20220420 reasoning framework analyzes what is gradient

What is gradient _guanhuazhan The blog of -CSDN Blog _ gradient

1、letterbox

Deep learning The size of the model input picture is square , The images in the dataset are generally rectangular , Rough resize It will distort the picture , use letterbox It can solve this problem better . This method can maintain the length width ratio of the picture , The rest is filled with gray .

 2、normalize

In deep learning , The image is usually normalized before model training , The normalization of the image is to adjust the size of the eigenvalue to a similar range , When not normalized , If the eigenvalue is large , The gradient value will also be large , When the eigenvalue is small , The gradient value will also be small . When the model is back propagated , The gradient value update is the same as the learning rate , When the learning rate is low , Smaller gradient values can cause slow updates , When the learning rate is high , Large gradient value will make the model difficult to converge , Therefore, in order to make the model training convergence stable , Normalize the image , Adjust the eigenvalues of different dimensions to a similar range , A unified learning rate acceleration model can be used to train , The formula is as follows ：

In style ,output Output for image pixel value ,input Enter... For image pixel values ,max And min Is the maximum pixel value and the minimum pixel value , After normalization , Adjust the pixel value to （0,1） Section .

 3、imagetotensor

Transformation tensor

4、onnx  r

tensorrt Optimized in those places ：

TensorRT course 1： Optimization principle _ Mistron _ZXL The blog of -CSDN Blog

5、nms

6、 Euclidean distance   Cosine distance

Recommended algorithm principle （ Two ） Euclidean distance calculates the similarity between objects - You know