Skip to content

Commit

Permalink
README
Browse files Browse the repository at this point in the history
  • Loading branch information
f4396 committed Jun 9, 2022
2 parents 15ad53d + a619f57 commit 66c0c00
Show file tree
Hide file tree
Showing 2 changed files with 85 additions and 40 deletions.
112 changes: 73 additions & 39 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -181,64 +181,63 @@

# 🍭训练篇&部署篇

## 0⃣️ 常见报错

- 🐾 [opencv读取数据--segmentation fault](train/preprocess/sefgmentationfault.md)

## 1⃣️ 检查数据

## 2⃣️训练模型--train e ↓

### 🍬数据预处理

#### 🍃 数据集介绍
- 🐾 [VOC](train/preprocess/voc.md)
- 🐾 [VOC](train/preprocess/voc.md)
- 🐾 [COCO](train/preprocess/coco.md)
- 🐾 [Cityscapaces](https://github.com/mcordts/cityscapesScripts)
- 🐾 [Cityscapaces](https://github.com/mcordts/cityscapesScripts)

#### 🍃 特征缩放[🐾](train/preprocess/featurescale.md)
#### 🍃 特征缩放

#### 🍃 Pytorch数据读取
- 🐾 [归一化和标准化](train/preprocess/featurescale.md)

#### 🍃 数据读取

- 🐾 [加速训练 —提高 GPU 利用率](train/preprocess/loaddata.md)
- 🐾 [加速训练—节约显存](train/preprocess/save_mem.md)
- 🐾 [加速训练—NN如何占用GPU显存以及如何节约显存](train/preprocess/save_mem.md)

<<<<<<< HEAD
#### 🍃 Python读取图片

- 🐾 [opencv读取数据--segmentation fault](train/preprocess/sefgmentationfault.md)
- 🐾 [直方图处理](train/preprocess/histogram.md)
- 🐾 [python读取图片的几种方式](train/preprocess/open_image.md)
=======
- 🐾 [使用不同的库读取图片](train/preprocess/open_image.md)
>>>>>>> a619f5763e328f7441ace9c037b9e528fb08afc3
### 🍬 网络模型

#### 🍃 机器学习

- 🐾 [最小二乘法](train/ml/least_square.md)
- 🐾 [kmeans](train/ml/kmeans.md)
- 🐾 [分类和回归的区别](https://www.cnblogs.com/anovana/p/8031724.html)

#### 🍃 NN

- 🐾 [NN发展史](train/nn/history.md)
- 🐾 [45分钟理解深度神经网络和深度学习-拟合角度](http://staff.ustc.edu.cn/~lgliu/Resources/DL/What_is_DeepLearning.html)
- 🐾 零基础入门深度学习- [感知器](https://www.zybuluo.com/hanbingtao/note/433855)-[线性单元和梯度下降](https://www.zybuluo.com/hanbingtao/note/448086)-[神经网络和反向传播算法](https://www.zybuluo.com/hanbingtao/note/476663)-[卷积神经网络](https://www.zybuluo.com/hanbingtao/note/485480)-[循环神经网络](https://zybuluo.com/hanbingtao/note/541458)-[长短时记忆网络(LSTM)](https://zybuluo.com/hanbingtao/note/581764)-[递归神经网络](https://zybuluo.com/hanbingtao/note/626300)
- 🐾 [NN发展史](train/nn/history.md)[45分钟理解深度神经网络和深度学习-拟合角度](http://staff.ustc.edu.cn/~lgliu/Resources/DL/What_is_DeepLearning.html)
- 🐾 零基础入门深度学习: [感知器](https://www.zybuluo.com/hanbingtao/note/433855)->[线性单元和梯度下降](https://www.zybuluo.com/hanbingtao/note/448086)->[神经网络和反向传播算法](https://www.zybuluo.com/hanbingtao/note/476663)->[卷积神经网络](https://www.zybuluo.com/hanbingtao/note/485480)->[循环神经网络](https://zybuluo.com/hanbingtao/note/541458)->[长短时记忆网络(LSTM)](https://zybuluo.com/hanbingtao/note/581764)->[递归神经网络](https://zybuluo.com/hanbingtao/note/626300)

