baowenbo
Why isn't ‘DAIN’ running very fast
On my TeslaT4 machine, a frame can be inserted in about 4 seconds. In other words, an hour of video costs me about 100 hours, and the entire packet size is about 200M. Where is the time bottleneck for model inference?
My Titan RTX is going at about 4s/it as well. 720p content /w half precision float enabled. CPU is a 5-ghz 8-core 9900K OC.
GPU and CPU isn't fully utilized; 20% and 50% peak usage reported, respectively. GPU VRAM usage never surpases 8-GB; got 24-gb in total.
The clip I'm trying to process is going to take 2 or 3 days minimum. I am not complaining; just sharing my results.
I'll try to run the code distributed and see what I get.
My Titan RTX is going at about 4s/it as well. 720p content /w half precision float enabled. CPU is a 5-ghz 8-core 9900K OC.
GPU and CPU isn't fully utilized; 20% and 50% peak usage reported, respectively. GPU VRAM usage never surpases 8-GB; got 24-gb in total.
The clip I'm trying to process is going to take 2 or 3 days minimum. I am not complaining; just sharing my results.
I'll try to run the code distributed and see what I get.
Did you find that the rendering effect of this model is not very friendly to subtitles?