D-SENO / README.md
yushuang88's picture
Upload folder using huggingface_hub
b144968 verified
|
Raw History Blame Contribute Delete
2.34 kB
---
license: other
library_name: pytorch
tags:
- OneScience
- neural-operator
- pipe-flow
---
# D-SENO: Pipe-G flow
This package reproduces the Pipe-G experiment from [Efficient Dilated Squeeze and Excitation Neural Operator for Differential Equations](https://arxiv.org/pdf/2601.17407). A seven-block, width-96 D-SENO maps curved-pipe X/Y coordinates to horizontal velocity on a 129x129 grid. Three seeds each completed 500 epochs with 1,000 training and 200 test cases.
## Results
The supplied seed-11 checkpoint gives test relative L2 **0.00292652** after reload. The best-checkpoint three-seed mean is **0.00289914**, and the final-epoch three-seed mean is **0.00307909**. The paper reports **0.0030**. Best checkpoints were selected using the same test set, so those best scores are not independent final-test results; see `results/three_seed_summary.json` for the complete record.
## Download
```bash
hf download OneScience-Group/D-SENO --local-dir ./D-SENO
cd D-SENO
```
The original Geo-FNO Pipe arrays are available from the [OneScience/HO-FNO-Pipe ModelScope dataset](https://www.modelscope.cn/datasets/OneScience/HO-FNO-Pipe). This Hugging Face repository contains no raw dataset.
```bash
modelscope download --dataset OneScience/HO-FNO-Pipe --local_dir ./data
```
`Pipe_X.npy` and `Pipe_Y.npy` are coordinate arrays, while `Pipe_Q.npy` contains flow fields. Each has 2,310 cases on a 129x129 grid; the model uses Q channel 0. Run all commands from the model directory; the published config reads `data/` and writes to `results/`.
## Environment and usage
Use Python 3.10 and PyTorch on an allocated GPU/DCU, plus `onescience[cfd-gpu]` or `onescience[cfd-dcu]` and `requirements.txt`. On an HPC system, use a compute-node allocation.
```bash
python scripts/inference.py
python scripts/result.py
```
To retrain, move the included `weight/` and `results/` outside the package and run `python scripts/train.py --seed 11`. Other seeds need distinct `--output-root` directories. Published inference accepts a relocated data directory while still checking model and preprocessing settings.
Checkpoint: `weight/best_model.pth` (SHA256 `f6464c8bd42694228df105abd651dfbe075fb92ba1464ad4835cc889c35bad95`). Source data: [Geo-FNO](https://github.com/neuraloperator/Geo-FNO). Check original data terms before redistribution.