#### 🍃CNN

- 🐾 [CNN介绍](train/cnn/introduce_cnn.md) --[🐾CNN网络代码讲解加部署 ⭐](https://github.com/WZMIAOMIAO/deep-learning-for-image-processing)
- 🐾 [LeNet, AlexNet, VGG](train/cnn/lenet_alexnet_vgg.md)
- 🐾 [DSC理解](train/cnn/dsc.md)
- 🐾 [Inception](train/cnn/inception.md)
- 🐾 [ResNet](train/cnn/resnet.md)
- 🐾 [DenseNet](train/cnn/densenet.md)
- 🐾 [MobileNet](train/cnn/mobilenet.md)
- 🐾 [SENet](train/cnn/senet.md)
- 🐾 [EfficientNet](train/cnn/efficientnet.md) 🐾
- 🐾 [conv 1* 1的作用](train/cnn/conv1.md)
- 🐾 [CNN介绍](train/cnn/introduce_cnn.md) ->[CNN网络代码讲解加部署 ⭐](https://github.com/WZMIAOMIAO/deep-learning-for-image-processing)
- 🐾 [LeNet, AlexNet, VGG](train/cnn/lenet_alexnet_vgg.md)-> [Inception](train/cnn/inception.md) -> [ResNet](train/cnn/resnet.md) -> [DSC理解](train/cnn/dsc.md) ->->[SENet](train/cnn/senet.md) [DenseNet](train/cnn/densenet.md)-> [MobileNet](train/cnn/mobilenet.md) ->[EfficientNet](train/cnn/efficientnet.md)
- 🐾. [conv 1* 1的作用](train/cnn/conv1.md)
- 🐾 [上采样](train/cnn/up.md)

#### 🍃[RNN/NLP](train/rnn/route.md)

- [RNN/LSTM/GRU](train/rnn/introduce_rnn.md)

- [Attention&Transformer](train/rnn/introduce_attention.md)
- [Bert]()
- 🐾 [RNN/LSTM/GRU](train/rnn/introduce_rnn.md)
- 🐾 [Attention&Transformer](train/rnn/introduce_attention.md)

#### 🍃GAN

Expand All @@ -260,19 +259,12 @@
#### 🍃Detection

- 🐾 [RCNNs](train/detection/rcnns.md)
- 🐾 YOLO
- [yolo v1](train/detection/yolov1.md)
- [yolo v2](train/detection/yolov2.md)
- [yolo v3](train/detection/yolov3.md)
- [yolo v4](train/detection/yolov4.md)
- [Yolox](train/detection/yolox.md) | [Github](https://github.com/Megvii-BaseDetection/YOLOX)
- 🐾 YOLO:[v1](train/detection/yolov1.md)->[yolo v2](train/detection/yolov2.md)->[yolo v3](train/detection/yolov3.md)->[yolo v4](train/detection/yolov4.md)->[Yolox](train/detection/yolox.md)

#### 🍃Segmentation

- 🐾 [分割综述(2020)](train/segmentation/introduce.md)
- 🐾 [UNet Family](https://github.com/ShawnBIT/UNet-family)
- [UNet](train/segmentation/unet.md)
- [UNet++](train/segmentation/unetpp.md)
- 🐾 [UNet Family](https://github.com/ShawnBIT/UNet-family):[UNet](train/segmentation/unet.md)->[UNet++](train/segmentation/unetpp.md)
- 🐾 [PSPNet](train/segmentation/pspnet.md)
- 🐾 [DUC & HDC](train/segmentation/duc_hdc.md)

