Computation and memory requirement analysis of classical CNN model
Table 1 Comparison of memory, computation amount and parameter number of CNN classical model
| AlexNet | VGG16 | Inception-v3 | |
|---|---|---|---|
| Model memory (MB) | > 200 | > 500 | 90-100. |
| Parameters (millions) | 60 | 138 | 23.2 |
| Calculations (millions) | 720 | 15300 | 5000 |
1. Specific analysis of CNN model (Taking AlexNet network model as an example)
1.1 Network Structure
Figure 1 AlexNet network structure
AlexNet has 5 convolution layers and 3 full connection layers
C1:96×11×11×3 (number of convolution kernels/width/height/depth) 34848
C2:256×5×5×48 (number of convolution kernels/width/height/depth) 307200
C3:384×3×3×256 (number of convolution kernels/width/height/depth) 884736
C4:384×3×3×192 (number of convolution kernels/width/height/depth) 663552
C5:256×3×3×192 (number of convolution kernels/width/height/depth) 442368
R1:4096×6×6×256 (number of convolution kernels/width/height/depth) 37748736
R2:4096×4096 16777216
R3:4096 x 1000 4096000
There are 60 million parameters
1.2 AlexNet model memory size calculation
60 million (parameters) x 32 bits (float32)= 1.92 billion bits ≈228.88MB
1.3 AlexNet model calculates force consumption
Figure 2 Number of floating point operations per second and number of parameters in each layer of AlexNet model
1.4 AlexNet network model Configuration
The AlexNet network model won the ImageNet competition in 2012. AlexNet uses two GTX580 graphics cards for training, and two Gpus each train part of the network. Only in the second convolution layer and the full connection layer can two Gpus communicate with each other.
