Technology Sharing

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

I have been studying Vincent graphs for some time, and have reproduced them in quite a few papers. This blog mainly records the pitfalls I encountered.

At present, the projects that implement Wenshengtu are mainly divided into two categories:

1. Implementation based on the original project file of Stable-diffusion

Original project address:https://github.com/Stability-AI/stablediffusion

Some of the paper's codes are developed based on this original project. It is unlikely that we will create a virtual environment for each project, so we will inevitably run multiple projects on the same environment. However, Python's environment is relatively not very friendly, and different versions of the same package may have differences. Therefore, some problems often occur.

Take my previous reproduction of Vico as an example. When I trained the model for inference, it gave me an error:

It said it couldn't download the model online, but I had clearly configured it to a local path in the project code, so it was impossible for this problem to occur again. Then I looked at its error backtrace, and I found that it actually ran the Dreambooth code in the same folder when it was running. At that time, I thought it was because the files had the same name, and the system automatically read them sequentially when obtaining the dependent path, so I printed the system path:

After I deleted everything related to Dreambooth, it still didn't work. I thought it must be an environment problem. Later I found out that it was because the local project ldm was not installed, and the environment was still using the previously installed ldm. In fact, the solution is also very simple, just execute pip install -e .