search

GuHuangAI / DiffusionEdge

Code for AAAI 2024 paper: "DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection"
Apache License 2.0
236 stars 25 forks source link

About train_num_steps #7

Closed iris0329 closed 9 months ago

iris0329 commented 9 months ago

Hi, Thanks a lot for the contribution of this code base; it makes it much easier to read and understand the pipeline of LDM. In the VAE training stage, I found you set the train_num_steps to 150000. Could you give me some idea of how you decided on this number? If I use my own dataset, how can I adapt this number?

Thanks in advance. Best wishes

GuHuangAI commented 9 months ago

@iris0329 For every 10000 steps, we test the performance on training dataset and save the checkpoint. Therefore, we can evaluate approximately the model by the sampling results. I suggest that you set the train_num_steps to 150000 as default, since it is enough for general reconstruction tasks. In addition, since the model uses the EMA strategy, the model of the subsequent iter usually performs better. Hope it can help you.

iris0329 commented 9 months ago

Ah, I see Thanks a lot for your reply, @GuHuangAI. Let me have a try

Update: it is truly enough to get a good reconstruction image on my dataset