OpenWAM-Stanford commited on
Commit
cb73196
·
verified ·
1 Parent(s): 36d736f

Upload verified LIBERO-CF64k files

Browse files
Files changed (2) hide show
  1. contexts.jsonl +0 -0
  2. dataset.py +178 -0
contexts.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
dataset.py ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Read and verify the common CF64k download using NumPy and the standard library."""
2
+
3
+ import argparse
4
+ from collections import Counter
5
+ import hashlib
6
+ import io
7
+ import json
8
+ from pathlib import Path
9
+ import tarfile
10
+
11
+ import numpy as np
12
+
13
+ VIEWS = ("observation.images.agentview_rgb", "observation.images.eye_in_hand_rgb")
14
+ SAMPLE_FIELDS = (*VIEWS, "future_actions", "observation.state", "frame_index", "timestamp")
15
+ CONTEXT_FIELDS = ("simulator_state",)
16
+
17
+
18
+ def digest(path):
19
+ with path.open("rb") as handle:
20
+ return hashlib.file_digest(handle, "sha256").hexdigest()
21
+
22
+
23
+ def validate_npz(data, *, context):
24
+ with np.load(io.BytesIO(data), allow_pickle=False) as payload:
25
+ fields = CONTEXT_FIELDS if context else SAMPLE_FIELDS
26
+ if set(payload.files) != set(fields) or len(payload.files) != len(fields):
27
+ raise ValueError("Unexpected public NPZ fields")
28
+ for field in fields:
29
+ array = payload[field]
30
+ if array.dtype.hasobject or not np.isfinite(array).all():
31
+ raise ValueError("Nonfinite or object public array")
32
+ if context:
33
+ state = payload["simulator_state"]
34
+ if state.ndim != 1 or state.size < 2 or state.dtype != np.float64:
35
+ raise ValueError("Invalid boundary physics state")
36
+ return {"state_length": state.size}
37
+ shapes = {**{key: ((129, 128, 128, 3), np.uint8) for key in VIEWS},
38
+ "future_actions": ((128, 7), np.float32),
39
+ "observation.state": ((129, 8), np.float32),
40
+ "frame_index": ((129,), np.int64), "timestamp": ((129,), np.float64)}
41
+ for key, (shape, dtype) in shapes.items():
42
+ if payload[key].shape != shape or payload[key].dtype != dtype:
43
+ raise ValueError(f"Invalid public array: {key}")
44
+ if not np.array_equal(payload["frame_index"], np.arange(129)) or not np.array_equal(
45
+ payload["timestamp"], np.arange(129) / 20.0
46
+ ):
47
+ raise ValueError("Invalid public control clock")
48
+ boundary = hashlib.sha256()
49
+ for key in (*VIEWS, "observation.state"):
50
+ boundary.update(payload[key][0].tobytes())
51
+ return {"boundary_sha256": boundary.hexdigest()}
52
+
53
+
54
+ class Dataset:
55
+ def __init__(self, root):
56
+ self.root = Path(root)
57
+ self.metadata = json.loads((self.root / "dataset.json").read_text())
58
+ if set(self.metadata) != {"samples", "contexts", "parent_trajectories", "split",
59
+ "control_frequency_hz", "duration_seconds", "frames", "actions",
60
+ "views", "perturbations", "tasks", "indices"}:
61
+ raise ValueError("Unexpected dataset metadata fields")
62
+ for key, value in {"samples": 64000, "contexts": 10000, "split": "train",
63
+ "control_frequency_hz": 20, "duration_seconds": 6.4,
64
+ "frames": 129, "actions": 128, "views": list(VIEWS)}.items():
65
+ if self.metadata[key] != value:
66
+ raise ValueError(f"Incorrect dataset contract: {key}")
67
+ for name in ("samples.jsonl", "contexts.jsonl", "shards.json"):
68
+ if digest(self.root / name) != self.metadata["indices"][name]:
69
+ raise ValueError("Index checksum mismatch")
70
+ read_rows = lambda name: [json.loads(line) for line in (self.root / name).read_text().splitlines()]
71
