1 Convolution channel cropping
1.1 Perceptual understanding of the number of convolution channels
This is the experimental result of the flying paddle:
2 Model quantification
Quantization is also effective on the GPU, you can look at the experimental results provided by the flying paddle,
2 Model distillation (Remarks: It can improve accuracy and cannot be accelerated)
Model distillation can be used to improve the accuracy of small models;