Pub Date : 2020-11-01DOI: 10.1109/DLS51937.2020.00010
Alberto Acevedo, Michael Curry, Shantanu H. Joshi, Brett Leroux, Nicholas Malaya
Solutions to the Schrödinger equation can be used to predict the electronic structure of molecules and materials and therefore infer their complex physical and chemical properties. Variational Quantum Monte Carlo (VMC) is a technique that can be used to solve the weak form of the Schrödinger equation. Applying VMC to systems with N electrons involves evaluating the determinant of an N by N matrix. The evaluation of this determinant scales as $O(N^{3})$ and is the main computational cost in the VMC process. In this work, we investigate an alternative VMC technique based on the Vandermonde determinant. The Vandermonde determinant is a product of pairwise differences and so evaluating it scales as $O(N^{2})$. Therefore, this approach reduces the computational cost by a factor of N. The Vandermonde determinant was implemented in PyTorch and the performance was assessed in approximating the ground state energy of various quantum systems against existing techniques. The performance is evaluated in a variety of systems, starting with the one-dimensional particle in a box, and then considering more complicated atomic systems with multiple particles. The Vandermonde determinant was also implemented in PauliNet, a deep-learning architecture for VMC. The new method is shown to be computationally efficient, and results in a speed-up as large as 5X. In these cases, the new ansatz obtains a reasonable approximation for wavefunctions of atomic systems, but does not reach the accuracy of the Hartree-Fock method that relies on the Slater determinant. It is observed that while the use of neural networks in VMC can result in highly accurate solutions, further work is necessary to determine an appropriate balance between computational time and accuracy.
{"title":"Vandermonde Wave Function Ansatz for Improved Variational Monte Carlo","authors":"Alberto Acevedo, Michael Curry, Shantanu H. Joshi, Brett Leroux, Nicholas Malaya","doi":"10.1109/DLS51937.2020.00010","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00010","url":null,"abstract":"Solutions to the Schrödinger equation can be used to predict the electronic structure of molecules and materials and therefore infer their complex physical and chemical properties. Variational Quantum Monte Carlo (VMC) is a technique that can be used to solve the weak form of the Schrödinger equation. Applying VMC to systems with N electrons involves evaluating the determinant of an N by N matrix. The evaluation of this determinant scales as $O(N^{3})$ and is the main computational cost in the VMC process. In this work, we investigate an alternative VMC technique based on the Vandermonde determinant. The Vandermonde determinant is a product of pairwise differences and so evaluating it scales as $O(N^{2})$. Therefore, this approach reduces the computational cost by a factor of N. The Vandermonde determinant was implemented in PyTorch and the performance was assessed in approximating the ground state energy of various quantum systems against existing techniques. The performance is evaluated in a variety of systems, starting with the one-dimensional particle in a box, and then considering more complicated atomic systems with multiple particles. The Vandermonde determinant was also implemented in PauliNet, a deep-learning architecture for VMC. The new method is shown to be computationally efficient, and results in a speed-up as large as 5X. In these cases, the new ansatz obtains a reasonable approximation for wavefunctions of atomic systems, but does not reach the accuracy of the Hartree-Fock method that relies on the Slater determinant. It is observed that while the use of neural networks in VMC can result in highly accurate solutions, further work is necessary to determine an appropriate balance between computational time and accuracy.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"31 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123859655","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Data parallel training is a powerful family of methods for the efficient training of deep neural networks on big data. Unfortunately, however, recent studies have shown that the merit of increased batch size in terms of both speed and model-performance diminishes rapidly beyond some point. This seem to apply to even LARS, the state-of-the-art large batch stochastic optimization method. In this paper, we combine LARS with online-codistillation, a recently developed, efficient deep learning algorithm built on a whole different philosophy of stabilizing the training procedure using a collaborative ensemble of models. We show that the combination of large-batch training and online-codistillation is much more efficient than either one alone. We also present a novel way of implementing the online-codistillation that can further speed up the computation. We will demonstrate the efficacy of our approach on various benchmark datasets.
