关于dpm-solver的问题

[ipwjincheng]

作者你好，我在训练ISIC数据集的时候，训练了2w7个step，使用ddim sample可以得到较好的结果，但是当我使用dpm-solver之后效果会变得很差，请问这是什么原因呢？
dpm-solver的配置是50个step。
期待您的回复

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

Title: Questions about dpm-solver

Hello author, when I trained the ISIC data set, I trained 2w7 steps. I can get better results using ddim sample, but when I use dpm-solver, the effect becomes very poor. What is the reason for this? ?
The configuration of dpm-solver is 50 steps.
Looking forward to your reply

[chenglr3]

@ipwjincheng 你用的是ISIC2016还是ISIC2018，数据量有多大，我训练的时候loss收敛的很快，但是loss_cal下不去，最终使用dpm-solver推理效果还可以，但是使用ddpm推理很差

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

@ipwjincheng Are you using ISIC2016 or ISIC2018? How big is the amount of data? When I trained, the loss converged quickly, but the loss_cal could not go down. In the end, the inference effect using dpm-solver was okay, but the inference using ddpm was very poor.