rarzumanyan
commented
2 years ago Hi @sotelo
A dummy solution to this problem would be to repeat the 3 channels and then save it as a usual BGR video.
So far that's the only way ;) If that's possible I'd recommend you to work this around in pytorch. Since I'm not a pytorch guru I can only suggest smth like this:
- Convert float
cuda tensorto uint8cuda tensor. - Repeat your float
cuda tensor3 times to create planar RGBcuda tensor. - Convert planar RGB
cuda tensorto planar RGBSurface. - Use
PySurfaceConverterclass to conver planar RGBSurface> YUV444Surface. - Feed YUV444
Surfaceto Nvenc (it supports YUV444 as well).
Looks like a long way, but at least everything will be done on GPU and it does these memcopy operations really fast.
P. S. Proper solution would be:
- Implement Y > YUV444 conversion in VPF (new feature).
- Convert float
cuda tensor> uint8cuda tensor. - Convert uint8
cuda tensor> YSurface. - Run Y > YUV444 conversion in VPF.
- Feed YUV444 to Nvenc.
It's feasible to implement, but not done yet in VPF.
sotelo
Hi @rarzumanyan!
First of all, let me thank you for the work that you've put into this library! It's very useful and we've been able to speed up our code significantly thanks to it.
On our current use case, we are processing videos and using a PyTorch model to extract a mask for them. The mask is a 1 float between 0 and 1 and has the same size as the resolution of the original video in terms of height and width. Then, the question is: What would be the recommended way to save the mask as a video file? (Ideally, we'd like to be able to visualize the mask as a black and white video in VLC)
A dummy solution to this problem would be to repeat the 3 channels and then save it as a usual BGR video. Is there a more efficient way to do this?
Thank you for your help and happy to provide more details. José