{"title":"Online-Codistillation Meets LARS, Going beyond the Limit of Data Parallelism in Deep Learning","authors":"Shogo Murai, Hiroaki Mikami, Masanori Koyama, Shuji Suzuki, Takuya Akiba","doi":"10.1109/DLS51937.2020.00006","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00006","url":null,"abstract":"Data parallel training is a powerful family of methods for the efficient training of deep neural networks on big data. Unfortunately, however, recent studies have shown that the merit of increased batch size in terms of both speed and model-performance diminishes rapidly beyond some point. This seem to apply to even LARS, the state-of-the-art large batch stochastic optimization method. In this paper, we combine LARS with online-codistillation, a recently developed, efficient deep learning algorithm built on a whole different philosophy of stabilizing the training procedure using a collaborative ensemble of models. We show that the combination of large-batch training and online-codistillation is much more efficient than either one alone. We also present a novel way of implementing the online-codistillation that can further speed up the computation. We will demonstrate the efficacy of our approach on various benchmark datasets.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"41 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116942769","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-11-01DOI: 10.1109/DLS51937.2020.00011
H. Venkataram, C. Mattmann, Scott Penberthy
We all have questions. About today’s temperature, scores of our favorite baseball team, the Universe, and about vaccine for COVID-19. Life, physical, and natural scientists have been trying to find answers to various topics using scientific methods and experiments, while computer scientists have built language models as a tiny step towards automatically answering all of these questions across domains given a little bit of context. In this paper, we propose an architecture using state-of-the-art Natural Language Processing language models namely Topic Models and Bidirectional Encoder Representations from Transformers (BERT) that can transparently and automatically retrieve articles of relevance to questions across domains, and fetch answers to topical questions related to COVID-19 current and historical medical research literature. We demonstrate the benefits of using domain-specific supercomputers like Tensor Processing Units (TPUs), residing on cloud-based infrastructure, using which we could achieve significant gains in training and inference times, also with very minimal cost.
{"title":"TopiQAL: Topic-aware Question Answering using Scalable Domain-specific Supercomputers","authors":"H. Venkataram, C. Mattmann, Scott Penberthy","doi":"10.1109/DLS51937.2020.00011","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00011","url":null,"abstract":"We all have questions. About today’s temperature, scores of our favorite baseball team, the Universe, and about vaccine for COVID-19. Life, physical, and natural scientists have been trying to find answers to various topics using scientific methods and experiments, while computer scientists have built language models as a tiny step towards automatically answering all of these questions across domains given a little bit of context. In this paper, we propose an architecture using state-of-the-art Natural Language Processing language models namely Topic Models and Bidirectional Encoder Representations from Transformers (BERT) that can transparently and automatically retrieve articles of relevance to questions across domains, and fetch answers to topical questions related to COVID-19 current and historical medical research literature. We demonstrate the benefits of using domain-specific supercomputers like Tensor Processing Units (TPUs), residing on cloud-based infrastructure, using which we could achieve significant gains in training and inference times, also with very minimal cost.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132474429","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-11-01DOI: 10.1109/DLS51937.2020.00009
Matthijs Jansen, V. Codreanu, A. Varbanescu
Due to its many applications across various fields of research, engineering, and daily life, deep learning has seen a surge in popularity. Therefore, larger and more expressive models have been proposed, with examples like Turing-NLG using as many as 17 billion parameters. Training these very large models becomes increasingly difficult due to the high computational costs and large memory footprint. Therefore, several approaches for distributed training based on data parallelism (e.g., Horovod) and model/pipeline parallelism (e.g., GPipe, PipeDream) have emerged. In this work, we focus on an in-depth comparison of three different parallelism models that address these needs: data, model and pipeline parallelism. To this end, we provide an analytical comparison of the three, both in terms of computation time and memory usage, and introduce DDLBench, a comprehensive (open-source1, ready-to-use) benchmark suite to quantify these differences in practice. Through in-depth performance analysis and experimentation with various models, datasets, distribution models and hardware systems, we demonstrate that DDLBench can accurately quantify the capability of a given system to perform distributed deep learning (DDL). By comparing our analytical models with the benchmarking results, we show how the performance of real-life implementations diverges from these analytical models, thus requiring benchmarking to capture the in-depth complexity of the frameworks themselves.1https://github.com/sara-nl/DDLBench
