1. GoogLeNet with parallel connections
The first neural network can do more than 100 layers.
The name of the Inception block, taken from Inception, is to go deeper and deeper into the dream.
Inception reduces the number of parameters and reduces the amount of computation. It is smaller than convolution with 3x3 Conv and 5x5 Conv directly.
https://cv.gluon.ai/model_zoo/classification.html
Inception V3 accuracy is much higher than VGG, from 0.7 to close to 0.8, the memory will be much larger and more complicated.
Mushen's network ResNeSt: Split-Attention Networks is the most accurate on this graph.
2. Code implementation
3. Q&A
-
- 1 x 1 Conv convolution is to reduce the number of channels channel
-
- The deeper the convolutional layer, the more features are learned.
reference
https://www.bilibili.com/video/BV1b5411g7Xo?p=1