当前位置:网站首页>pytorch_ Yolox pruning [with code]
pytorch_ Yolox pruning [with code]
2022-07-06 22:25:00 【Meat loving Peng】
Catalog
Fine tuning training after pruning
Conv And BN The fusion reasoning of layer accelerates
Environmental Science
pytorch 1.7
loguru 0.5.3
NVIDIA 1650 4G
intel i5-9th
torch-pruning 0.2.7
Installation package
pip install torch_pruning
Note: This project is in b standing up Lord Bubbliiiing Heyuan YOLOX The official code is integrated .
1. Added feature Visualization
2. It can be turned on during training EMA function
3. Network pruning ( Support s,m,l,x)
3.1 Support single convolution pruning
3.2 Support network layer pruning
4. Fine tuning training after pruning
5.Conv And BN The fusion reasoning of layer accelerates
6. preservation log Information
Dataset format : use voc Dataset format
feature Visualization
stay tools/Net_Vision.py Implement for visual code . You can import NV function , Realize channel visualization .
eg:
features = [out_features[f] for f in self.in_features]
[x2, x1, x0] = features # shape is (batch_size,channels,w,h)
NV(x2)
Network pruning
Reference paper :Pruning Filters for Efficient ConvNets
Import the pruning tool
import torch_pruning as tp
If you need to see yolov4 Of , You can see :YOLOv4 prune 【 The attached code 】_ Meat loving Peng's blog -CSDN Blog _yolov4 prune
Adopt channel pruning , Instead of weight pruning .
Before pruning, you need to pass tools/prunmodel.py save_whole_model(weights_path, num_classes) Function saves the weight and structure of the model .
weights_path: Weight path
num_classes: Own category quantity
model = YOLOX(num_classes, 's') # It needs to be modified according to the number of classes s finger yolox-s
Support the pruning of a convolution : call Conv_pruning(whole_model_weights):
pruning_idxs = strategy(v, amount=0.4) # 0.4 Is the pruning rate Modify as needed , The larger the number, the more you cut
For the pruning of a single convolution , Two local values need to be modified , The convolution layer here needs to be obtained by printing the model , Don't guess blindly :
if k == 'backbone.backbone.dark2.0.conv.weight'pruning_plan = DG.get_pruning_plan((model.backbone.backbone.dark2)[0].conv, tp.prune_conv, idxs=pruning_idxs)
Support network layer pruning : call layer_pruning(whole_model_weights):
included_layers = list((model.backbone.backbone.dark2.modules())) # Prune for a certain layer
Note: After successful pruning , It will print the parameter variation of the model ! If not printed , That means you cut it wrong , Check it out !
After pruning log The log file will be saved in logs Under the document
Fine tuning training after pruning
take train.py Medium pruned_train Set to True.
False For normal training , And then modify it yourself batch_size.
Pay attention to revision model_path and classes_path, Otherwise, it will report a mistake !
The size of network input before pruning must be consistent with that during fine tuning training and reasoning !
Train your own dataset
If you have Bubbliiiing up Code of master , You will soon be able to get started . The dataset uses VOC In the form of
VOCdevkit/ `-- VOC2007 |-- Annotations ( Deposit xml Label file ) |-- ImageSets | `-- Main `-- JPEGImages ( Store image )
stay model_data Create a new one in new_classes.txt, Write your own class inside . function voc_annotation.py, Will generate... In the current directory 2007_train.txt Document and 2007_val.txt file .( You can check whether it is successfully generated )
stay train.py in , take classes_path It is amended as follows model_data/new_classes.txt【 Wait for the prediction , Also needed in the yolo.py Modify here 】
Then modify other super parameters as needed to train , Training weights will be saved in logs In file ( Save weights by default , Without network structure )
forecast
Parameter description : The inputs of the following terminals are optional
--predict: Prediction model
--pruned: Turn on pruning prediction or training
--image: Image detection
--video: Start video detection
--video_path: Video path
--camid: camera id Default 0
--fps:FPS test
--dir_predict: Predict the image under a folder
--phi: You can choose s,m,l,x etc.
