FuXi-Ocean

KO | EN

gitlite — search

#javascript #ai #typescript #llm #golang #go #android #linux #docker #windows #kubernetes #vue #awesome-list #ios

(★ 25)

Ocean forecasting and RMSE evaluation scripts.

FuXi-Ocean 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# FuXi-Ocean

Official open-source inference code for **FuXi-Ocean**.

FuXi-Ocean has been accepted as an Oral presentation at NeurIPS 2025. See the NeurIPS virtual page:
https://neurips.cc/virtual/2025/loc/san-diego/poster/117603

This repository provides ONNX-based ocean forecasting inference and RMSE evaluation utilities for HYCOM-format ocean data.

## Release Plan

- [x] Inference code
- [x] Test code
- [ ] ONNX weight files

## Repository Layout

```text
.
├── inference_hycom_aim_full.py  # ONNX inference script
├── infer_hycom_full.sh          # Example inference launcher
├── eval_rmse.py                 # RMSE evaluation script
├── run_eval_rmse.sh             # Example evaluation launcher
├── re.txt                       # Python dependencies
└── README.md                    # Project documentation
```

## Environment

Install the Python dependencies:

```bash
pip install -r re.txt
```

Recommended runtime requirements:

- Python 3.8+
- 32+ CPU cores for CPU inference
- 128 GB+ memory for full-resolution HYCOM fields

## Inference

`inference_hycom_aim_full.py` runs autoregressive ONNX inference on normalized HYCOM ocean data.

### Arguments

| Argument | Type | Default | Description |
| --- | --- | --- | --- |
| `--input` | str | required | Input Zarr dataset path |
| `--model` | str | required | ONNX model path |
| `--mask` | str | required | Land-mask NetCDF path |
| `--output` | str | required | Output directory |
| `--device` | str | `cuda` | ONNX Runtime execution provider preference: `cuda` or `cpu` |
| `--total_step` | int | `40` | Number of autoregressive forecast steps |
| `--hour_interval` | int | `6` | Forecast interval in hours |
| `--start_time` | str | `20150101` | Inclusive start time |
| `--end_time` | str | `20151230` | Inclusive end time |
| `--stride` | int | `1` | Stride over candidate initialization windows |

### Input Data

The expected input is a normalized Zarr dataset with dimensions `(time, channel, lat, lon)`.

Example structure:

```text
hycom.6h.2006_2015.2160_4320.c105/
├── .zarray
├── .zattrs
├── channel/
├── lat/
├── lon/
├── time/
├── data/
├── mean.nc
├── std.nc
└── land_mask.nc
```

Expected data shape:

```python
<xarray.Dataset>
Dimensions:  (time: 14044, channel: 105, lat: 2160, lon: 4320)
Coordinates:
  * time     (time) datetime64[ns] 2006-01-01T12:00:00 ... 2015-12-31
  * channel  (channel) <U6 's0' 's2' 's4' ... 'sv1500' 'sel'
  * lat      (lat) float32 90.0 89.92 ... -89.83 -89.92
  * lon      (lon) float32 0.0 0.0833 ... 359.833 359.917
Data variables:
    data     (time, channel, lat, lon) float16
```

### Run Inference

Use the launcher script:

```bash
bash infer_hycom_full.sh
```

Or run the Python script directly:

```bash
python inference_hycom_aim_full.py \
    --input /path/to/hycom/data \
    --model /path/to/ocean.onnx \
    --mask /path/to/land_mask.nc \
    --output /path/to/output \
    --total_step 40 \
    --device cpu
```

### Output Format

Inference outputs are written by initialization time:

```text
output/
├── 20150107T00/
│   ├── step_001.nc
│   ├── step_002.nc
│   └── ...
├── 20150107T12/
│   ├── step_001.nc
│   └── ...
└── ...
```

Each prediction file stores a full-resolution forecast field:

```python
<xarray.DataArray 'pred' (channel: 105, lat: 2160, lon: 4320)>
Coordinates:
  * channel  (channel) <U6 's0' 's2' ... 'sel'
  * lat      (lat) float32 90.0 ... -89.92
  * lon      (lon) float32 0.0 ... 359.917
```

## Evaluation

`eval_rmse.py` computes weighted RMSE between predictions and ground truth.

### Arguments

| Argument | Type | Default | Description |
| --- | --- | --- | --- |
| `--pred_root` | str | required | Prediction root directory |
| `--gt_path` | str | required | Ground-truth Zarr dataset path |
| `--mask_path` | str | required | Land-mask NetCDF path |
| `--weight_path` | str | required | Spatial weight NetCDF path |
| `--output` | str | required | RMSE result output directory |
| `--hour_interval` | int | `6` | Forecast interval in hours |

### Run Evaluation

Use the launcher script:

```bash
bash run_eval_rmse.sh
```

Or run the Python script directly:

```bash
python eval_rmse.py \
    --pred_root /path/to/predictions \
    --gt_path /path/to/hycom/data \
    --mask_path /path/to/land_mask.nc \
    --weight_path /path/to/weight.nc \
    --output /path/to/rmse_results \
    --hour_interval 6
```

The main output is `rmse_results.csv`, a wide-format table where each row is a forecast step and each channel is a metric column.

## RMSE Definition

For each channel:

```text
RMSE_channel = sqrt(
    sum((pred - gt)^2 * weight * mask) / sum(mask)
)
```

Where:

- `pred`: predicted field with shape `(C, H, W)`
- `gt`: ground-truth field with shape `(C, H, W)`
- `mask`: ocean mask, where ocean is `1` and land is `0`
- `weight`: latitude-based spatial weight, usually `cos(lat)`

## Notes

- ONNX weight files are not included yet and are listed in the release plan.
- The example shell scripts contain site-specific paths. Update them before running on a new system.

FuXi-Ocean

FuXi-Ocean 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

FuXi-Ocean

FuXi-Ocean 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README