华为云深度学习在文本分类中的实践 华为 Cloud&AI 李明磊 3 2 3 1 4 分类 算法 简史 深度 学习 架构 难点 应用 案例 目录 4 文本分类介绍 内容：  买没几天就降价一点都不开心，闪存跑分就五百多点点 ---  外观漂亮音质不错，现在电子产品基本上都是华为的了 ---  汽车不错，省油，性价比高 ---  这个政策好啊，利国利民 --- 数据不均衡 预处理方法  上采样  下采样  SMOTE  数据增广 集成方法  SMOTEbagging 改损失函数  Focal loss “An Insight into Classification with Imbalanced Data: Empirical Results and Current Trends on Using Data Intrinsic

efficient too. We hope that this project gave you an insight into how to use distillation for your tasks. The next section gives a quick insight into some of the research into distillation related methods

of the loss function is changing with respect to and . The second derivative gives us a clearer insight into how important might be to minimize the loss. Since computing pairwise second-derivatives for

Preparing to use NVIDIA Containers Getting Started Guide. ‣ For non-DGX users, see NVIDIA ® GPU Cloud ™ (NGC) container registry installation documentation based on your platform. ‣ Ensure that you system depends on the DGX OS version that you installed (for DGX systems), the NGC Cloud Image that was provided by a Cloud Service Provider, or the software that you installed to prepare to run NGC containers shipped in /workspace/nvidia-examples). You can obtain the models from Github or the NVIDIA GPU Cloud (NGC) instead. Some Python packages that were included in previous containers to support these example

The original paper reports BERT-Base requiring 4 Cloud TPU Pods (4 chips each, total 16 chips) over 4 days for a total of 1,536 TPU hours. Each Cloud TPU chip is priced at $3.22 / hr6, which means the the training would take ~ 1536 * 3.22 = $4,945.92. BERT-Large requires 16 Cloud TPU Pods for 4 days, which turns out to be 6,144 TPU hours and $19,783.68 to train. Other pre-trained models can be a couple 68_A-12/4 7 GPU pricing source: https://cloud.google.com/compute/gpus-pricing. Numbers reported from October 2022. 6 Cloud TPU pricing source: https://cloud.google.com/tpu/pricing. Numbers reported from

(Tensor Processing Units) Google Cloud TPU. https://cloud.google.com/tpu NVIDIA V100 TPU v2 TPU v3 Hardware Architecture NVIDIA Volta GPU Google Cloud TPU Google Cloud TPU Memory 16GB / 32GB 64GB 128GB DL: 112 TFLOPS 180 TFLOPS 420 TFLOPS 深度学习的硬件 27 • 提问：训练一个模型需要多大开销？ • 以训练 BERT-large 模型为例, 16 Cloud TPUs = 16 * 4.5 = 72 USD / hour One-day cost = 72 * 24 = 1,728 USD Four-day cost = 1,728 USD *

SACC2017 图像技术的三个核心难点>>小、快、准 小模型 线上速度快 预测准 Frequent remote upgrade CPU-constrained, real-time Cloud processing SACC2017 视觉感知模型 分割 Forward Block Forward Block deconvolution deconvolution convolution Frame Predictor 检测 RNN SACC2017 360小水滴摄像机：视觉大不同 小水滴·360智能摄像机 视觉大不同 你不在家时有它在 通过语音人工智能实现求救与留言功能 Cloud-API 每天调用1.5亿次!2000QPS! SACC2017 系统框架 n 根据业务需求，对图像人脸进行识别，将结果推送到业务端 n 基于深度学习的准确的人脸检测、特征抽取 n 人脸检测占用95%计算资源 SACC2017 图像技术的三个核心难点>>小、快、准 小模型 线上速度快 预测准 Frequent remote upgrade CPU-constrained, real-time Cloud processing SACC2017 图像技术的三个核心难点>>小、快、准 模型 数据 工程 模型缩减 结构演进 SACC2017 单尺度卷积核 多尺度卷积核 视觉感知的三个核心难点>>小、快、准

................................................................................... 6 Anaconda Cloud ................................................................................................. run the below command to quit the environment: deactivate Anaconda Cloud We believe that you have installed anaconda cloud on your machine. If anaconda is not installed, then visit the official

AI+智慧城市 2015-2017 l单机、简易分布式人脸检测、跟踪、比对平台 l处理数十路到数百路监控摄像头数据 l千万级别深度学习特征检索 l行业试水 2018-2019 l云原生Cloud-Native超大规模视图存储、处理、检 索 l处理数万到数十万路，城市范围级别监控、门禁摄 像头数据 l10-100 Billion级别深度学习特征检索 - PB以上级别数据库存储 -

相关技术正在进⼊推荐领域 问题1. 推荐链路的漏⽃ 是对资源的巨⼤浪费 问题2. 结果利⽤ 不充分，响应不 够快 [2021] MC2 -SF: Slow-Fast Learning for Mobile-Cloud Collaborative Recommendation 问题3. ⼏⼗个场 景，独⽴链路 总结 � 千亿级推荐模型应⽤ O1. 千亿级特征（TB级）的模型的在线/离线训练， 在线推理服务和持续上线