{"title":"DDLBench: Towards a Scalable Benchmarking Infrastructure for Distributed Deep Learning","authors":"Matthijs Jansen, V. Codreanu, A. Varbanescu","doi":"10.1109/DLS51937.2020.00009","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00009","url":null,"abstract":"Due to its many applications across various fields of research, engineering, and daily life, deep learning has seen a surge in popularity. Therefore, larger and more expressive models have been proposed, with examples like Turing-NLG using as many as 17 billion parameters. Training these very large models becomes increasingly difficult due to the high computational costs and large memory footprint. Therefore, several approaches for distributed training based on data parallelism (e.g., Horovod) and model/pipeline parallelism (e.g., GPipe, PipeDream) have emerged. In this work, we focus on an in-depth comparison of three different parallelism models that address these needs: data, model and pipeline parallelism. To this end, we provide an analytical comparison of the three, both in terms of computation time and memory usage, and introduce DDLBench, a comprehensive (open-source1, ready-to-use) benchmark suite to quantify these differences in practice. Through in-depth performance analysis and experimentation with various models, datasets, distribution models and hardware systems, we demonstrate that DDLBench can accurately quantify the capability of a given system to perform distributed deep learning (DDL). By comparing our analytical models with the benchmarking results, we show how the performance of real-life implementations diverges from these analytical models, thus requiring benchmarking to capture the in-depth complexity of the frameworks themselves.1https://github.com/sara-nl/DDLBench","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132634827","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-11-01DOI: 10.1109/dls51937.2020.00002
{"title":"[Copyright notice]","authors":"","doi":"10.1109/dls51937.2020.00002","DOIUrl":"https://doi.org/10.1109/dls51937.2020.00002","url":null,"abstract":"","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"52 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133256750","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-10-22DOI: 10.1109/DLS51937.2020.00012
M. Cai, Jeroen B'edorf, V. Saletore, V. Codreanu, Damian Podareanu, Adel Chaibi, P. X. Qian
Galaxy mergers, the dynamical process during which two galaxies collide, are among the most spectacular phenomena in the Universe. During this process, the two colliding galaxies are tidally disrupted, producing significant visual features that evolve as a function of time. These visual features contain valuable clues for deducing the physical properties of the galaxy mergers. In this work, we propose DeepGalaxy, a visual analysis framework trained to predict the physical properties of galaxy mergers based on their morphology. Based on an encoder-decoder architecture, DeepGalaxy encodes the input images to a compressed latent space z, and determines the similarity of images according to the latent-space distance. DeepGalaxy consists of a fully convolutional autoencoder (FCAE) which generates activation maps at its 3D latent-space, and a variational autoencoder (VAE) which compresses the activation maps into a 1D vector, and a classifier that generates labels from the activation maps. The backbone of the FCAE can be fully customized according to the complexity of the images. DeepGalaxy demonstrates excellent scaling performance on parallel machines. On the Endeavour supercomputer, the scaling efficiency exceeds 0.93 when trained on 128 workers, and it maintains above 0.73 when trained with 512 workers. Without having to carry out expensive numerical simulations, DeepGalaxy makes inferences of the physical properties of galaxy mergers directly from images, and thereby achieves a speedup factor of ~105.
{"title":"DeepGalaxy: Deducing the Properties of Galaxy Mergers from Images Using Deep Neural Networks","authors":"M. Cai, Jeroen B'edorf, V. Saletore, V. Codreanu, Damian Podareanu, Adel Chaibi, P. X. Qian","doi":"10.1109/DLS51937.2020.00012","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00012","url":null,"abstract":"Galaxy mergers, the dynamical process during which two galaxies collide, are among the most spectacular phenomena in the Universe. During this process, the two colliding galaxies are tidally disrupted, producing significant visual features that evolve as a function of time. These visual features contain valuable clues for deducing the physical properties of the galaxy mergers. In this work, we propose DeepGalaxy, a visual analysis framework trained to predict the physical properties of galaxy mergers based on their morphology. Based on an encoder-decoder architecture, DeepGalaxy encodes the input images to a compressed latent space z, and determines the similarity of images according to the latent-space distance. DeepGalaxy consists of a fully convolutional autoencoder (FCAE) which generates activation maps at its 3D latent-space, and a variational autoencoder (VAE) which compresses the activation maps into a 1D vector, and a classifier that generates labels from the activation maps. The backbone of the FCAE can be fully customized according to the complexity of the images. DeepGalaxy demonstrates excellent scaling performance on parallel machines. On the Endeavour supercomputer, the scaling efficiency exceeds 0.93 when trained on 128 workers, and it maintains above 0.73 when trained with 512 workers. Without having to carry out expensive numerical simulations, DeepGalaxy makes inferences of the physical properties of galaxy mergers directly from images, and thereby achieves a speedup factor of ~105.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133451533","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-10-06DOI: 10.1109/DLS51937.2020.00008
Bita Hasheminezhad, S. Shirzad, Nanmiao Wu, Patrick Diehl, Hannes Schulz, Hartmut Kaiser
Although recent scaling up approaches to train deep neural networks have proven to be effective, the computational intensity of large and complex models, as well as the availability of large-scale datasets require deep learning frameworks to utilize scaling out techniques. Parallelization approaches and distribution requirements are not considered in the primary designs of most available distributed deep learning frameworks and most of them still are not able to perform effective and efficient fine-grained inter-node communication. We present Phylanx that has the potential to alleviate these shortcomings. Phylanx presents a productivity-oriented frontend where user Python code is translated to a futurized execution tree that can be executed efficiently on multiple nodes using the C++ standard library for parallelism and concurrency (HPX), leveraging fine-grained threading and an active messaging task-based runtime system.
