Open RunpuWei opened 11 months ago
你好,你的图像生成的很清晰,但我生成的却是以下两个样子的,一个是用savedmodel000000.pt采样的,一个是emasavedmodel_0.9999_000000.pt采用,想知道这是不是训练不够导致的,因为代码中作者在训练中没有给截止条件,这很迷惑
Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑🤝🧑👫🧑🏿🤝🧑🏻👩🏾🤝👨🏿👬🏿
Hello, your image is very clear, but what I generated is the following two. One is sampled with savedmodel000000.pt, and the other is sampled with emasavedmodel_0.9999_000000.pt. I want to know if this is caused by insufficient training. , because the author in the code did not give a cut-off condition during training, which is very confusing.
I trained the model on the ISIC dataset with the default experimental settings provided in the readme and sampled the same image using the weight files of 15,000, 65,000 and 100,000 iterations. Finally, I got the following three segmentation prediction images.
Obviously, it can be seen that the effect of 65,000 iterations is better than 15,000 iterations, but the result obtained by 100,000 iterations is a pure noise image. I wonder if you have encountered this problem, I suspect that diffusion also has a problem with over-fitting and I would like to ask if there is any way to mitigate this problem. Hope to get your answer, thank you!
15,000 iter
65,000 iter
100,000 iter