历史匿名回执 ·
对课程本身不做评价，只是给想知道课程内容的同学一个参考。前面的点评都太久远了……没办法作为参考，久远到什么程度呢，差不多属于考古层，挖出来也拼不出这课现在长啥样。 ### 本学期上课时间： 周三16:00-17:40{1-8周}，注意，是1-8周，第9周开始这课在课表上就原地蒸发，别到时候满校园找人。 ### 课程内容： Week 1 Course Introduction Week 2 Introduction to the 1st competition topic and related background content Week 3 Introduction to the 2nd competition topic and related background content Week 4 Introduction to the 3rd competition topic and related background content Week 5 Introduction to Federated Learning basics Week 6 FL project(individual) introduction Week 7 Federated Learning Related Techniques Week 8 Presentation investigating the frontiers of Federal Learning Week 11 Group project (AI competition) Report (Online) Week 12 Individual project DDL 前四周基本是题目介绍连连看，一周抛一个比赛；第五周开始画风一转，直接把你按进联邦学习的海洋里，别问，问就是basics。老师说，最好在前六周完成小组项目，后六周完成个人项目——这话翻译一下就是：建议你前半程冲刺、后半程再冲刺，全程都在冲刺。 ### 考核标准： Daily performance(说是出勤): 10% Group presentation(Presentation investigating the frontiers of Federal Learning): 20% Group project(AI competition 3选1): 35% Individual project(Simulation of Federated Learning Process): 35% Bonus(Correcting errors/Answering question in class/Raising good question in class): 5 points at most on final score，最多加5分，改错、答题、问好问题都能薅，属于课堂上开源节流式攒分。 ### 项目内容： *所有project要求使用latex撰写报告，代码上传github并在报告中附上链接。* latex写报告、github传代码、报告里还得贴链接，一套连招下来，代码没跑通，排版先跑通了。 + Group project 1. Ultra-wide Angle OCTA image classification and lesion region segmentation competition[https://drac22.grand-challenge.org/] *Task 1*: Lesion segmentation. Realize the segmentation of lesions in ultra-wide Angle OCTA images. There are three categories of lesions: intraretinal microvascular abnormalities (IRMA), nonperfusion areas (NPA) and neovascularization (NV). *Task 2*: Image quality assessment. The images are classified into three quality levels: poor level, good level, excellent level. *Task 3*: DR Grading. The ultra-wide Angle OCTA images are classified into three categories: non-DR, NPDR and PDR. NPDR: non-proliferatived diabetic retinophathy, PDR: proliferatived diabetic retinophathy. **以上Task 3选1** 2. Hypertensive Retinopathy Diagnostic Competition[https://codalab.lisn.upsaclay.fr/competitions/11877] *Task 1*: Detection of hypertensive retinopathy. This is a dichotomous task. Given a fundus image of a patient, the aim of the task is to determine whether this eye has hypertensive retinopathy. *Task 2*: Prediction of blood pressure. This is a regression task. Given the fundus image of a patient, the aim of the task is to predict the values of systolic blood pressure (SBP) and diastolic blood pressure (DBP). **以上Task 2选1** 3. Text Generation Sketch Challenge(TGSC)[https://codalab.lisn.upsaclay.fr/competitions/13111?secret_key=cca0f3a6-4a08-4638-8446-dd8118bffae5] The main task of this challenge is to input a piece of text containing limited facial features and output a high-quality sketch portrait corresponding to that text. 小组项目三个比赛摆在你面前，像三扇门，一扇是眼底图鉴，一扇是血压预言机，一扇是文字变脸，选哪个都得自己扛。 + Group presentation Investigate the frontiers of Federal Learning, present at Week 8. 第8周上台讲联邦学习前沿，前沿到底在哪不知道，反正你得先站上去。 + Individual project The task itself is a simple CIFAR10 classification task. With horizontal federation, each client owns a non-overlapping set of data. Local clients train their own models and upload them to the server to complete model aggregation. After that, the server sends the model to the client to continue training. We need to complete the simulation of both full customer participation in the model update and partial customer participation. The specific details of the task will be presented in the subsequent course. 个人项目说是simple的CIFAR10分类，配合横向联邦，每个client抱住一块不重叠的数据，本地练模型、传给server聚合、server再发回来接着练，核心就是让你把“全员参与更新”和“部分客户参与更新”两种模式都模拟一遍，细节后续课上再说——经典“先别急，后面讲”，后面讲没讲，反正课上会讲。 **以上内容均来源于 Week1 Course Introduciton 的PPT** 注意是Introduciton，PPT上就这么拼的，我也就这么照着往里放，主打一个原汁原味。