大會實作坊與訓練

The CUDA computing platform enables the acceleration of CPU-only applications to run on the world’s fastest massively parallel GPUs. Experience C/C++ application acceleration by:

- Accelerating CPU-only applications to run their latent parallelism on GPUs
- Utilizing essential CUDA memory management techniques to optimize accelerated applications
- Exposing accelerated application potential for concurrency and exploiting it with CUDA streams
- Leveraging Nsight Systems to guide and check your work

Upon completion, you’ll be able to accelerate and optimize existing C/C++ CPU-only applications using the most essential CUDA techniques and Nsight Systems. You’ll understand an iterative style of CUDA development that will allow you to ship accelerated applications fast.

RAPIDS is a collection of data science libraries that allows end-to-end GPU acceleration for data science workflows. In this training, you'll:

- Use cuDF and Dask to ingest and manipulate massive datasets directly on the GPU
- Apply a wide variety of GPU-accelerated machine learning algorithms, including XGBoost, cuGRAPH, and cuML, to perform data analysis at massive scale
- Perform multiple analysis tasks on massive datasets in an effort to stave off a simulated epidemic outbreak affecting the UK

Upon completion, you'll be able to load, manipulate, and analyze data orders of magnitude faster than before, enabling more iteration cycles and drastically improving productivity.

You’ll learn how to use Transformer-based natural language processing models for text classification tasks, such as categorizing documents. You will also learn how to leverage Transformer-based models for named-entity recognition (NER) tasks and how to analyze various model features, constraints, and characteristics to determine which model is best suited for a particular use case based on metrics, domain specificity, and available resources.

By participating in this workshop, you’ll be able to:

- Understand how text embeddings have rapidly evolved in NLP tasks such as Word2Vec, recurrent neural network (RNN)-based embeddings, and Transformers
- See how Transformer architecture features, especially self-attention, are used to create language models without RNNs
- Use self-supervision to improve the Transformer architecture in BERT, Megatron, and other variants for superior NLP results
- Leverage pre-trained, modern NLP models to solve multiple tasks such as text classification, NER, and question answering
- Manage inference challenges and deploy refined models for live applications

The amount of information moving through our world’s telecommunications infrastructure makes it one of the most complex and dynamic systems that humanity has ever built. In this workshop, you’ll implement multiple AI-based solutions to solve an important telecommunications problem: identifying network intrusions.

In this workshop, you’ll:

- Implement three different anomaly detection techniques: accelerated XGBoost, deep learning-based autoencoders, and generative adversarial networks (GANs)
- Build and compare supervised learning with unsupervised learning-based solutions
- Discuss other use cases within your industry that could benefit from modern computing approaches

Upon completion, you'll be able to detect anomalies within large datasets using supervised and unsupervised machine learning.

Modern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. In this course, you will learn how to scale deep learning training to multiple GPUs. Using multiple GPUs for deep learning can significantly shorten the time required to train lots of data, making solving complex problems with deep learning feasible.

This course will teach you how to use multiple GPUs to train neural networks. You'll learn:

- Approaches to multi-GPU training
- Algorithmic and engineering challenges to large-scale training
- Key techniques used to overcome the challenges mentioned above

Upon completion, you'll be able to effectively parallelize training of deep neural networks using Horovod.

Learn how to design, train, and deploy deep neural networks for autonomous vehicles using the DRIVE development platform.

You'll learn how to:

- Work with CUDA code, memory management, and GPU acceleration on the DRIVE AGX
- Train a semantic segmentation neural network
- Optimize, validate, and deploy a trained neural network using TensorRT

Upon completion, you'll be able to create and optimize perception components for autonomous vehicles using NVIDIA DRIVE.

Writing CUDA C++ applications that efficiently and correctly utilize all available GPUs on a node drastically improves performance over single-GPU code, and makes the most cost-effective use out of compute nodes with multiple GPUs. In this workshop you will learn to utilize multiple GPUs on a single node by:

- Learning how to launch kernels on multiple GPUs, each working on a subsection of the required work
- Learning how to use concurrent CUDA Streams to overlap memory copy with computation on multiple GPUs

Upon completion, you will be able to build robust and efficient CUDA C++ applications that can leverage all available GPUs on a single node.

Discover how deep learning works through hands-on exercises in computer vision and natural language processing. You’ll train deep learning models from scratch, learning tools and tricks to achieve highly accurate results. You’ll also learn to leverage freely available, state-of-the-art pre-trained models to save time and get your deep learning application up and running quickly.

By participating in this is workshop, you’ll:
> Learn the fundamental techniques and tools required to train a deep learning model
> Gain experience with common deep learning data types and model architectures
> Enhance datasets through data augmentation to improve model accuracy
> Leverage transfer learning between models to achieve efficient results with less data and computation
> Build confidence to take on your own project with a modern deep learning framework

Learn how to identify anomalies and failures in time-series data, estimate the remaining useful life of the corresponding parts, and use this information to map anomalies to failure conditions.

