当前位置: 首页 > 工具软件 > Forward DL > 使用案例 >

DL知识总结

水焱
2023-12-01

1、优化方法

https://blog.csdn.net/yinyu19950811/article/details/90476956---精读,讲的非常好

https://blog.csdn.net/u011984148/article/details/99800544--RADAM

http://wulc.me/2019/03/18/Adam%E9%82%A3%E4%B9%88%E6%A3%92%EF%BC%8C%E4%B8%BA%E4%BB%80%E4%B9%88%E8%BF%98%E5%AF%B9SGD%E5%BF%B5%E5%BF%B5%E4%B8%8D%E5%BF%98/

https://mp.weixin.qq.com/s?__biz=MzA3NDIyMjM1NA==&mid=2649031658&idx=1&sn=fd1b54b24b607a9d28dc4e83ecc480fb&chksm=8712bd97b065348132d8261907c56ce14077646dfc9c7531a4c3f1ecf6da1a488450428e4580&scene=21#wechat_redirect

https://www.cnblogs.com/wishchin/p/9199825.html

https://baijiahao.baidu.com/s?id=1643019716161086382&wfr=spider&for=pc--RAdam和LOOKHEAD结合

2、权重衰减、L2正则、

3、权重裁剪、梯度裁剪、学习率裁剪

4、数据增强(通用的数据增强(传统的、新提出的mixup等)、特点任务的数据增强、基于GAN的数据增强)

https://blog.csdn.net/hacker_long/article/details/89104317

5、ROC、froc、pr曲线、精确率、召回率、灵敏度、特异度、假阳率、准确率

https://blog.csdn.net/rocling/article/details/93602303

https://www.cnblogs.com/freebird92/p/9021405.html

https://blog.csdn.net/b876144622/article/details/80009867

https://www.thepaper.cn/newsDetail_forward_6848696---froc

https://www.lmlphp.com/user/56/article/item/1756/

6、偏差和方差

https://blog.csdn.net/qq_41951186/article/details/82534050

 

7、学习率和batch size

https://zhuanlan.zhihu.com/p/64864995

https://blog.csdn.net/fanhongyuan21/article/details/103100398

8、模型的宽度和深度影响

https://blog.csdn.net/hacker_long/article/details/89608750

https://blog.csdn.net/hacker_long/article/details/100138274

 

9、FPN(top-down等)、图像金字塔(多尺度输入、输出预测)、多层预测如SSD、正常卷积特征最后一层预测(如检测faster-rcnn的公共特征图)、NMS(softnms等改进)、OHEM、focal loss

https://blog.csdn.net/u012426298/article/details/81773319

https://blog.csdn.net/thisiszdy/article/details/89950030?depth_1-utm_source=distribute.pc_relevant.none-task-blog-OPENSEARCH-1&utm_source=distribute.pc_relevant.none-task-blog-OPENSEARCH-1

https://blog.csdn.net/qq_43243022/article/details/88895057---NMS及python代码

10、样本不平衡解决(样本端、loss端)

要分清是样本类别不均衡还是检测中的正负样本(前景、背景不均衡)

数据端:

离线+在线数据增强(imgaug)

loss端:

focal loss

https://zhuanlan.zhihu.com/p/49981234

ohem

GHM

circle loss--将softmax+交叉熵推广到多标签分类,缓解不平衡(一般的多标签分类采用sigmoid+交叉熵,然后用阈值卡,无法缓解不平衡),circle-loss是度量学习的loss思想

libra-rcnn:检测中的不平衡问题的解决(sample不平衡、low-high level的特征不平衡、分类、检测loss的不平衡)

lovrax-softmax:分割中用bce无法解决不同样本图片中label区域大小带来的loss计算的不平衡,有wight bce loss,直接用IOU loss训练会带来严重的震荡,loss曲线不平滑,因此有lovrax-softmax loss

