Closed CellNw closed 5 months ago
It probably was forgotten to take out of the command. Looking at the code, the teacher_model will not be used if the training mode is consistency_training: https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/train_util.py#L488
Only the model and target model will be passed to the loss calculation https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L106 And I believe target_model is just the ema model of model (?)
It probably was forgotten to take out of the command. Looking at the code, the teacher_model will not be used if the training mode is consistency_training:
Only the model and target model will be passed to the loss calculation
And I believe target_model is just the ema model of model (?)
Thanks for the reply, I checked the code you mentioned https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L106
The euler_solver and henu_solver will pass the teacher_model. heun_solver: https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L148 euler_solver: https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L168
and those solver will be used to gen 'x_t2'. https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L193-L196
then 'x_t2' will be passed target_model and finally pass to the loss calculation. https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L199
It probably was forgotten to take out of the command. Looking at the code, the teacher_model will not be used if the training mode is consistency_training: https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/train_util.py#L488
Only the model and target model will be passed to the loss calculation https://github.com/openai/consistency_models/blob/e32b69ee436d518377db86fb2127a3972d0d8716/cm/karras_diffusion.py#L106
And I believe target_model is just the ema model of model (?)
Thanks for the reply, I checked the code you mentioned
The euler_solver and henu_solver will pass the teacher_model. heun_solver:
euler_solver:
and those solver will be used to gen 'x_t2'.
then 'x_t2' will be passed target_model and finally pass to the loss calculation.
Thank u, so does the consistency training mode need the teacher model path ?
And I also met the "shape mismatch" in the update ema process of the target model, do u know why?
Thank u, so does the consistency training mode need the teacher model path ?
no,consistency training mode dont't need the teacher model.
And I also met the "shape mismatch" in the update ema process of the target model, do u know why?
Have you solved the problem?I also met this error.
writen In
launch.sh
section "Consistency training on class-conditional ImageNet-64, and LSUN 256"so,Why use teacher model in Consistency training.In my understanding,Consistency training is training model isolate. is anything wrong? here is Consistency Training (CT) Algorithm in paper