search

sail-sg / MDT

Masked Diffusion Transformer is the SOTA for image synthesis. (ICCV 2023)
Apache License 2.0
500 stars 35 forks source link

Training Problem #32

Closed xv994 closed 5 months ago

xv994 commented 5 months ago

Hey guys. I want to know how the algorithm judge when should stop the training. It seems like that there is no ending code. Thank you for your patience.

gasvn commented 5 months ago

Based on our observation, the model keeps improve the performance even with a long training interation. So it depends on your gpu resources. For ablation, I suggest you pick a suitable ending points, such 600k, 1000k, or 2000k iterations.