11、attention(通道和图像层面)、hard/soft attention等

https://www.jianshu.com/p/366b8e9bfb99

https://zhuanlan.zhihu.com/p/65459972---SEnet(https://www.sohu.com/a/161633191_465975)

https://blog.csdn.net/qixutuo6087/article/details/88822428---senet、sknet

https://blog.csdn.net/elaine_bao/article/details/80821306----non-local(https://zhuanlan.zhihu.com/p/33345791)

https://zhuanlan.zhihu.com/p/33345791

https://blog.csdn.net/u011345885/article/details/96566339----GCNet

https://blog.csdn.net/weixin_42662358/article/details/90676272?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1---GCNet

https://blog.csdn.net/l2181265/article/details/86754225-----CCNet

https://blog.csdn.net/czp_374/article/details/87600944?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-2&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-2---CCNet

https://www.lizenghai.com/archives/45076.html---scse attention

https://blog.csdn.net/qq_32768091/article/details/86612132---BAM attention

https://blog.csdn.net/u013738531/article/details/82731257----CBAM attention

https://www.jianshu.com/p/06742d8bd8ff---Dual attention语义分割

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247487322&idx=1&sn=c2704e63c0a2555a6e30e2e3840c5e94&chksm=ec1fe2a3db686bb52cc7efb39bfc71140c6a52d31d4158558adce3799620f472e675041482de&token=1455375513&lang=zh_CN&scene=21#wechat_redirect----OCNet

https://blog.csdn.net/mieleizhi0522/article/details/84873101--OCNet

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247487328&idx=1&sn=48a3fac2350c0c0f08abf9b6b4f1b213&chksm=ec1fe299db686b8f3bf38af83bf2b4301fb9fb69ef51c9c79fb0a49e6401e36f50ff38ce8c3a&token=938535871&lang=zh_CN&scene=21#wechat_redirect-----Fast-OCNet

https://zhuanlan.zhihu.com/p/53466363-------RCAN:超分辨率中的注意力机制

12、dicom、nii、nrrd等医学图像格式

https://www.jianshu.com/p/9bf31ac4f184--dicom信息

Pydicom+SimpleITK操作DICOM图像数据和TAG

pydicom和SimpleITK分别解析医学影像中dicom文件

软件:ITK-SNAP(https://blog.csdn.net/qq_33254870/article/details/100125788)

常见三种医学图像格式(DICOM、Nrrd、Nifti)

https://www.cnblogs.com/duter/p/4974755.html

Python可视化mhd格式和raw格式的医学图像并保存

https://www.cnblogs.com/XDU-Lakers/p/10781321.html(dcm转mhd、raw)

http://blog.sina.com.cn/s/blog_4bce5f4b0100rl7t.html

https://www.manongdao.com/article-1134706.html

https://blog.csdn.net/caixiajia/article/details/53749495

https://blog.csdn.net/shaguawo/article/details/80609810?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2-----[Python][DICOM]去除一个Series下重复Instance Number的DICOM图像文件

https://blog.csdn.net/qq_26293147/article/details/68924393

https://zhidao.baidu.com/question/745943474307776412.html

3d-slicer、MITK

13、unet、resunet、attention-unet、3d-unet、2.5d、vnet、nnunet等

https://blog.csdn.net/zhang_jing5/article/details/90452924

14、faster-rcnn系列、mask-rcnn、yolo系列、deeplab系列

15、roi-pooling、roi-align、prroi-pooling等

16、upsamplining、uppooling、反卷积、pixel-shuffle等

17、sigmoid、softmax、relu、prelu、maxout等

18、1D/2D/3D CNN、十大拍案叫绝的卷积设计(组卷积、深度可分离卷积)等

https://blog.csdn.net/u010801994/article/details/85005979------卷积原理:几种常用的卷积(标准卷积、深度卷积、组卷积、扩展卷积、反卷积)

https://blog.csdn.net/isMarvellous/article/details/80087705-----关于转置卷积(反卷积)的理解---https://blog.csdn.net/tsyccnh/article/details/87357447

https://blog.csdn.net/zhulf0804/article/details/86676279?utm_medium=distribute.pc_relevant.none-task-blog-baidujs-3----3D卷积和去(反)卷积----https://zhuanlan.zhihu.com/p/31841353----https://blog.csdn.net/liuweiyuxiang/article/details/84202352----https://zhuanlan.zhihu.com/p/47490523

https://blog.csdn.net/m0_37833297/article/details/89214342?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase-----2D、3D卷积后特征图大小公式

