(Focus on the conclusion of the experiment and find thinking)
Three classic lightweight networks:
Squeezenet compression expansion Fire Module
MobileNet Depthwise deep separable convolution
ShuffleNet channel cleaning
Read today Squeezenet
Squeeze Net: Alexnet-level accuracy with 50X fewer parameters and <0.5MB model size , Forrest N. Iandoal, Song Han, 2017
Innovation
Squeezenet is a well-designed lightweight network that uses common compression techniques: SVD, Pruning, and quantized 8-bit quantization to further compress the model without losing accuracy. It is widely used for mobile object detection.
(It is much more efficient to train with a small model at the beginning than pruning a large model into a small model)
1 Benefits of small models
- 1 Distribute training more efficiently
- 2 Update model parameters faster in the cloud
- 3 Small storage and memory usage is easy to apply to FPGA field programmable gate array
2 Neural network design space exploration
- 1 Microscopic exploration: the dimension and configuration of each layer
affect the model size and accuracy of the cnn structure by the
same amount of parameters: deep and narrow models are better than wide and shallow models - 2 Macro exploration: end-to-end organization module or other layers
3 SqueezeNet high-precision less parameter strategy
3.1 Innovation strategy
- 1 Decrease parameters: replace 3 3 with 1 1 convolution kernel (9x reduction)
- 2 Reduce the parameter: reduce the number of input channels use Squeeze layer (1 * 1 convolution)
- Parameter calculation input chanels * ker-sizes squared * filters nums
- 3 Post-positioning: making feature maps large and rich
- Use multiple sizes of convolution kernel calculations to retain more information and improve classification accuracy
3.2 Fire Module
Each Fire module has three hyperparameters
1 input channel (96) with fire M (M <96) 1 1 convolution to reduce the dimension
2 Expand Use 1 1 and 3 * 3 convolution in parallel to obtain different receptive field feature map Similar to inception—)
3 Concat: splicing feature maps
3.3Sqeeze net (input 224 * 224)
The bypass on the left is not used, the bypass in the middle, and the complex on the right is a
module that has been provided functions in the following boxes
4 evaluation
Here is a point: Squeezenet is a small model but it can also be quantified and compressed but the accuracy has little effect