments of the invention provide a programming model for CPU-GPU platforms. In particular, embodiments of the invention provide a uniform programming model for both integrated and discrete devices. The model also works uniformly for multiple GPU cards and hybrid GPU systems (discrete and integrated). This allows software vendors to write a single application stack and target it to all the different platforms. Additionally, embodiments of the invention provide a shared memory model between the CPU and GPU. Instead of sharing the entire virtual address space, only a part of the virtual address space needs to be shared. This allows efficient implementation in both discrete and integrated settings.

Shared virtual memory

The client computing platform is moving towards a heterogeneous architecture consisting of a combination of cores focused on scalar performance, and a set of throughput-oriented cores. The throughput oriented cores (e.g. a GPU) may be connected over both coherent and non-coherent interconnects, and have different ISAs. This paper describes a programming model for such heterogeneous platforms. We discuss the language constructs, runtime implementation, and the memory model for such a programming environment. We implemented this programming environment in a x86 heterogeneous platform simulator. We ported a number of workloads to our programming environment, and present the performance of our programming environment on these workloads.

Programming model for a heterogeneous x86 platform

JavaScript frameworks, such as jQuery, are widely used for developing web applications. To facilitate using these JavaScript frameworks to implement a feature (e.g., functionality), a large number of programmers often search for code snippets that implement the same or similar feature. However, existing code search approaches tend to be ineffective, without taking into account the fact that JavaScript code snippets often implement a feature based on various relationships (e.g., sequencing, condition, and callback relationships) among the invoked framework API methods. To address this issue, we present a novel Relationship-Aware Code Search (RACS) approach for finding code snippets that use JavaScript frameworks to implement a specific feature. In advance, RACS collects a large number of code snippets that use some JavaScript frameworks, mines API usage patterns from the collected code snippets, and represents the mined patterns with method call relationship (MCR) graphs, which capture framework API methods’ signatures and their relationships. Given a natural language (NL) search query issued by a programmer, RACS conducts NL processing to automatically extract an action relationship (AR) graph, which consists of actions and their relationships inferred from the query. In this way, RACS reduces code search to the problem of graph search: finding similar MCR graphs for a given AR graph. We conduct evaluations against representative real-world jQuery questions posted on Stack Overflow, based on 308,294 code snippets collected from over 81,540 files on the Internet. The evaluation results show the effectiveness of RACS: the top 1 snippet produced by RACS matches the target code snippet for 46% questions, compared to only 4% achieved by a relationship-oblivious approach.

Shoumeng Yan

Papers

Learning Efficient Convolutional Networks through Network Slimming

Learning Efficient Convolutional Networks through Network Slimming

Shared virtual memory

Programming model for a heterogeneous x86 platform

Relationship-aware code search for JavaScript frameworks