Expand All @@ -286,7 +278,7 @@
- 🐾 [小样本与元学习](train/fsl/fsl_metalearning.md)
- 🐾 [综述---《Generalizing from a Few Examples: A Survey on Few-Shot Learning》学习](train/fsl/introduce.md)|[小样本论文收集-GitHub](https://github.com/tata1661/FewShotPapers) |[视频](https://www.bilibili.com/video/BV1Jh411X7FG?from=search&seid=4539837598698253223)
- 🐾 [元学习的方法实现小样本分类](fsl_metalearning2.md)
- 🐾 [meta-learning系列](metalearning.md)
- 🐾 meta-learning系列
- 元学习系列(零):小样本学习与元学习综述
- 元学习系列(一):Siamese Network(孪生网络)[1](https://zhuanlan.zhihu.com/p/35040994)|[2](https://www.cnblogs.com/wj-1314/p/11556107.html)|[3](https://zhuanlan.zhihu.com/p/142381922)
- 元学习系列(二):Prototypical Networks(原型网络)
Expand All @@ -300,18 +292,41 @@

#### 🍃 3D

- [3D学习路线-1](https://zhuanlan.zhihu.com/p/97299116)[2](https://github.com/qxiaofan/awesome_slam_computer_vision_resources)
- [3D图像的表示-1](https://www.cnblogs.com/geeksongs/p/13498145.html)[2](https://zhuanlan.zhihu.com/p/42772630)[3](https://www.linkresearcher.com/information/ed62a994-a3e5-4d9a-987e-2b2a67748a14)
- [3D成像-1](https://www.sohu.com/a/309203564_100166336)[2](https://bbs.huaweicloud.com/blogs/180872)
- [3D历史-1](https://www.pianshen.com/article/10101195970/)
- [PointNet 视频B站](https://www.bilibili.com/video/BV1Pp4y1473K?from=search&seid=4936235752772443171&spm_id_from=333.337.0.0)
- 🐾 [3D学习路线](https://zhuanlan.zhihu.com/p/97299116)->[3D视觉GitHub](https://github.com/qxiaofan/awesome_slam_computer_vision_resources)
- 🐾 [3D图像的表示](https://www.cnblogs.com/geeksongs/p/13498145.html) -> [3D成像](https://bbs.huaweicloud.com/blogs/180872)

#### 🍃 Anomaly Detection

<<<<<<< HEAD
- [综述与研究进展](https://blog.csdn.net/qq_36560894/article/details/120698709)


### 🍬 损失函数:[详细见AICore中的loss](https://github.com/FelixFu520/AICore/blob/main/dao/losses/__init__.py)
=======
- 🐾 综述与研究进展--[1](https://blog.csdn.net/qq_36560894/article/details/120698709)
- [CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows]()
- [PaDiM](train/anomalydetection/padim.md)


### 🍬 损失函数

- 🐾 [距离计算方法总结](train/loss/distance.md) (2022-01-06)

- 🐾 [交叉墒](train/loss/entropy.md)
- 🐾 [Pytorch的损失函数](train/loss/losses_pytorch.md)
- [三个相关系数](https://www.cnblogs.com/yjd_hycf_space/p/11537153.html)
- [机器学习-距离度量](https://blog.csdn.net/xjp_xujiping/article/details/108576496)|[2](https://zhuanlan.zhihu.com/p/46626607)
- [方差](https://blog.csdn.net/lijinxiu123/article/details/52450858)|[2](https://geek.digiasset.org/pages/mathbasic/correlation-co-variances-variance-coeffi_21Mar07115144053588/#%E4%B8%89%E5%8D%8F%E6%96%B9%E5%B7%AE)|[3](https://zhuanlan.zhihu.com/p/86181679)|[4](https://zhuanlan.zhihu.com/p/68967565)|[5](https://blog.csdn.net/lilong117194/article/details/78399568)
- Lovasz-Softmax Loss
- Exponential Logarithmic loss
- Focal Loss + Dice Loss
- BCE + Dice Loss
- Generalized Dice loss
- Tversky Loss
- IOU Loss
- Dice Loss
- Focal Loss
>>>>>>> a619f5763e328f7441ace9c037b9e528fb08afc3
### 🍬 优化器