 

 

 

19、inception系列、resnet系列、inception-resnet系列

https://blog.csdn.net/u010712012/article/details/84862797

20、dropout、dropblock、常见的normalization(BN\LN)等

https://blog.csdn.net/qq_26598445/article/details/81950116

21、标注工具labelme、labelimg、CVAT等,以及遥感、文本中的一些旋转标注工具

22、检测专题直到anchor-free、分割专题(语义分割、实例分割、全景分割)、分类(单分类、多分类、多标签分类、人脸属性识别)等

分类

https://zhuanlan.zhihu.com/p/22214112-------CNN--结构上的思考

论文笔记:CNN经典结构1(AlexNet,ZFNet,OverFeat,VGG,GoogleNet,ResNet)   ---写的非常好

论文笔记:CNN经典结构2(WideResNet,FractalNet,DenseNet,ResNeXt,DPN,SENet)-----精读

https://blog.csdn.net/u011501388/article/details/79262115---DenseNets

https://blog.csdn.net/winycg/article/details/86760274------Pytorch实现ResNet V2-Pre-activation ResNet

https://blog.csdn.net/zkq_1986/article/details/84142047------【深度学习】alexnet、vgg19_bn、ResNet-110、PreResNet-110、ResNeXt-29, 8x64等模型性能对比

https://www.cnblogs.com/huanxifan/p/12626863.html----Wide Resnet(Resnet的有趣变种:WRN--https://blog.csdn.net/shwan_ma/article/details/78168629

http://www.mamicode.com/info-detail-2293789.html---总结近期CNN模型的发展(一)---- ResNet [1, 2] Wide ResNet [3] ResNeXt [4] DenseNet [5] DPNet [9] NASNet [10] SENet [11] Capsules [12]

https://blog.csdn.net/ruoruojiaojiao/article/details/89074763----Res2Net------论文理解

https://blog.csdn.net/hejin_some/article/details/80743818-----ResNeXt算法详解(resnet提升篇)(组卷积)

https://blog.csdn.net/qq_39478403/article/details/105567088?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-1.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-1.nonecase------【论文阅读】Improved Residual Networks for Image and Video Recognition (iResNet)----https://github.com/iduta/iresnet(分布式训练)

https://www.jianshu.com/p/46d76bd56766------干货 | 深度详解ResNet及其六大变体

Resnest

 

Xception、Mobienet是google利用深度可分离卷积提出的分类网络

https://blog.csdn.net/lk3030/article/details/84847879

https://blog.csdn.net/u014380165/article/details/75142710?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1.nonecase

https://www.jianshu.com/p/4708a09c4352

 

 

 

检测

https://blog.csdn.net/baidu_30594023/article/details/82623623---FPN

https://github.com/AlexeyAB/darknet----yolov4

https://www.cnblogs.com/silence-cho/p/11717135.html----yolov1-v3精读

https://zhuanlan.zhihu.com/p/59910080------RetinaNet(Focal Loss)

https://baijiahao.baidu.com/s?id=1651514751262272881&wfr=spider&for=pc-------比当前SOTA小4倍、计算量少9倍,谷歌最新目标检测器EfficientDet

https://blog.csdn.net/czp_374/article/details/81162923-----人脸检测和识别

https://blog.csdn.net/weixin_43489950/article/details/94720958---AP loss(https://blog.csdn.net/jiaoyangwm/article/details/91479594

Anchor free

http://www.dataguru.cn/article-14837-1.html--------重磅!13篇基于Anchor free的目标检测方法

https://zhuanlan.zhihu.com/p/93703085------一文看尽8篇目标检测最新论文(EfficientDet/EdgeNet/ASFF/RoIMix等)

 

 

小目标检测(通用的小目标检测、人脸小目标检测):

https://blog.csdn.net/mary_0830/article/details/103100001?utm_medium=distribute.pc_relevant.none-task-blog-baidujs-7-----ICCV2019 | 目标检测论文阅读 SCRDet:Towards More Robust Detection for Small, Cluttered and Rotated Objects

https://www.jianshu.com/p/7e7f677bc8eb----Efficient Coarse-to-Fine Non-Local Module for the Detection of Small Objects

tinyperson

 

