Update dataset reader and split coverage

README.md

@@ -34,6 +34,23 @@ python read_example.py
 
 训练输入包括雷达 cube、RGB、相机内外参和图像前端的深度、语义特征。监督包括占据层、回波角色、反射面、折回目标、车辆与静态恢复标签。3,486 个训练帧另有六类语义、高度和车道监督，用 `dense_present` 判断。测试输入和评分标签分别存放。
 
+## 训练与评测
+
+下载 [xfield_sim.py](xfield_sim.py)，安装 `numpy zstandard filelock huggingface_hub`。下面的代码取一份已发布的训练样本。读取时只下载对应分片。
+
+```python
+from xfield_sim import XFieldSim
+
+train = XFieldSim("fit")
+key = train.available_keys()[0]
+frame = train[key]
+radar = frame["inputs"]["cube"]
+rgb = frame["inputs"]["image_rgb"]
+labels = frame["targets"]
+```
+
+本地下载完整目录后可使用 `XFieldSim("fit", local_dir="./xfield-sim-core")`。源数据按实际访问口径约 248.48 GB，其中训练约 212.01 GB。下载大小取决于分片压缩。
+
 [当前发布进度](STATUS.md) · [字段与使用](FIELDS.md) · [训练与评测划分](SPLITS.md) · [复现代码](REPRODUCE.md)
 
 仿真采集来源为 [CARLA X-band Radar NLOS](https://huggingface.co/datasets/KAS2003/carla-xband-radar-nlos)。

REPRODUCE.md

@@ -2,7 +2,18 @@
 
 与此划分相配的 XField 代码版本为 [XField_PredCore](https://github.com/kasssssss/XField_PredCore/tree/9edbf4ba909533f8f0df8132726d9811ea0169dd)。仿真 cube+RGB 和 RA+RGB 使用共享的 Stage 1、学习折回、Stage 2 和 Bipolar。
 
-`read_example.py` 可独立运行，无需模型或训练机器上的数据目录。对比方法可以使用相同的 `key`、输入和原生输出接口。复现原训练时使用 `training.rgb_required.batchify` 组批，按代码执行 RGB 归一化。
+`read_example.py` 可独立运行。复现原训练时，在 XField 代码环境中使用下面的入口。它将 RGB 按原训练方式归一化，并保留存在的语义、高度和车道目标。
+
+```python
+from xfield_sim import XFieldSim
+from training.rgb_required import batchify
+
+train = XFieldSim("fit")
+keys = train.available_keys()[:4]
+batch = batchify([train.xfield_sample(k) for k in keys])
+```
+
+锁定测试先用 `XFieldSim("paired_eval")[key]` 产生预测，再用 `test_truth(key)` 读取独立车辆和整场占据标签。
 
 雷达编码器的早期预训练、Fold 初始化、Stage 2 缓存和补全头使用过不同的监督。需要从头复做这些阶段时，可参考[阶段依赖及成员清单](https://huggingface.co/datasets/KAS2003/carla-xband-radar-nlos/blob/385666cfcdc2eb466429760b0783ff711c3fc056/releases/2026-11-03/final/STAGE_DATA.md)。新方法可以直接从本仓的固定训练集开始。
 

STATUS.md

@@ -1,4 +1,11 @@
 # 发布进度
 
-训练示例、固定名单和数据分片分批发布。逐划分可读数量将在对应分片完成公开回读后更新。
+2026-11-03 20:59 UTC+8
 
+| 划分 | 可读帧 | 总帧数 |
+|---|---:|---:|
+| fit | 192 | 29,223 |
+| dev | 1,484 | 1,484 |
+| paired_eval | 3,270 | 3,270 |
+
+训练帧的 RGB／监督和雷达分片同时存在才计为可读。测试真值存放在 [test_labels](test_labels)。