You'll learn how to:

> Leverage predictive maintenance to manage failures and avoid costly unplanned downtimes
> Identify key challenges around identifying anomalies that can lead to costly breakdowns
> Use time-series data to predict outcomes using machine learning classification models with XGBoost
> Apply predictive maintenance procedures by using a long short-term memory ( LSTM)-based model to predict device failure
> Experiment with autoencoders to detect anomalies by using the time-series sequences from the previous steps

Upon completion, you’ll understand how to use AI to predict the condition of equipment and estimate when maintenance should be performed.

The CUDA computing platform enables the acceleration of CPU-only applications to run on the world’s fastest massively parallel GPUs. Experience C/C++ application acceleration by:

> Accelerating CPU-only applications to run their latent parallelism on GPUs
> Utilizing essential CUDA memory management techniques to optimize accelerated applications
> Exposing accelerated application potential for concurrency and exploiting it with CUDA streams
> Leveraging Nsight Systems to guide and check your work

Upon completion, you’ll be able to accelerate and optimize existing C/C++ CPU-only applications using the most essential CUDA techniques and Nsight Systems. You’ll understand an iterative style of CUDA development that will allow you to ship accelerated applications fast.

You’ll learn how to use Transformer-based natural language processing models for text classification tasks, such as categorizing documents. You will also learn how to leverage Transformer-based models for named-entity recognition (NER) tasks and how to analyze various model features, constraints, and characteristics to determine which model is best suited for a particular use case based on metrics, domain specificity, and available resources.

By participating in this workshop, you’ll be able to:
> Understand how text embeddings have rapidly evolved in NLP tasks such as Word2Vec, recurrent neural network (RNN)-based embeddings, and Transformers
> See how Transformer architecture features, especially self-attention, are used to create language models without RNNs
> Use self-supervision to improve the Transformer architecture in BERT, Megatron, and other variants for superior NLP results
> Leverage pre-trained, modern NLP models to solve multiple tasks such as text classification, NER, and question answering
> Manage inference challenges and deploy refined models for live applications

Learn how GPUs change the game for Data Engineering on the hardware level and how to optimize pipelines with CUDA Python Libraries.

Practice GPU based Data Engineering by:

> Surveying different Data Formats and their strengths and weaknesses
> Using statistics to get a "feel" for the data and visualizing it in a dashboard
> Scaling an ETL pipeline to multiple GPUs using NVTabular

Upon completion, learners will be able to make appropriate hardware and software choices to optimize the speed of their data pipelines.

Writing CUDA C++ applications that efficiently and correctly utilize all available GPUs on a node drastically improves performance over single-GPU code, and makes the most cost-effective use out of compute nodes with multiple GPUs. In this workshop you will learn to utilize multiple GPUs on a single node by:

- Learning how to launch kernels on multiple GPUs, each working on a subsection of the required work
- Learning how to use concurrent CUDA Streams to overlap memory copy with computation on multiple GPUs

Upon completion, you will be able to build robust and efficient CUDA C++ applications that can leverage all available GPUs on a single node.

RAPIDS is a collection of data science libraries that allows end-to-end GPU acceleration for data science workflows. In this training, you'll:

> Use cuDF and Dask to ingest and manipulate massive datasets directly on the GPU
> Apply a wide variety of GPU-accelerated machine learning algorithms, including XGBoost, cuGRAPH, and cuML, to perform data analysis at massive scale
> Perform multiple analysis tasks on massive datasets in an effort to stave off a simulated epidemic outbreak affecting the UK

Upon completion, you'll be able to load, manipulate, and analyze data orders of magnitude faster than before, enabling more iteration cycles and drastically improving productivity.

深度學習技術的推薦系統，正是提供個人化線上體驗的秘方，也是零售、娛樂、醫療照護、金融與其他產業的強大決策支援工具。本次實作坊的內容，涵蓋了建立高效推薦系統所需的基本工具和技術，並會說明如何部署 GPU 加速的解決方案以即時提供推薦。

參加本實作坊可學會：

> 使用開放原始碼 cuDF 函式庫和 Apache Arrow 建立以內容為基礎的推薦系統
> 使用交替最小平方（alternating least squares，ALS）和 CuPy 建立協作式篩選推薦系統
> 使用 TensorFlow 2 設計出兼具寬度和深度的神經網路，打造混合式的推薦系統
> 使用大型的稀疏資料集，達到最佳的訓練和推論效能
> 將推薦模型部署為高效能網路服務

在本實作坊中，會學習如何將以 Transformer 為基礎的自然語言處理模型運用於文字分類工作，例如分類文件。也會學到如何利用以 Transformer 為基礎的模型來處理命名實體識別（named-entity recognition，NER）工作，以及如何分析各種模型功能、限制和特性，以便依據指標、領域獨特性和可用資源，判斷何種模型最適合特定使用案例。

