Input Normalization/Standardization Alex and Caffe initialization parameters are normalized based on the mean value. if normalization is not done, training will fail because the input is half as large.This is also why Caffe forces the calculation of image mean values for samples.
In this way, the pixel value [ 0,255 ] is adjusted to approximately [-128,128 ].Although the format of the image data is regular, it is quite useful to normalize it.
Face Key Point Network FAN: 5s DAN: time 1.5s openpose face: poor results 3000FPS: can only look at the face, opencv or dlib has integration, fast speed, 10ms Van Face: https://github.com/lsy17096535/face-landmark Allegedly 5ms (Lightweight Network) Vanilla CNN：
https://github.com/cunjian/face_alignment look at boundary:https://github.com/wywu/LAB face-landmark： https://github.com/jiangwqcooler/face-landmark-localization The Nest of the Great Spirit Among them, 5 and 6 are very powerful. VANface is quite good except for the side face effect.
Keras Deep Learning Framework can be used to understand what deep learning can be used to do. The following is an introduction to some basic image classification using Keras. Welcome to exchange. Reference: HTTPS://Moranzhou.github.io/Tutorials/Machine-Learning/Keras/2-3-CNN/ Version I used: Python2.7, numpy1.13.1, Theano0.9.0, Keras2.0.6, h5py2.5.0, opencv2.4.13, WIN7 system. In order to classify images, data sets are needed first, and the downloaded image data sets need to be converted into numpy matrices that Keras
This article is an improved version of my previous blog. Due to limited time, only a simple optimization has been made. Related Articles: Import excel into Database On April 1, 2018, add a new download address link: click Open Source Download Address I am very sorry that this link address has not been published in this article.I hope it's not too late. The last article was about such a data
MultiLabel means that a sample may belong to multiple classes at the same time, that is, there are multiple labels.For example, for an L-sized cotton-padded jacket, the sample has at least two labels-model number: L and type: winter clothing.Only two links are posted here for reference:To solve the problem of multi-label classificationLearning Problems of multi-label Data
1. Create an xmlhttpRequest object 2. Set callback monitoring 3. Open a connection Accept two parameters: 1.http method2.httpurl 4. Set the request header (get does not have this step) Notify the browser of the relevant settings for the requestor 5. Making requests Parameters: specific value sent readystate: 0-(uninitialized) send () method has not been called 1-(load) send () method has been called, request is being sent 2-(load complete) send ()
use function mapminmax
1 The default map range is [-1,1 ], so if [ 0,1 ] is required, provide the parameters in this format:
MappedData = mapminmax(OriginalData, 0, 1);
2 is only normalized by rows. If it is a matrix, each row is normalized separately. If the whole matrix needs to be normalized, use the following method:
FlattenedData = OriginalData(:)';The % expansion matrix is a column and then transposed to a row.
The Method of Running libsvm under weka http://datamining.xmu.edu.cn/bbs/forum.php?mod=viewthread&tid=120 weka 3.6.9 failed to install using the " old method" below!" problemevaluating classificator rand" error thrown!
1. first install WEKA3.7.9, install libsvm with tools-package manager, and then find libsvm.jar in wekafiles\packages\LibSVM\lib under the user directory.
2. Install WEKA3.6.9, and put libsvm.jar in the first step into the installation directory of WEKA3.6.9.（）
3. amend the RunWeka.ini document of WEKA3.
1, block level element display:block a, each block-level element starts from a new row, and the following elements also start from another row.(A block-level element has an exclusive row)
b, element height, width, row height, and top and bottom margin can be set. c, element width is 100% of its own parent container (consistent with the width of the parent element) without setting, unless a width is set.