search

orangeduck / Motion-Matching

Learned Motion Matching example implementation and source code for the article "Code vs Data Driven Displacement"
MIT License
694 stars 105 forks source link

How much the loss is acceptable? #18

Closed Maekdzp closed 1 year ago

Maekdzp commented 1 year ago

I'm a rookie in neural network, so the question may be stupid. I changed the training data from referenced dataset and found the loss of decompressor reaches 1.9 after 500k iterations. Generated .bvh files also shows obvious difference. So I wonder if I choose bad training data and what can I do to reduce loss.

orangeduck commented 1 year ago

Ultimately the visual result is what matters when deciding how much difference is acceptable. I'm surprised you get an obvious difference on the referenced dataset. When I train this locally yes there is a small different between the generated bvh files but it is extremely subtle. Did you change any of the parameters or data for training?

Maekdzp commented 1 year ago

Yeah, I changed batch size to 256 and use cuda instead of cpu to accelerate training. But it should not affect the training result too much. I have tried several times using different pieces of dataset and reached better results. Maybe I just used bad datasets when I tried first time. Thank you for your reply.