GeoVectorMatrix
commented
7 months ago For the DDPM part, the batch size depends on the GPU memory. It might be set to 10 for the 3090 (refer to the GPU monitor). A training session may take several days (e.g., >=6 days, >=6000 epochs with one 3090). We haven't extensively explored this part. Empirically, one may decide whether to stop training by inspecting the intermediate images generated during the training process.
我使用MSRS数据集，train_ddpm.json训练新的ddpm.pth，然而似乎效果不好，loss函数是这样的：
请问训练时的超参数就是按照train_ddpm.json中设置的那样吗？大概需要训练多久呢，几天？