Ice9Coffee

I have the same problem. I'm training with the Scribble Dataset. The loss of G_L1 is always 0.00. Fake_B (view in visdom) are first just the input with a gray...

Looking forward to your training dateset & code, as well as the pytorch version.

尝试一下 Quart==0.14.1 Werkzeug==1.0.1

感谢！正式迁移时可作为参考。 service层最近有些改动，同时导致了bug #193 ，可能无法直接合入

Good idea. Daemon is very useful! But for docker, Hoshino need to direct access to custom RESOURCE_DIR, you should also make a link. More tests required. Hoshino is designed for...

I know. But RESOURCE_DIR is custom. It doesn't have to in `~/.hoshino`. You should add more tips in README.md

我觉得docker比较适合作为发行版，或许可以直接放到docker hub上，但缺点在于需要经常去维护。 现在的Dockerfile是将代码和配置文件拷贝到容器中，这样并不方便去修改零碎的config.json配置文件，或许可以考虑开一个新的branch，Dockerfile从这个branch拉取精简版的hoshino（仅包含不需要额外config.json的功能）来实现一行命令直接部署；同时也免去了升级的问题，删掉旧容器重建即可。

config.json是各插件独立的，插件的功能不同其中的结构也不同。比较简单的仅存放一个AUTH_KEY，而复杂的则存了多个列表、字典（如gacha/config.json），我想这不是简单修改环境变量就能调整的。 我的想法有2个，一种办法是提供精简版的docker image，仅保留无外部依赖的功能；另一种办法是将代码目录也进行映射，这样以后维护config.json时才能更加方便。 或许你可以参考NoneBot的文档->[部署](https://nonebot.cqp.moe/advanced/deployment.html)，将酷Q与Hoshino一起部署。

Yes, you delete it indeed, but it is still tracked by git. The size of the repo increased, and someone else could revert this commit.

群中看到了你们的讨论，感觉可行。 提供下代码片段？