Need to optimize the memory allocation of VGG16 model in Fluid
Created by: kuke
Maximum feasible batch size on TITAN X (Pascal, 12GB global memory)
| cifar10 ( 3 x 32 x 32 ) | flowers (3 x 224 x 224) | |
|---|---|---|
| Fluid | 966 | 19 |
| TensorFlow | 1618 | 33 |
- CMAKE_BUILD_TYPE:STRING=Release
- CUDNN: v6
Compared to TensorFLow, the VGG16 model in Fluid can only run with much smaller batch size right now.