參加本實作坊可學會：

> 瞭解文字內嵌如何在自然語言處理工作中迅速進化，例如 Word2Vec、遞歸神經網路（RNN）型內嵌及 Transformer
> 瞭解 Transformer 架構功能（尤其是自我注意力機制）是如何用於建立語言模型，而無需遞歸神經網路
> 使用自我監督技術提升 BERT、Megatron 和其他變體中的 Transformer 架構，以達到優異的自然語言處理成果
> 利用預先訓練的現代自然語言處理模型處理多項工作，例如文字分類、命名實體識別和問答系統
> 管理推論挑戰，並部署即時應用程式的微調模型

深度學習技術的推薦系統，正是提供個人化線上體驗的秘方，也是零售、娛樂、醫療照護、金融與其他產業的強大決策支援工具。本次實作坊的內容，涵蓋了建立高效推薦系統所需的基本工具和技術，並會說明如何部署 GPU 加速的解決方案以即時提供推薦。

參加本實作坊可學會：

> 使用開放原始碼 cuDF 函式庫和 Apache Arrow 建立以內容為基礎的推薦系統
> 使用交替最小平方（alternating least squares，ALS）和 CuPy 建立協作式篩選推薦系統
> 使用 TensorFlow 2 設計出兼具寬度和深度的神經網路，打造混合式的推薦系統
> 使用大型的稀疏資料集，達到最佳的訓練和推論效能
> 將推薦模型部署為高效能網路服務

With the increase in traffic cameras, growing prospect of autonomous vehicles, and promising outlook of smart cities, there's a rise in demand for faster and more efficient object detection and tracking models. This involves identification, tracking, segmentation and prediction of different types of objects within video frames.

In this workshop, you’ll learn how to:

- Efficiently process and prepare video feeds using hardware accelerated decoding methods
- Train and evaluate deep learning models and leverage ""transfer learning"" techniques to elevate efficiency and accuracy of these models and mitigate data sparsity issues
- Explore the strategies and trade-offs involved in developing high-quality neural network models to track moving objects in large-scale video datasets
- Optimize and deploy video analytics inference engines by acquiring the DeepStream SDK

Upon completion, you'll be able to design, train, test and deploy building blocks of a hardware-accelerated traffic management system based on parking lot camera feeds.

Learn how to identify anomalies and failures in time-series data, estimate the remaining useful life of the corresponding parts, and use this information to map anomalies to failure conditions.

You'll learn how to:

> Leverage predictive maintenance to manage failures and avoid costly unplanned downtimes
> Identify key challenges around identifying anomalies that can lead to costly breakdowns
> Use time-series data to predict outcomes using machine learning classification models with XGBoost
> Apply predictive maintenance procedures by using a long short-term memory ( LSTM)-based model to predict device failure
> Experiment with autoencoders to detect anomalies by using the time-series sequences from the previous steps

Upon completion, you’ll understand how to use AI to predict the condition of equipment and estimate when maintenance should be performed.

Modern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. In this course, you will learn how to scale deep learning training to multiple GPUs. Using multiple GPUs for deep learning can significantly shorten the time required to train lots of data, making solving complex problems with deep learning feasible.

This course will teach you how to use multiple GPUs to train neural networks. You'll learn:

- Approaches to multi-GPU training
- Algorithmic and engineering challenges to large-scale training
- Key techniques used to overcome the challenges mentioned above

Upon completion, you'll be able to effectively parallelize training of deep neural networks using Horovod.

You’ll learn how to use Transformer-based natural language processing models for text classification tasks, such as categorizing documents. You will also learn how to leverage Transformer-based models for named-entity recognition (NER) tasks and how to analyze various model features, constraints, and characteristics to determine which model is best suited for a particular use case based on metrics, domain specificity, and available resources.

By participating in the workshop you will:

> Understand how text embeddings have rapidly evolved in NLP tasks such as Word2Vec, recurrent neural network (RNN)-based embeddings, and Transformers
> See how Transformer architecture features, especially self-attention, are used to create language models without RNNs
> Use self-supervision to improve the Transformer architecture in BERT, Megatron, and other variants for superior NLP results 
> Leverage pre-trained, modern NLP models to solve multiple tasks such as text classification, NER, and question answering
> Manage inference challenges and deploy refined models for live applications

Modern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. In this course, you will learn how to scale deep learning training to multiple GPUs. Using multiple GPUs for deep learning can significantly shorten the time required to train lots of data, making solving complex problems with deep learning feasible.

This course will teach you how to use multiple GPUs to train neural networks. You'll learn:

> Approaches to multi-GPU training
> Algorithmic and engineering challenges to large-scale training
> Key techniques used to overcome the challenges mentioned above

Upon completion, you'll be able to effectively parallelize training of deep neural networks using Horovod.