prepare_rotary_positional_embeddings

[HaoZhang990127]

Feature request / 功能建议

train video with different resolution

Motivation / 动机

train video with different resolution

Your contribution / 您的贡献

Hi,

Thank you for your nice work again. I am confuse about the prepare_rotary_positional_embeddings function,

CogVideo/finetune/train_cogvideox_lora.py

Line 822 in da5fc1c

def prepare_rotary_positional_embeddings(

, in here you use base_height: int = 480, base_width: int = 720 for your fix resolution training. Could you explain the meaning of the base_height and base_width? If i want to train a video with height=720 and width=480, what base_height and base_width should be set in here? And if I want to train the video with random resolution, what base_height and base_width should be set in here?

Thank you

[foreverpiano]

same problem