README.md

metadata

dataset_info:
  features:
    - name: sample_id
      dtype: string
    - name: energy
      sequence: float32
      length: 3384
    - name: n_freqs
      dtype: int32
    - name: n_dirs
      dtype: int32
    - name: source
      dtype: string
    - name: station
      dtype: string
    - name: n_anchors
      dtype: int32
    - name: anchors_json
      dtype: string
    - name: Hs
      dtype: float32
    - name: Tp
      dtype: float32
    - name: Dp
      dtype: float32
    - name: total_energy
      dtype: float32
  splits:
    - name: train
      num_examples: 37412
    - name: validation
      num_examples: 15625
    - name: test
      num_examples: 15411
license: cc-by-4.0
task_categories:
  - image-to-image
tags:
  - oceanography
  - wave-spectrum
  - compression
  - diffusion-model

ATLAS-WDS: Wave Directional Spectrum Dataset

海浪方向谱压缩回传训练数据集。

数据格式

每条记录包含一个 47×72 能量矩阵（展平为 3384 维 float32 数组） 及对应的 斜高斯锚点参数。

快速加载

from datasets import load_dataset
import numpy as np, json

ds = load_dataset("wuff-mann/ATLAS-WDS", split="train", streaming=True)

for sample in ds:
    # 还原能量矩阵
    energy = np.array(sample["energy"], dtype=np.float32).reshape(
        sample["n_freqs"], sample["n_dirs"])  # (47, 72)
    # 锚点参数
    anchors = json.loads(sample["anchors_json"])
    # 物理参数
    Hs, Tp, Dp = sample["Hs"], sample["Tp"], sample["Dp"]

三阶段训练使用

# Stage 1: cLDM 预训练 — 只用能量矩阵
for sample in ds:
    matrix = np.array(sample["energy"]).reshape(47, 72)

# Stage 2: Swin 编码器 — 矩阵 + 锚点
for sample in ds:
    matrix = np.array(sample["energy"]).reshape(47, 72)
    anchors = json.loads(sample["anchors_json"])

# Stage 3: 端到端对齐 — 仅真实数据
ds_real = ds.filter(lambda x: x["source"] != "synthetic")