实时目标检测

https://blog.csdn.net/mingqi1996/article/details/91873719-----Receptive Field Block Net for Accurate and Fast Object Detection论文笔记

 

 

分割:

https://github.com/ShawnBIT/Loss-family---分割的各种loss

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486129&idx=1&sn=67bdf950e107ae2508fb9763450c4871&chksm=ec1fe748db686e5e6a2aa5af4ca39c85a5e6e8d9cf6386bd62fd4f46a117da49a5133868d2e2&scene=21#wechat_redirect------CVPR 2018|分割算法——可以分割一切目标(附各种分割总结)

https://www.sohu.com/a/314483400_823210---精度高、模型小、速度快!梯形DenseNets结构实现语义分割新高度!

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486129&idx=1&sn=67bdf950e107ae2508fb9763450c4871&chksm=ec1fe748db686e5e6a2aa5af4ca39c85a5e6e8d9cf6386bd62fd4f46a117da49a5133868d2e2&scene=21#wechat_redirect

https://blog.csdn.net/digu6003/article/details/101283920?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1---综述未看

https://blog.csdn.net/Najlepszy/article/details/104212394?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-2&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-2------图像分割论文“Adaptive Context Network for Scene Parsing”

https://blog.csdn.net/guleileo/article/details/80544835?depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-7&utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-7---------北大、北理工、旷视联手:用于图像语义分割的金字塔注意力网络(PANet)

https://blog.csdn.net/mieleizhi0522/article/details/83651195-------MS-NFN Model for Retinal Vessel Segmentation(血管分割)

https://blog.csdn.net/Formlsl/article/details/80373200---Unet

https://www.jianshu.com/p/a782aba60ede-----Unet---写的很细,包括Over-tile策略、边缘像素加权

大规模数据目标检测

https://blog.csdn.net/c9Yv2cf9I06K2A9E/article/details/105648318

concurrent-softmax loss---大规模多标签目标检测--CVPR2020 Oral

 

 

 

 

实时语义分割:

https://blog.csdn.net/Arron_hou/article/details/101073177?depth_1-utm_source=distribute.pc_relevant.none-task-blog-OPENSEARCH-2&utm_source=distribute.pc_relevant.none-task-blog-OPENSEARCH-2-----未看

https://blog.csdn.net/Freeverc/article/details/83591210---Bisenet

https://zhuanlan.zhihu.com/p/129803377---Bisenetv2

YOLACT

https://blog.csdn.net/sinat_37532065/article/details/89415374

YOLCAT++(实时实例分割)、SOLO、SOLOv2

多标签分类:

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486947&idx=2&sn=159c8a4b76c0993fe38389f7bd7f1b5d&chksm=ec1fe01adb68690c22b65a8cc4c6631a9d8c96fb526643c42ef79552b018329c474006f6241a&token=1382801619&lang=zh_CN&scene=21#wechat_redirect---腾讯AI lab开源的多标签分类数据集--未看

OCR

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486894&idx=1&sn=8cc95793f82dd54d1766c87c96c6adb0&chksm=ec1fe057db686941ccf784d866b8d0624461f625d7a10cdf6dde1ad364639372f9cd900b988a&token=1821394195&lang=zh_CN&scene=21#wechat_redirect---未看

Reid

https://blog.csdn.net/qq_28266311/article/details/85856772?depth_1-utm_source=distribute.pc_relevant.none-task&utm_source=distribute.pc_relevant.none-task

https://blog.csdn.net/m0_37615398/article/details/98040663

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486867&idx=1&sn=f7ff524ac53fa12572308e13e73cc6c4&chksm=ec1fe06adb68697c49da589f0431572fd77e8689cbd318b560868f44c18a41e3b1b736c15495&scene=21#wechat_redirect--未看

23、GAN系列,原理loss上的改进、特定任务的结构(如风格迁移、样本生成、用GAN做异常检测)

24、OCR系列(检测、识别、end-to-end或者按照水平、倾斜、弯曲等划分)

