当前位置:网站首页>solver. Learning notes of prototxt file parameters
solver. Learning notes of prototxt file parameters
2022-07-05 07:47:00 【Fall in love with wx】
Reference resources :https://blog.csdn.net/xygl2009/article/details/77484522 Thank you very much !
# -*- coding: UTF-8 -*-
train_net: "examples/YOLO_Helmet/train.prototxt" # Training profile
test_net: "examples/YOLO_Helmet/test.prototxt" # Test profile
test_iter: 4952 # Represents the number of tests ; such as , Yours test Stage batchsize=100, And your test data is 10000 A picture , Then your test times are 10000/100=100 Time ; namely , Yours test_iter=100
test_interval: 2000 # It indicates how many times your network iterations are tested , You can set up a generation of online training , Just do a test .
base_lr: 0.0005 # Indicates the basic learning rate , In the process of parameter gradient descent optimization , The learning rate will be adjusted , And the adjustment strategy can be passed lr_policy This parameter is set ;
display: 10
max_iter: 120000 # Maximum number of iterations
lr_policy: "multistep" # Set to multistep, You also need to set a stepvalue. The sum of this parameter step Very similar ,step It's a uniform, equidistant change , and mult-step It is based on stepvalue Value change
gamma: 0.5
weight_decay: 0.00005 # Indicates weight attenuation , Used to prevent over fitting
snapshot: 2000 # Save model interval
snapshot_prefix: "models/MobileNet/mobilenet_yolov3_deploy" # Save the prefix of the model
solver_mode: GPU # Whether to use GPU
debug_info: false
snapshot_after_train: true
test_initialization: false
average_loss: 10 # Take multiple times foward Of loss Average , Display output
stepvalue: 25000
stepvalue: 50000
stepvalue: 75000
iter_size: 4 # Handle batchsize*itersize After a picture , Just once ApplyUpdate The function depends on the learning rate 、method(SGD、AdaSGD etc. ) Make a gradient descent
type: "RMSProp"
eval_type: "detection"
ap_version: "11point"
show_per_class_result: true
边栏推荐
- MySQL - storage engine
- A simple method to prove 1/t Fourier transform
- RTOS in the development of STM32 single chip microcomputer
- Exit of pyGame, idle and pycharm
- About the problem that MySQL connector net cannot be cleared in MySQL
- Esmini longspeedaction modification
- What is deep learning?
- Ten thousand words detailed eight sorting must read (code + dynamic diagram demonstration)
- I 用c l 栈与队列的相互实现
- Package ‘*****‘ has no installation candidate
猜你喜欢
Acwing-宠物小精灵之收服-(多维01背包+正序倒序+两种形式dp求答案)
Significance and requirements of semiconductor particle control
With the help of Navicat for MySQL software, the data of a database table in different or the same database link is copied to another database table
Numpy——1.数组的创建
I 用c l 栈与队列的相互实现
Daily Practice:Codeforces Round #794 (Div. 2)(A~D)
Detailed explanation of miracast Technology (I): Wi Fi display
[neo4j] common operations of neo4j cypher and py2neo
Altium Designer 19.1.18 - 更改铺铜的透明度
UE5热更新-远端服务器自动下载和版本检测(SimpleHotUpdate)
随机推荐
Self summary of college life - freshman
Numpy——1.数组的创建
万字详解八大排序 必读(代码+动图演示)
Deepin get file (folder) list
Linked list (establishment, deletion, insertion and printing of one-way linked list)
Could NOT find XXX (missing: XXX_LIBRARY XXX_DIR)
Package ‘*****‘ has no installation candidate
Altium Designer 19.1.18 - 导入板框
I 用c l 栈与队列的相互实现
Day01 markdown log entry tips
Rename directory in C [closed] - renaming a directory in C [closed]
·Practical website·
TCP and UDP
Openxlsx field reading problem
assert_ Usage of param function
Temperature sensor DS18B20 principle, with STM32 routine code
Threads and processes
Count and sort the occurrence times of specific fields through SQL statements
Calibre garbled
PIL's image tool image reduction and splicing.