catalog/availability.json

@@ -0,0 +1,24 @@
+{
+  "as_of": "2026-11-03 20:59 UTC+8",
+  "state": "PUBLISHING",
+  "data_revision": "9eebef5702531f2809fb7ac5c9fbe104eb2a149c",
+  "coverage": {
+    "fit": {
+      "planned": 29223,
+      "ready": 192,
+      "shards": 88
+    },
+    "dev": {
+      "planned": 1484,
+      "ready": 1484,
+      "shards": 16
+    },
+    "paired_eval": {
+      "planned": 3270,
+      "ready": 3270,
+      "shards": 69
+    }
+  },
+  "test_labels_present": true,
+  "independent_read_revision": null
+}

catalog/frames.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:462a2132a0e2aba0e88c69178832c6ecea5e2d76958f289272a6da1abce55e68
+size 1736020

xfield_sim.py

@@ -0,0 +1,140 @@
+"""Read XField simulation frames from a downloaded dataset or a pinned public Hub revision."""
+from functools import lru_cache
+import gzip
+import hashlib
+import io
+import json
+import os
+from pathlib import Path
+import tarfile
+import tempfile
+
+import numpy as np
+import zstandard as zstd
+from filelock import FileLock
+from huggingface_hub import HfApi, hf_hub_download
+
+REPO = 'KAS2003/xfield-sim-core'
+SPLITS = ('fit', 'dev', 'paired_eval')
+SENSORS = {'image_rgb', 'img', 'cube', 'depth', 'sem', 'K', 'T', 'steer'}
+
+@lru_cache(maxsize=8)
+def _tar(path, pid):
+    return tarfile.open(path, mode='r:')
+
+class XFieldSim:
+    """New methods get separate input and target dictionaries. The test sensor files contain no GT."""
+
+    def __init__(self, split='fit', *, revision=None, local_dir=None, cache_dir=None):
+        if split not in SPLITS:
+            raise ValueError(f'Unknown split {split!r}')
+        self.split = split
+        self.local_dir = Path(local_dir).resolve() if local_dir else None
+        self.cache = Path(cache_dir or os.environ.get('XFIELD_CACHE', Path.home()/'.cache/xfield-sim'))
+        if self.local_dir is None:
+            self.revision = HfApi(token=False).dataset_info(
+                REPO, revision=revision or 'refs/heads/main', expand=['sha']).sha
+        else:
+            self.revision = revision
+        with gzip.open(self.file('catalog/frames.jsonl.gz'), 'rt') as f:
+            self.rows = [r for line in f if (r := json.loads(line))['split'] == split]
+        with self.file(f'splits/{split}.txt').open() as f:
+            keys = [s.rstrip('\n') for s in f]
+        if keys != [r['key'] for r in self.rows] or len(set(keys)) != len(keys):
+            raise ValueError('Frame index does not match the split')
+        self.by_key = {r['key']:i for i,r in enumerate(self.rows)}
+
+    def __len__(self):
+        return len(self.rows)
+
+    def file(self, name):
+        if self.local_dir:
+            return self.local_dir/name
+        return Path(hf_hub_download(REPO, name, repo_type='dataset', revision=self.revision, token=False))
+
+    def archive(self, name):
+        path = self.file(name)
+        if name.endswith('.tar.zst'):
+            # A single seekable copy per immutable source file, shared by data-loader workers.
+            scope = self.revision or 'download-'+hashlib.sha256(str(self.local_dir).encode()).hexdigest()[:16]
+            unpacked = (self.cache/scope/name).with_suffix('')
+            unpacked.parent.mkdir(parents=True, exist_ok=True)
+            with FileLock(str(unpacked)+'.lock'):
+                if not unpacked.exists():
+                    with tempfile.NamedTemporaryFile(dir=unpacked.parent, delete=False) as out:
+                        tmp = Path(out.name)
+                        try:
+                            with path.open('rb') as stream:
+                                zstd.ZstdDecompressor().copy_stream(stream, out)
+                        except BaseException:
+                            tmp.unlink(missing_ok=True)
+                            raise
+                    tmp.replace(unpacked)
+            path = unpacked
+        return _tar(str(path), os.getpid())
