search

universome / stylegan-v

[CVPR 2022] StyleGAN-V: A Continuous Video Generator with the Price, Image Quality and Perks of StyleGAN2
https://universome.github.io/stylegan-v
333 stars 36 forks source link

Questions about details of hyperparameters in AlignedTimeEncoder #21

Closed SongDoHou closed 2 years ago

SongDoHou commented 2 years ago

Hi, thank you for sharing this amazing works!!

Can I ask you about the hyperparameters in AlignedTimeEncoder?

When we train the model on SkyTimelapse dataset, what is the values of the member variables (time_enc_dim, min_period_len, max_period_len) in AlignedTimeEncoder?

I think each values are 256, 256, 1024. Is it correct?

(Current I'm on limited (HW and Internet) situation, It's hard to run it.)

universome commented 2 years ago

Hi, thank you! Yes, the hyperparameters for SkyTimelapse are the ones which you specified

SongDoHou commented 2 years ago

Thank you so much!!

Can I ask you one more??

How to set the parameters value when we inference using "generate.py" for an hour?

universome commented 2 years ago

You mean, how to generate an hour-long video using the inference script? That's tricky to be honest. By default, the video length is controlled by the --video_len argument (and also --fps of course, but let's consider it is fixed to 25 frames per second). If you want to generate a video which lasts N seconds, then you need to set --video_len to N * 25. But generating an hour-long video (i.e. 3600 * 25 = 90000 frames) would take very much time and will likely result in RAM running out of space. So you can generate it in chunks (controlling the start frame index with --time_offset argument) and then concatenate the videos with something like ffmpeg.

SongDoHou commented 2 years ago

Thank you for letting me know.

I appreciate it!!