Search

Conditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more compute-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method finds efficient architectures via neural architecture search. To accelerate the search process, we decouple the model training and search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings, network architectures, and learning methods. Without losing image quality, we reduce the computation of CycleGAN by 21×, Pix2pix by 12×, MUNIT by 29×, and GauGAN by 9×, paving the way for interactive image synthesis. [ABSTRACT FROM AUTHOR]

Conditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more compute-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method finds efficient architectures via neural architecture search. To accelerate the search process, we decouple the model training and search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings, network architectures, and learning methods. Without losing image quality, we reduce the computation of CycleGAN by 21×, Pix2pix by 12×, MUNIT by 29×, and GauGAN by 9×, paving the way for interactive image synthesis. [ABSTRACT FROM AUTHOR]

Polarization information can greatly improve the ability of detection, parameter retrieval, and classification for entomological radar. In recent, the instantaneous fully polarimetric measurement technique is used in entomological radar, so as to relax the restriction of invariant orientation during observation compared to the typical vertical-looking radar with linear-polarized rotation configuration. However, the fully polarimetric measurement is based on a multichannels’ radar system, and the imbalance and crosstalk between channels, namely, polarimetric errors, will severely affect the measurement of the polarimetric scattering matrix (PSM). Thus, polarimetric calibration is essential to obtain the accurate polarization information of targets. In contrast to inefficient and manual calibration methods, a data-driven polarimetric calibration method is proposed based on the assumption of reciprocity and bilateral symmetry of insects in this article. The polarimetric errors are elaborately decomposed into two components, which could be estimated by reciprocity and bilateral symmetry, respectively. In addition, because the proposed method is data-driven, the calibration process could be automatically executed to compensate for the temporal-variant polarimetric errors induced by temperature change and, thus, suitable for the long-term operation of entomological radar. Simulations and experiments are carried out to evaluate the performance of the proposed calibration method. Results show that it could achieve high accuracy when the number of insects used as calibrators is large enough. The proposed method has been applied in Ku-band high-resolution fully polarimetric entomological radars for cross-border migratory insect observation in Yunnan province, China. [ABSTRACT FROM AUTHOR]

In this paper, a robust parametric model of a brushless permanent magnet machine with fractional-slot concentrated windings, which was developed for automated design optimization is presented. A computationally efficient finite-element analysis method was employed to estimate the dq-axes inductances, the induced voltage and torque ripple waveforms, and losses of the machine. A method for minimum effort calculation of the torque angle corresponding to the maximum torque per ampere load condition was developed. A differential evolution algorithm was implemented for the global design optimization with two concurrent objectives of minimum losses and minimum material cost. An engineering decision process based on the Pareto-optimal front for 3,500 candidate designs is presented together with discussions on the tradeoffs between cost and performance. One optimal design was finally selected, prototyped and successfully tested. [ABSTRACT FROM AUTHOR]

In this paper, we present generic cloud performance models for evaluating Iaas, PaaS, SaaS, and mashup or hybrid clouds. We test clouds with real-life benchmark programs and propose some new performance metrics. Our benchmark experiments are conducted mainly on IaaS cloud platforms over scale-out and scale-up workloads. Cloud benchmarking results are analyzed with the efficiency, elasticity, QoS, productivity, and scalability of cloud performance. Five cloud benchmarks were tested on Amazon IaaS EC2 cloud: namely YCSB, CloudSuite, HiBench, BenchClouds, and TPC-W. To satisfy production services, the choice of scale-up or scale-out solutions should be made primarily by the workload patterns and resources utilization rates required. Scaling-out machine instances have much lower overhead than those experienced in scale-up experiments. However, scaling up is found more cost-effective in sustaining heavier workload. The cloud productivity is greatly attributed to system elasticity, efficiency, QoS and scalability. We find that auto-scaling is easy to implement but tends to over provision the resources. Lower resource utilization rate may result from auto-scaling, compared with using scale-out or scale-up strategies. We also demonstrate that the proposed cloud performance models are applicable to evaluate PaaS, SaaS and hybrid clouds as well. [ABSTRACT FROM AUTHOR]