+
+    def read(self, index):
+        row = self.rows[self.by_key[index] if isinstance(index, str) else index]
+        ar = self.archive(row['shard'])
+        with ar.extractfile(row['member']) as member, np.load(io.BytesIO(member.read()), allow_pickle=False) as z:
+            sample = {k:z[k] for k in z.files}
+        if str(sample['key']) != row['key']:
+            raise ValueError('Prepared frame identity mismatch')
+        if self.split == 'fit':
+            with self.archive(row['radar']).extractfile(row['radar_member']) as member:
+                data = zstd.ZstdDecompressor().decompress(member.read())
+            sample['cube'] = np.load(io.BytesIO(data), allow_pickle=False)
+        if 'img' in sample:
+            sample['image_rgb'] = sample.pop('img')
+        cube = sample['cube']
+        if cube.shape != (512,25,64) or cube.dtype != np.complex64 or sample['image_rgb'].shape != (3,360,640):
+            raise ValueError('Invalid radar/RGB dimensions')
+        if self.split == 'fit' and bool(sample['dense_present']) != row['dense']:
+            raise ValueError('Dense supervision flag mismatch')
+        inputs = {k:v for k,v in sample.items() if k in SENSORS}
+        labels = {k:v for k,v in sample.items() if k not in SENSORS|{'key','route'}}
+        if self.split == 'paired_eval' and labels:
+            raise ValueError('Unexpected labels in test sensors')
+        return {'key':row['key'], 'route':str(sample.get('route',row['episode'])), 'inputs':inputs, 'targets':labels}
+
+    __getitem__ = read
+
+    def available_keys(self):
+        """Keys whose required shards exist at the selected revision or in the local download."""
+        if self.local_dir:
+            exists = lambda name: (self.local_dir/name).is_file()
+        else:
+            files = set(HfApi(token=False).list_repo_files(REPO, repo_type='dataset', revision=self.revision))
+            exists = files.__contains__
+        return [r['key'] for r in self.rows if exists(r['shard']) and ('radar' not in r or exists(r['radar']))]
+
+    def xfield_sample(self,index):
+        """A fit/dev item accepted by XField_PredCore training.rgb_required.batchify."""
+        if self.split == 'paired_eval':
+            raise ValueError('Use read() for paired test inference')
+        frame = self.read(index)
+        item = {**frame['inputs'], **frame['targets'], 'key':frame['key'], 'route':frame['route']}
+        rgb = item.pop('image_rgb').astype(np.float32)/255
+        mean=np.array([.485,.456,.406], dtype=np.float32)[:,None,None]
+        std=np.array([.229,.224,.225], dtype=np.float32)[:,None,None]
+        item['image']=(rgb-mean)/std
+        item['dense'] = tuple(item.pop(n) for n in ('dense_sem','dense_height','dense_lane')) if bool(item.pop('dense_present')) else None
+        return item
+
+    def test_truth(self,index):
+        """Read the separately stored targets when scoring frozen test predictions."""
+        if self.split != 'paired_eval':
+            raise ValueError('Test truth is available only for paired_eval')
+        row=self.rows[self.by_key[index] if isinstance(index,str) else index]
+        if not hasattr(self,'_vehicles'):
+            with gzip.open(self.file('test_labels/vehicles.jsonl.gz'),'rt') as f:
+                self._vehicles={r['key']:r for r in map(json.loads,f)}
+        out={'vehicles':self._vehicles[row['key']]}
+        if row['present']:
+            with tarfile.open(self.file('test_labels/scene_occupancy.tar.gz'),'r:gz') as ar:
+                with np.load(io.BytesIO(ar.extractfile(row['key']+'.npz').read()),allow_pickle=False) as z:
+                    if str(z['key'])!=row['key']:raise ValueError('Test label identity mismatch')
+                    shape=tuple(z['shape']);n=int(np.prod(shape))
+                    out['occupancy']={k:np.unpackbits(z[k])[:n].reshape(shape).astype(bool) for k in ('obs','any')}
+        return out
+