how to change the demo.py for my own scenario

[ElmiraNa]

Hi every one In my application checkpoint file and dataset.yaml are always the same. Also I want to use the same video for all the images. Actually just the image is the variable input in each run now, I think can use some of the results in all of my executions for different images, for example there is no need to create "generator, kp_detector, in load_checkpoint " or kp_driving in make_animation.... in every time of execution.

I'm not familiar with torch and other ML frameworks... How can I change the code and use the results of mentioned variables in all of my execution to save time consumption.

thank you in advance

[AliaksandrSiarohin]

You can save all the models + kp_driving in memory.

For each source you could also save encoding of the source.

[ElmiraNa]

thanks for your regard @AliaksandrSiarohin I extract kp_driving and it is a dictionary type, I load it as pickle file and when run it, the execution time has been reduced about 6 seconds, but there was no change in gpu consumption. Now please if you wouldn't mind, answer these question:

1. You said: save all models.....I did not get the point. We always load 'config/vox-256.yaml' and 'vox-cpk.pth.tar' as checkpoint, what should I exactly do for save the model?

2. In demo.py in line77 there is a code line: predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]) it takes about 16 sec to run. what exactly it does and is there a way to reduce its running time?

[ElmiraNa]

@AliaksandrSiarohin I really need your help about above questions, please

[AliaksandrSiarohin]

1. I mean load the models only once for all videos, if you already do it than nvm.
2. It saves the frame into a list in order to save it to the disk later.

[ElmiraNa]

thank you so much for your quick answer. and excuse me if I bother you with my questions prediction.append.... transfer data from gpu to cpu? it takes too much time and is it any way to do it faster?

[AliaksandrSiarohin]

.cpu() transfers data. You can try to create a pytorch tensor t and save all the frames there, then use io.mimsave('q.mp4', t.cpu().numpy()) maybe this will be faster.

[ElmiraNa]

thanks again, i will inform you about result