Video Model Benchmark

Environments

• CUDA Toolkit 10.1
• CuDNN 7.6.2

Install dependencies for SlowFast. Then create link by

ln -s SlowFast/slowfast ./slowfast

Getting started

Profile by running

python main.py --model i3d

Example output

Profile model i3d

Params: 28,043,472
GPU Mem: 0.10470056533813477 MB
FLOPs: 56.814501888 GFLOPs

Module                | Self CPU total | CPU total | CUDA total | Occurrences
----------------------|----------------|-----------|------------|------------
ResNet                |                |           |            |            
├── s1                |                |           |            |            
│├── pathway0_stem    |                |           |            |            
││├── conv            |        1.448ms |   5.687ms |  102.623ms |           1
││├── bn              |        1.314ms |   3.763ms |   10.265ms |           1
││├── relu            |       48.309us |  48.309us |    1.073ms |           1
││└── pool_layer      |        1.579ms |   3.144ms |    4.938ms |           1
│ ...
└── head              |                |           |            |            
 ├── pathway0_avgpool |       67.693us |  67.693us |  961.536us |           1
 ├── dropout          |       71.013us | 127.499us |  131.072us |           1
 ├── projection       |      250.353us | 399.099us |  410.624us |           1
 └── act              |        0.000us |   0.000us |    0.000us |           0

View trace in chrome://tracing or edge://tracing.