{"title":"Towards a Scalable and Distributed Infrastructure for Deep Learning Applications","authors":"Bita Hasheminezhad, S. Shirzad, Nanmiao Wu, Patrick Diehl, Hannes Schulz, Hartmut Kaiser","doi":"10.1109/DLS51937.2020.00008","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00008","url":null,"abstract":"Although recent scaling up approaches to train deep neural networks have proven to be effective, the computational intensity of large and complex models, as well as the availability of large-scale datasets require deep learning frameworks to utilize scaling out techniques. Parallelization approaches and distribution requirements are not considered in the primary designs of most available distributed deep learning frameworks and most of them still are not able to perform effective and efficient fine-grained inter-node communication. We present Phylanx that has the potential to alleviate these shortcomings. Phylanx presents a productivity-oriented frontend where user Python code is translated to a futurized execution tree that can be executed efficiently on multiple nodes using the C++ standard library for parallelism and concurrency (HPX), leveraging fine-grained threading and an active messaging task-based runtime system.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126725342","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2020-09-09DOI: 10.1109/DLS51937.2020.00007
Yunsong Wang, Charlene Yang, S. Farrell, Yan Zhang, T. Kurth, Samuel Williams
Deep learning applications based on neural networks are generating considerable interest in various fields due to their high accuracy. Such an application is usually very compute-intensive thus requires a long run time. Researchers and engineers are actively exploring new solutions to this issue from both hardware and software/algorithm sides. However, little previous work has focused on providing a practical methodology to characterize deep learning performance bottlenecks and potentially guide the following optimization efforts. In this paper, we introduce an extension of the Roofline model and use it to analyze two representative computation kernels in deep learning, 2D convolution and long short-term memory, on NVIDIA GPUs. This new time-based Roofline model incorporates both compute/bandwidth complexity and run time in its formulae to demonstrate performance issues that cannot be reflected by the classic Roofline. Factors such as arithmetic intensity, data transfer, kernel launch overhead, and the Tensor Core usage will be examined by varying different parameters such as batch size and feature size, etc. This work helped form a more systematic way to understand the performance issue of deep learning applications. Last but not least, this generic performance model can be applied to a wide category of applications besides deep learning as well.
{"title":"Time-Based Roofline for Deep Learning Performance Analysis","authors":"Yunsong Wang, Charlene Yang, S. Farrell, Yan Zhang, T. Kurth, Samuel Williams","doi":"10.1109/DLS51937.2020.00007","DOIUrl":"https://doi.org/10.1109/DLS51937.2020.00007","url":null,"abstract":"Deep learning applications based on neural networks are generating considerable interest in various fields due to their high accuracy. Such an application is usually very compute-intensive thus requires a long run time. Researchers and engineers are actively exploring new solutions to this issue from both hardware and software/algorithm sides. However, little previous work has focused on providing a practical methodology to characterize deep learning performance bottlenecks and potentially guide the following optimization efforts. In this paper, we introduce an extension of the Roofline model and use it to analyze two representative computation kernels in deep learning, 2D convolution and long short-term memory, on NVIDIA GPUs. This new time-based Roofline model incorporates both compute/bandwidth complexity and run time in its formulae to demonstrate performance issues that cannot be reflected by the classic Roofline. Factors such as arithmetic intensity, data transfer, kernel launch overhead, and the Tensor Core usage will be examined by varying different parameters such as batch size and feature size, etc. This work helped form a more systematic way to understand the performance issue of deep learning applications. Last but not least, this generic performance model can be applied to a wide category of applications besides deep learning as well.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"290 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117311979","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
science. Other papers describe algorithms and systems for large-scale training on supercomputers, and approaches to performance benchmarking on the most powerful HPC systems. We thank both the authors and reviewers for their contributions to the workshop.
{"title":"Message from the Workshop Chairs","authors":"E. Biersack, P. Rodriguez","doi":"10.1109/PERCOMW.2006.85","DOIUrl":"https://doi.org/10.1109/PERCOMW.2006.85","url":null,"abstract":"science. Other papers describe algorithms and systems for large-scale training on supercomputers, and approaches to performance benchmarking on the most powerful HPC systems. We thank both the authors and reviewers for their contributions to the workshop.","PeriodicalId":185533,"journal":{"name":"2020 IEEE/ACM Fourth Workshop on Deep Learning on Supercomputers (DLS)","volume":"14 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130481232","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：