real-stanford / diffusion_policy

[RSS 2023] Diffusion Policy Visuomotor Policy Learning via Action Diffusion
https://diffusion-policy.cs.columbia.edu/
MIT License
1.39k stars 260 forks source link

Is EMA used to update the training weights? #63

Open StarCycle opened 5 months ago

StarCycle commented 5 months ago

Hello @cheng-chi ,

In your colab notebook I find that you mention the EMA model can "accelerate training and improves stability":

b632a5a0185bbd8b7dde82955fef541

However, in the training loop, you may only update the EMA model with the training model:

659b2caf46c1603a22cfdbdefd5c275

The ema.step function will not update/smooth the training weights. Why can it accelerate training?

If my understanding is wrong, please let me know

Best, StarCycle