Expand All @@ -326,14 +341,19 @@

### 🍬 评价指标

<<<<<<< HEAD
- 🐾 [分割评价指标](train/assessment/seg_metris.md)
- 🐾 [目标检测评价指标-mAP](train/assessment/map.md) [code](https://github.com/TD-4/mAP)
=======
- 🐾 [语义分割评价指标](train/assessment/seg_metris.md)
- 🐾 [mAP](train/assessment/map.md) [code](https://github.com/TD-4/mAP)
>>>>>>> a619f5763e328f7441ace9c037b9e528fb08afc3
### 🍬 网络可视化

#### 🍃权重、梯度、特征图可视化

- 🐾 [网络中间层显示工具CAM](https://github.com/frgfm/torch-cam) 🐾 [blog](https://cloud.tencent.com/developer/article/1674200)
- 🐾 [网络中间层显示工具CAM](https://github.com/frgfm/torch-cam) -> [blog](https://cloud.tencent.com/developer/article/1674200)
- 🐾 [模型权重理解DeepDream](https://github.com/TD-4/Pytorch-Deep-Dream)
- 🐾 [网络结构显示工具权重/梯度/特征图/混淆矩阵tensorboard](https://github.com/TD-4/PyTorch_Tutorial)
- 🐾 [网络显示工具visdon](https://github.com/fossasia/visdom)
Expand All @@ -357,7 +377,17 @@

## 3⃣️测试模型--train e↔︎test e ↓

<<<<<<< HEAD
### 🍬 [欠拟合&过拟合](train/overfitting/introduce_overfitting.md)
=======
### 🍬 过拟合[🐾](train/overfitting/introduce_overfitting.md)

- 🐾 [正则化](train/overfitting/regularization.md)
- 🐾 early stop
- 🐾 数据增强
- 🐾 dropout
- ...
>>>>>>> a619f5763e328f7441ace9c037b9e528fb08afc3
## 4⃣️ 微调/迁移学习

Expand All @@ -368,7 +398,11 @@

### 🍬 压缩

<<<<<<< HEAD
- 🐾 [序列化](infer/serialization/introduce.md) -{[yaml](infer/serialization/yaml.md)|[protobuf](infer/serialization/protobuf.md)|[grpc](infer/serialization/grpc.md)}
=======
- 🐾 [序列化](infer/serialization/introduce.md)->[yaml](infer/serialization/yaml.md)->[protobuf](infer/serialization/protobuf.md)->[grpc](infer/serialization/grpc.md)
>>>>>>> a619f5763e328f7441ace9c037b9e528fb08afc3
### 🍬 数据后处理

Expand Down
13 changes: 12 additions & 1 deletion train/preprocess/loaddata.md
Original file line number Diff line number Diff line change
Expand Up @@ -186,11 +186,22 @@ tmpfs /tmp tmpfs size=30G 0 0
1. 尽量简化预处理的操作,使用 numpy、opencv 等优化过的库,多多利用向量化代码,提升代码运行效率;
2. 尽量缩减数据大小,不要传输无用信息。

### 2.6 把数据弄成缓存文件

参考链接:https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/data/datasets/coco.py

### 2.7 DALI库

将数据预处理部分放到GPU上执行:

- github:https://github.com/NVIDIA/DALI
- docs:https://docs.nvidia.com/deeplearning/dali/user-guide/docs/index.html
- 博客:https://www.aiuai.cn/aifarm1755.html

### 2.6 其他

1. 使用 `TFRecord` 或者 `LMDB` 等,减少小文件的读写;
2. 使用 `apex.DistributedDataParallel` 替代 `torch.DataParallel`,使用 `apex` 进行加速;
3. 使用 `dali` 库,在 gpu 上直接进行数据预处理。

## 3. 实验

Expand Down

0 comments on commit 66c0c00

Please sign in to comment.