FaceSpace dataset training takes too long

[Volleria]

Hello, dear author of mofanerf:
your paper mentions that it takes about 2 days for training on dual NVIDIA GTX3090 GPUS , however , after l adjusting the chunk、netchunk and N-rand to the recommended minimum, I found that I still need to train nearly 7700 hours on a single NVIDIA GTX 2080ti，are there any configuration errors, how can l reduce training time？

[yiyuzhuang]

Hi! Could you show me your config file?

[Volleria]

Hi! Could you show me your config file?

[gongwkang]

你好，请问如下报错是因为我数据集下载少了还是代码本身问题？

[Volleria]

你好，请问如下报错是因为我数据集下载少了还是代码本身问题？

应该是代码的问题吧， tranforms_train_tranforms_all_16.json.json 这个文件名肯定是有问题的，看起来应该是 transforms_all_16.json 或者是 transforms_train_16.json

[yiyuzhuang]

你好！ 请问你数据的组织格式是对的吗？另外这个合理的返回结果应该是"basedir/transfroms_{}_{}.json"，感觉是其中的"id"的值出错了？

…

------------------ Original ------------------ From: "zhuhao-nju/mofanerf" ***@***.***>; Date: Sat, Apr 8, 2023 03:39 PM ***@***.***>; ***@***.******@***.***>; Subject: Re: [zhuhao-nju/mofanerf] FaceSpace dataset training takes too long (Issue #12) 你好，请问如下报错是因为我数据集下载少了还是代码本身问题？ 应该是代码的问题吧， tranforms_train_tranforms_all_16.json.json 这个文件名肯定是有问题的，看起来应该是 transforms_all_16.json 或者是 transforms_train_16.json — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[yiyuzhuang]

Hi! Could you show me your config file?

I have reviewed the config file and have noticed that the settings for "netchunk" and "chunk" may potentially slow down the performance of the program. It may be worth considering maximizing these settings based on your machine specifications.

[gongwkang]

作者您好，依然是训练时间过长的问题。我用24G的4卡titan RTX以如下配置训练，依然需要将近2000h
，这与您的训练时间相差过大，请帮我看看是否存在配置问题。

[gongwkang]

Hi! Could you show me your config file?

请问你通过增加chunk和net_chunk有明显提升训练速度么

[Volleria]

Hi! Could you show me your config file?

请问你通过增加chunk和net_chunk有明显提升训练速度么

还没有试，实验室资源比较紧张，我现在在跑别的东西

[yiyuzhuang]

作者您好，依然是训练时间过长的问题。我用24G的4卡titan RTX以如下配置训练，依然需要将近2000h ，这与您的训练时间相差过大，请帮我看看是否存在配置问题。

你好，我注意到“i_testset=1，i_weights=1”设置了“验证指标，保存模型的间隔为1”，这两个应该设置大一些，如100,000。每次迭代都渲染一张完整的图像会占用大量运算时间，而和chunk的影响可能不大。