Hyperparameters for convolutional layers:
- Kernel convolution kernel size
- Padding
- Stride
1. Padding in the convolutional layer
2 Stride in the convolutional layer
3. Code implementation
3. Q&A
-
- Hyperparameter kernel size, the filling is generally kernel size -1, and the stride generally depends on whether the speed of calculation needs to be optimized.
-
- The variable length of the convolution kernel is generally chosen to be an odd number, which is convenient for calculation.
-
- Generally, everyone directly uses the classic network structure, such as ResNet.
-
- The essence of machine learning is a compression algorithm. Machine learning will always lose information, just like information theory.
-
- The bottom layer uses a large Kernel, and the upper layer uses a small Kernel
-
- Different convolutional layers can be understood as identifying multiple different textures of an image.
-
- The Industrial Revolution is about replacing human beings with machines. Science is also a process from expensive to cheap.
reference
https://www.bilibili.com/video/BV1Th411U7UN?p=1