25、人群计数、用分割代替检测用于数钢筋等

26、NAS(NAS-FPN、Auto-deeplab)、AutoML自动调参等

27、模型量化、压缩、知识蒸馏(teacher-student\mean-teacher)等、mimic

28、tensorRT、ONNX、tf-lite、tensorRT-inference-server、tf-serving等部署相关

https://blog.csdn.net/lidawei0124/article/details/90116124-----跑通Jetson Nano TensorRt sampleSSD例程

 

29、netro等网络结构可视化、模型转换工具、t-sne、large-vis数据可视化

30、全监督、半监督、弱监督、无监督、自监督(用于训练得到预训练模型)和各个领域的结合(如MAML),例如无监督领域自适应也可以看成一定的半监督问题(DA根据数据范围不同也可以分类、openset、closed-set的DA等)、多示例学习

31、深度强化学习系列

32、图像复原(去雾、去雨、去模糊、去噪等)、行人重识别reid、单目标、多目标跟踪、人脸检测、人脸识别、图像、视频检索、行为识别、视频关键帧、图像视频描述、关键点、姿态估计、超分辨率、显著性检测

姿态估计

https://blog.csdn.net/qq_14845119/article/details/104701103?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1----

高IOU遮挡下的姿态估计(HintPose)

https://blog.csdn.net/lh641446825/article/details/103110837?utm_medium=distribute.pc_relevant.none-task-blog-baidujs-6-----[ICCV 2019] 解决位姿估计中遮挡、对称、无纹理物体的方法

 

目标跟踪

MDNet

https://blog.csdn.net/elaine_bao/article/details/78824357-----视频物体检测(VID)系列 NoScope:1000x的视频检索加速算法

医学图像异常检测

https://blog.csdn.net/inkky/article/details/88943533---Skip-GANomaly

 

33、医学图像处理(2维X光、3维CT、MR、病理切片、OCT、糖网眼底彩照)、遥感图像处理是否可借鉴病理切片、工业检测

34、CAM、Grad-CAM热图绘制

35、训练时单机单卡、单机多卡、多级多卡分布式、多线程处理、训练加速、混合精度优化

https://baijiahao.baidu.com/s?id=1610222052714852444&wfr=spider&for=pc

https://blog.csdn.net/xiaojiajia007/article/details/103472850/---一个 Pytorch 训练实践 (分布式训练 + 半精度/混合精度训练)

https://blog.csdn.net/m0_37192554/article/details/106078207?utm_source=blogxgwz7----pytorch半精度,混合精度,单精度训练的区别amp.initialize

https://www.sohu.com/a/197649510_465975----学界 | 减少模型半数内存用量:百度&英伟达提出混合精度训练法 

http://nvidia.zhidx.com/content-6-1651-1.html-----单精度、双精度、多精度和混合精度计算的区别是什么?

https://blog.csdn.net/zyy617532750/article/details/104219708----GPU加速实战——混合精度训练

https://www.sohu.com/a/292425196_114877------实战 PK!RTX2080Ti 对比 GTX1080Ti 的 CIFAR100 混合精度训练 

http://www.gpus.cn/gpus_list_page_techno_support_content?id=61-----利用Tensor Core优化GPU性能的几个小窍门---https://www.cnblogs.com/marsggbo/p/11686184.html

https://baike.baidu.com/item/%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6%E8%AE%AD%E7%BB%83/22717597?fr=aladdin

 

36、padding等模式、图像卷积后尺寸计算

37、梯度爆炸、梯度消失等

38、NLP相关知识、知识图谱

https://zhuanlan.zhihu.com/p/107917373-----从transformer到bert

 

39、不确定度解决领域自适应(MICCAI眼科文章)、主动学习等

40、小目标检测专题、图像中多尺度物体检测

41、课程学习、自步学习、在线学习、持续学习等

42、三大特性含义:旋转不变性、平移不变性等

43、比赛方案分析总结

44、感受野、参数量、flops等计算

45、无人驾驶、车道线检测、点云等、轻量级的网络设计(mobilnetv1-v3、shufflenet、EfficientNet、FBNetv2,PVANet等)

46、latex论文写作使用

47、docker、nvidia-docker使用、git使用

48、不同特定任务各种loss、比如分割的dice、bce、soft-dice等

49、jupyter notebook、pycharm、markdown使用

50、各种任务评价指标的计算

51、GNN、GCN、胶囊网络等、强化学习药物研发

52、element-add、element-相乘、concat用处区别(好像add用的多,是梯度反传时比较好吗?)

