Arithmetic intensity of some vision models
GFLOPS and params are obtained from Models and pre-trained weights — Torchvision main documentation. Parameters are all in float32 (4 bytes).
model | Params | GFLOPS | FLOPS/byte |
---|---|---|---|
MobileNet_V3_Large | 5483032 | 0.22 | 10.0 |
ShuffleNet_V2_X2_0 | 7393996 | 0.58 | 19.6 |
ResNet152 | 60192808 | 11.51 | 47.8 |
VGG19 | 143667240 | 19.63 | 34.2 |
EfficientNet_V2_L | 118515272 | 56.08 | 118.3 |
ViT_B_16 | 86567656 | 17.56 | 50.7 |
Swin_V2_B | 87930848 | 20.32 | 57.8 |