luo-yining / cfdbench Goto Github PK

Dear Yining Luo and Yingfa Chen:
Thanks for your excellent codes shared with us, but I had some questions about it. That is below:
中文版：
在CFDBench-main/src/models/pinn.py文件中
from .base_model import CfdNN
但是没有CfdNN这个函数或模块
English Version:
In the CFDBench-main/src/models/pinn.py modules
from .base_model import CfdNN
I didn't find out the module named CfdNN

大佬您好：
我训练时候的src/args.py中的mode: str = "train-test"，能训练并预测模型，这不会出错。但当我将src/args.py中的mode: str = "test"，我于src.utils.py中的get_best_ckpt()中ckpt_dirs = sorted(output_dir.glob("ckpt-*"))断点发现，ckpt_dirs返回空字符串，我搞不清楚为什么会这样，希望大佬能解惑。
非常感谢！

另外，我做了一张该代码的uml图，如有不正确的地方，希望各位指正~
https://drive.google.com/file/d/1jc0of_o-XIJvoqIjQMnoM6aUn5m3bm6j/view?usp=sharing

Please ask big brother, the code used in the environment version can be marked, I python3.12 and torch2.1 when running many versions conflict

In dam's packet, only subfolders such as case001 are entered, and there are no subsets of bc, geo, and prop after opening, but a series of folders at the beginning of case, and the code cannot be executed against dam, the error is as follows:
python train_auto.py --model fno --data_name dam_case
in get_dam_auto_datasets
assert case_dirs
AssertionError

Originally posted by @justgoinggoxixi in #1 (comment)

在24行
assert problem_name in ["cylinder", "tube", "dam", "cylinder"]
以及34行
elif problem_name == "cylinder": # data_dir = Path("data", data_name) train_data, dev_data, test_data = get_cavity_datasets( data_dir / problem_name, case_name=subset_name, norm_props=norm_props, norm_bc=norm_bc, ) return train_data, dev_data, test_data
还有94行
elif problem_name == "cylinder": train_data, dev_data, test_data = get_cavity_auto_datasets( data_dir / problem_name, case_name=subset_name, norm_props=norm_props, norm_bc=norm_bc, delta_time=delta_time, ) return train_data, dev_data, test_data
这里的problem_name与下方的get_cavity_auto_datasets不符，也许需要改成"cavity"

https://github.com/luo-yining/CFDBench/blob/88ec058ac6e9b1003ba4fb46ea13654888bc538c/src/models/fno/fno2d.py#L168C10-L168C10

When inputting, U (u, v) is trained as two channels by the network, but in the output image, only the velocity field u is present without v. As shown in the train auto code: inputloss: dict=model. loss_fn（
Labels=labels [:,: 1], preds=inputs [:,: 1], the author only uses the output of the first channel for loss, so why do we need to input the v velocity field?