--input_shape: Network input size , Default 640
--confidence: Confidence threshold
--nms_iou:iou threshold
--num_classes: Number of categories , Default 80
--fuse: Whether the convolution layer and BN Layer fusion accelerates , Default False
The input terminal :
# Image prediction python demo.py --predict --image
# Video prediction python demo.py --predict --video --camdi 0
# fps test python demo.py --predict --fps
The default forecasts are yolox_s, If you want to specify another network , Input :( It is important to note that in yolo.py Modify the weight path , If it is your own data set , It needs to be revised classes_path)
# Use yolox_l forecast python demo.py --predict --image --phi l
Conv And BN The fusion reasoning of layer accelerates
Other commands can be used together , For example, using conv and bn Reasoning in the way of fusion
python demo.py --predict --image --fuse
Found by test FPS Promoted 3 frame /s about ( my GPU yes 1650)
Saving of log files
This project adopts loguru Tools capture logs , Some log records during detection and training will be automatically recorded , Save in logs Under the document , One log The maximum size of the file I set is 1 MB, If beyond this range , Will automatically generate a new .log file , You can modify this value by yourself , Or modify the log save time ( So as not to save too many logs ). If you don't want this feature , You can find the corresponding position and comment it out .
Here is just to help you build wheels , Implement some functions with as simple code as possible , You don't have to look at complex engineering code anymore , The final effect needs to be adjusted patiently , Take your time “ Alchemy ”!
The weight
link : Baidu SkyDrive Please enter the extraction code https://pan.baidu.com/s/1Jbq8dCv893rZ7RkaANUZgQ Extraction code :yypn
Code ( If it helps , Excuse me star Chant ~):
边栏推荐
- Management background --3, modify classification
- Inno Setup 打包及签名指南
- 2500个常用中文字符 + 130常用中英文字符
- [sdx62] wcn685x will bdwlan Bin and bdwlan Txt mutual conversion operation method
- [sciter bug] multi line hiding
- Applet system update prompt, and force the applet to restart and use the new version
- 2500 common Chinese characters + 130 common Chinese and English characters
- QT | UDP broadcast communication, simple use case
- [sciter]: encapsulate the notification bar component based on sciter
- Assembly and Interface Technology Experiment 6 - ADDA conversion experiment, AD acquisition system in interrupt mode
猜你喜欢
PVL EDI project case
That's why you can't understand recursion
[linear algebra] determinant of order 1.3 n
Hardware development notes (10): basic process of hardware development, making a USB to RS232 module (9): create ch340g/max232 package library sop-16 and associate principle primitive devices
MySQL数据库基本操作-DML
将MySQL的表数据纯净方式导出
2500 common Chinese characters + 130 common Chinese and English characters
硬件開發筆記(十): 硬件開發基本流程,制作一個USB轉RS232的模塊(九):創建CH340G/MAX232封裝庫sop-16並關聯原理圖元器件
在IPv6中 链路本地地址的优势
第4章:再谈类的加载器
随机推荐
数据处理技巧(7):MATLAB 读取数字字符串混杂的文本文件txt中的数据
[linear algebra] determinant of order 1.3 n
anaconda安装第三方包
(十八)LCD1602实验
CCNA-思科网络 EIGRP协议
Common sense: what is "preservation" in insurance?
labelimg的安装与使用
变量与“零值”的比较
剪映+json解析将视频中的声音转换成文本
AI enterprise multi cloud storage architecture practice | Shenzhen potential technology sharing
[sdx62] wcn685x will bdwlan Bin and bdwlan Txt mutual conversion operation method
PVL EDI 项目案例
GNN, please deepen your network layer~
PVL EDI project case
MySQL----初识MySQL
将MySQL的表数据纯净方式导出
每日一题:力扣:225:用队列实现栈
GPS from getting started to giving up (XI), differential GPS
LeetCode 练习——剑指 Offer 26. 树的子结构
2022-07-05 使用tpcc对stonedb进行子查询测试