1. Network NiN in Network
The network NiN in the network is based on the attempted improvement of AlexNet, and the fully connected layer is removed.
The network NiN in the network is not used much later, but the idea of removing the fully connected layer is adopted.
The biggest problem with the fully connected layer is that
- uses a lot of memory
- Takes up a lot of computing bandwidth
- Convergence is fast, but can lead to overfitting
2. Code implementation
3. Q&A
-
nvidia-smi
It is nvidia's system management interface, where smi is the abbreviation of System management interface, which can collect various levels of information and view video memory usage.
-
- AlexNet, NiN still need Softmax and need to be classified.
-
- Why does NiN need two 1x1 convolutional layers, the high probability is the result of training parameters.
reference
https://www.bilibili.com/video/BV1Uv411G71b?p=1