VideoAI-Speedrun

VQA & VSR Efficiency and Runtime Evaluation

AIS: Vision, Graphics and AI for Streaming Workshop at CVPR

python runtime_vqa.py --frames 30 --imsize 1920 1080 --repeat 10 --fp16

We run the model 10 times using a [batch, frames, 3, H, W] input. We warmup and syncronize GPU times to make sure that we get an accurate runtime. In the example the input to the model VQAModel is a tensor [1, 30, 3, 1920, 1080] representing a clip of 30 frames a FHD.

Sample output:

INFO:AIS24-VQA:------> INPUT 1920x1080, 30 frames, 1 clip
INFO:AIS24-VQA:------> Average runtime on clip 30-frames  of (test_model) is : 139.535997 ms
INFO:AIS24-VQA:------> Average runtime per frame of (test_model) is : 4.651200 ms
INFO:AIS24-VQA:------> Average FPS of (test_model) is : 214.998285 FPS
INFO:AIS24-VQA:------> MACs per clip 30-frames : 397.309594881 [G]
INFO:AIS24-VQA:------> MACs per frame : 13.243653162700001 [G]
INFO:AIS24-VQA:------> #Params 2.225153 [M]

We report MACs. One MACs equals roughly two FLOPs.

Rquirements

• ptflops --- for calculating MACs / FLOPs.