search

mingyuan-zhang / MotionDiffuse

MotionDiffuse: Text-Driven Human Motion Generation with Diffusion Model
https://mingyuan-zhang.github.io/projects/MotionDiffuse.html
Other
835 stars 74 forks source link

What is "zero_module" for? #13

Open JeremyCJM opened 1 year ago

JeremyCJM commented 1 year ago

Hi Mingyuan,

Why zero out the parameters of the "self.out" projection module in transformers.py?

Thanks, Jeremy

mingyuan-zhang commented 1 year ago

This trick is also used in GLIDE and other Text-to-Image Generation. Zero out self.out can enforce the output to be a zero vector. The learning target of our model is the added Gaussian Noise, whose expected mean is also zero. Therefore, our model can be trained steadily with this trick.