53、BN和激活函数的顺序

54、模型的特征提取能力、模型的特征上下文信息、语义信息、模型的泛化能力、感受野、特征的reduction程度、通道和空间层面的特征表示含义、模型的计算量GFLOPS和参数量#P(卷积和全连接层不一样)、显存占用、flops、模型的吞吐量

https://zhuanlan.zhihu.com/p/86587652---参数量、FLOPs计算https://blog.csdn.net/Seven_year_Promise/article/details/69360488------resnet50参数量计算(https://blog.csdn.net/qq_42278791/article/details/90690747https://blog.csdn.net/qq_36401512/article/details/105640808)

http://www.elecfans.com/d/598165.html----五种CNN模型的尺寸,计算量和参数数量对比详解

https://www.jianshu.com/p/883df8e81450-----史上最全的cnn参数计算详解

 

 

55、zero-shot、one-shot、few-shot、小样本学习、元学习

56、5x5卷积可以由几个3x3卷积替代(作用:降低参数量等)、或者由3x3卷积,空洞率为2的空洞卷积替代、1X1卷积的用处等

57、孪生网络、实时检测、分割等网络

58、卷积神经网络的两个特点,增大感受野的方式(卷积层的堆叠)、训练网络的tricks(亚马逊的分类、检测tricks论文)

分割技巧:

https://mp.weixin.qq.com/s?__biz=MzI5MDUyMDIxNA==&mid=2247486978&idx=1&sn=5a7da98eb3a69a3260a4f128bd88770b&chksm=ec1fe3fbdb686aed3c74c0bcf611ddb4c91c20a5203fc0f2f67f14175af8348403c555ba12ee&token=1382801619&lang=zh_CN&scene=21#wechat_redirect

https://www.aiuai.cn/aifarm826.html

https://www.zhihu.com/question/272988870/answer/562262315

 

59、各种网络模型初始化策略(正太分布、kaiming等初始化)、模型预训练模型加载、冻结指定层训练等、warm up等技巧

60、传统图像特征

https://www.jianshu.com/p/2a06c68f6c14--边缘检测(sobel、canny、laplacian)

http://www.luyixian.cn/news_show_24847.aspx

https://www.cnblogs.com/henuliulei/p/10797634.html---图像的Garbor和hog特征

https://blog.csdn.net/lien0906/article/details/39023271?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromBaidu-1

 

61、项目开发中的版本控制

https://learngitbranching.js.org/?locale=zh_CN---git命令图形化学习

https://hacker-tools.github.io/version-control/

https://www.jianshu.com/p/ea6341224e89--------Git中.gitignore的配置语法

 

 

 

62、一些新颖的产品背后的技术

AI换脸

https://www.zhihu.com/question/316591736/answer/778920100

https://blog.csdn.net/qq_36625422/article/details/90053241---DeepFack实现

https://www.boxuegu.com/news/2156.html

https://www.zhihu.com/question/308193404

 

63、模型轻量化改进思路

https://zhuanlan.zhihu.com/p/33345791---5X5--->3X3(https://www.zhihu.com/question/265791259)

 

64、生产环境、测试环境、开发环境

https://blog.csdn.net/qq_30715329/article/details/79363691?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase

https://blog.csdn.net/yanglangdan/article/details/102552388

 

 

65、imgcat终端查看图片、jupyter notebook

https://blog.csdn.net/SoftPoeter/article/details/88954089?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.nonecase

https://blog.csdn.net/Li_suhuan/article/details/100281263

https://www.jianshu.com/p/91365f343585------Jupyter Notebook介绍、安装及使用教程

 

66、宽度学习

https://blog.csdn.net/weixin_40802676/article/details/100161494

 类似资料: