MOTIVATION: With the advancement of cryo-electron microscopy (cryo-EM) into the atomic resolution era, accurate Cα atom modeling has become essential for macromolecular structure determination. However, existing evaluation systems overly rely on full-atom metrics and lack a dedicated, comprehensive benchmark for assessing Cα prediction modules within automated modeling tools. RESULTS: To address this gap, we establish a rigorous benchmark to evaluate the Cα prediction performance of four prominent deep learning-based methods (ModelAngelo, DeepMainMast, EModelX, and CryoAtom) across multiple dimensions. We construct a diverse dataset covering a wide range of resolutions (1-8 Å), molecular weights, and noise levels. A novel evaluation framework is introduced, incorporating multi-threshold RMSD-based metrics (1-3 Å) alongside advanced point-cloud similarity measures (Chamfer Distance, Earth Mover's Distance) for quantitative and nuanced assessment. Our results reveal that method performance is highly dependent on the chosen evaluation criteria and intrinsic data characteristics. ModelAngelo excels under loose thresholds with high-quality data but shows sensitivity to resolution degradation; CryoAtom demonstrates notable computational efficiency, however, its completeness-oriented design leads to a certain loss of precision; EModelX demonstrates balanced generalization across varied conditions; DeepMainMast achieves high localization accuracy under stringent criteria but incurs a high computational cost. AVAILABILITY AND IMPLEMENTATION: This work provides a reproducible, Cα-centric evaluation framework to guide method development and advance automated cryo-EM structure determination. The source code for the benchmark and evaluation metrics is freely available at https://github.com/zhtianz/Benchmarking\_CA.

ByteScheduler partitions and rearranges tensor transmissions to improve the communication efficiency of distributed Deep Neural Network (DNN) training. The configuration of hyper-parameters (i.e., the partition size and the credit size) is critical to the effectiveness of partitioning and rearrangement. Currently ByteScheduler adopts Bayesian Optimization (BO) to find the optimal configuration for the hyper-parameters beforehand. In practice, however, various runtime factors (such as worker node status and network conditions) change over time, making the statically-determined one-shot configuration result suboptimal for real-world DNN training. To address this problem, in this paper we present a realtime configuration method (called AutoByte) that automatically and timely searches the optimal hyper-parameters as the training systems dynamically change. AutoByte extends the ByteScheduler framework with a meta network, which takes the systems’ runtime statistics as its input, dynamically adjusts the triggering threshold based on system environment characteristics, and outputs predictions for speedups under specific configurations. Evaluation results on various DNN models show that AutoByte can dynamically tune the hyper-parameters with low resource usage, and deliver up to 33.2% higher performance than the best static configuration method on the ByteScheduler framework.

In high-power DC transformer systems, large-capacity medium-frequency transformers serve as critical components for energy conversion and electrical isolation, with their electrical performance and structural design directly influencing overall system efficiency and reliability. To address the challenges of increased losses, localized heating, and insulation stress under medium-frequency, non-sinusoidal excitation, this paper presents the optimized design and prototype development of a 4.1 MVA, 1 kHz MFT. The study focuses on material selection and structural refinement, including the use of amorphous alloy cores, copper foil windings, insulation configuration, and thermal management. Based on the proposed design methodology, a dry-type, high turns-ratio MFT prototype rated at 1.5 kV/10.5 kV, 4.1 MVA, and 1 kHz was successfully developed. Experimental results show that the transformer achieves an efficiency exceeding 99.2% under full-power operation, demonstrating the effectiveness of the proposed design approach.

This paper investigates the loss behavior of the clamping circuit in a high step-up resonant DC transformer (IGCT-SRDCT) based on integrated gate commutated thyristor (IGCT) and series-connected diodes. The proposed topology features a simple structure, high voltage gain, and high efficiency, making it highly suitable for medium-voltage DC collection in renewable energy systems. Due to the zero-current switching (ZCS) characteristics of the IGCT-SRDCT, the clamping circuit is triggered only during the zero-current turn-on of IGCT, where it suppresses the currents caused by the parasitic junction capacitor of the IGCT and diode. A theoretical analysis is conducted to examine the triggering behavior and energy dissipation of the clamping circuit under varying load current rise rates (di/dt). A mathematical model is developed based on the peak current of the clamping circuit to estimate the clamping losses. The analysis reveals that increasing di/dt following zero-current turn-on of IGCT significantly reduces clamping loss. A prototype of the 1.5kV/30kV/10MW IGCT-SRDCT was developed to validate the proposed theory, demonstrating that the clamping loss is limited to only 72 W.

This paper considers the problem of global asymptotic stabilization (GAS) for second-order linearly uncontrol-lable/unobservable systems with long delays in state and input by designing a sampled-data homogeneous feedback controller. To transform a large delay nonlinear integrator chain into a small delay nonlinear system, the time rescaling and nonsingular transformations are presented. The homogeneous domination approach and Lyapunov-Krasovskii(L-K) stability theorem are employed in building a sampled-data homogeneous controller that achieves GAS of the time-delayed nonlinear systems. In order to verify the effectiveness of the algorithm, a simulation is conducted.

Decades of research on Internet congestion control (CC) have produced a plethora of algorithms that optimize for different performance objectives. Applications face the challenge of choosing the most suitable algorithm based on their needs, and it takes tremendous efforts and expertise to customize CC algorithms when new demands emerge. In this paper, we explore a basic question: can we design a single CC algorithm to satisfy different objectives?

ByteScheduler partitions and rearranges tensor transmissions to improve the communication efficiency of distributed Deep Neural Network (DNN) training. The configuration of hyper-parameters (i.e., the partition size and the credit size) is critical to the effectiveness of partitioning and rearrangement. Currently, ByteScheduler adopts Bayesian Optimization (BO) to find the optimal configuration for the hyper-parameters beforehand. In practice, however, various runtime factors (e.g., worker node status and network conditions) change over time, making the statically-determined one-shot configuration result suboptimal for real-world DNN training.To address this problem, we present a real-time configuration method (called AutoByte) that automatically and timely searches the optimal hyper-parameters as the training systems dynamically change. AutoByte extends the ByteScheduler framework with a meta-network, which takes the system’s runtime statistics as its input and outputs predictions for speedups under specific configurations. Evaluation results on various DNN models show that AutoByte can dynamically tune the hyper-parameters with low resource usage, and deliver up to 33.2% higher performance than the best static configuration in ByteScheduler.

Modeling hairstyles for classification, synthesis and image editing has many practical applications. However, existing hairstyle datasets, such as the Beauty e-Expert dataset, are too small for developing and evaluating computer vision models, especially the recent deep generative models such as generative adversarial network (GAN). In this paper, we contribute a new large-scale hairstyle dataset called Hairstyle30k, which is composed of 30k images containing 64 different types of hairstyles. To enable automated generating and modifying hairstyles in images, we also propose a novel GAN model termed Hairstyle GAN (H-GAN) which can be learned efficiently. Extensive experiments on the new dataset as well as existing benchmark datasets demonstrate the effectiveness of proposed H-GAN model