+ self.samples = read_rows("samples.jsonl")
72
+ self.contexts = read_rows("contexts.jsonl")
73
+ self.shards = json.loads((self.root / "shards.json").read_text())
74
+ self.entries = {}
75
+ for kind, rows, count, fields in (
76
+ ("samples", self.samples, 64000, {"sample_id", "context_id", "task_id", "perturbation"}),
77
+ ("contexts", self.contexts, 10000, {"context_id", "task_id", "parent_id", "state_length"}),
78
+ ):
79
+ id_key = "sample_id" if kind == "samples" else "context_id"
80
+ if [row[id_key] for row in rows] != list(range(count)):
81
+ raise ValueError("Incomplete or duplicate membership")
82
+ for row in rows:
83
+ if set(row) != fields | {"path", "shard", "sha256", "bytes"}:
84
+ raise ValueError("Unexpected public index fields")
85
+ if row["path"] != f"{kind}/{row[id_key]:05d}.npz":
86
+ raise ValueError("Invalid member path")
87
+ if row["shard"] != f"shards/{kind}-{row[id_key] // 512:05d}.tar":
88
+ raise ValueError("Invalid shard assignment")
89
+ self.entries[row["path"]] = row
90
+ if Counter(row["task_id"] for row in self.samples) != Counter({i: 6400 for i in range(10)}):
91
+ raise ValueError("Invalid task composition")
92
+ if set(row["parent_id"] for row in self.contexts) != set(range(self.metadata["parent_trajectories"])):
93
+ raise ValueError("Invalid parent groups")
94
+ for row in self.samples:
95
+ if type(row["context_id"]) is not int or not 0 <= row["context_id"] < 10000:
96
+ raise ValueError("Invalid context reference")
97
+ context = self.contexts[row["context_id"]]
98
+ if row["task_id"] != context["task_id"]:
99
+ raise ValueError("Sample/context task mismatch")
100
+ if dict(Counter(row["perturbation"] for row in self.samples)) != self.metadata["perturbations"]:
101
+ raise ValueError("Perturbation counts disagree")
102
+ if [row["task_id"] for row in self.metadata["tasks"]] != list(range(10)):
103
+ raise ValueError("Invalid task table")
104
+ for task in self.metadata["tasks"]:
105
+ if set(task) != {"task_id", "prompt", "samples", "simulator_state_length"} or task["samples"] != 6400:
106
+ raise ValueError("Invalid task metadata")
107
+ expected_shards = {row["shard"] for row in self.entries.values()}
108
+ if len(self.shards) != len(expected_shards) or {row["path"] for row in self.shards} != expected_shards:
109
+ raise ValueError("Missing or duplicate shard index")
110
+
111
+ def read(self, sample_id):
112
+ """Read one lossless sequence without extracting a tar to disk."""
113
+ row = self.samples[sample_id]
114
+ return row, self._read_member(row)
115
+
116
+ def read_context(self, context_id):
117
+ row = self.contexts[context_id]
118
+ return row, self._read_member(row)
119
+
120
+ def _read_member(self, row):
121
+ with tarfile.open(self.root / row["shard"]) as archive:
122
+ member = archive.getmember(row["path"])
123
+ if not member.isfile() or member.size != row["bytes"]:
124
+ raise ValueError("Invalid member")
125
+ with archive.extractfile(member) as handle:
126
+ data = handle.read()
127
+ if hashlib.sha256(data).hexdigest() != row["sha256"]:
128
+ raise ValueError("Member checksum mismatch")
129
+ validate_npz(data, context="sample_id" not in row)
130
+ with np.load(io.BytesIO(data), allow_pickle=False) as payload:
131
+ return {key: payload[key] for key in payload.files}
132
+
133
+ def verify(self):
134
+ seen, boundaries = set(), {}
135
+ layouts = {task["task_id"]: task["simulator_state_length"] for task in self.metadata["tasks"]}
136
+ for index, shard in enumerate(self.shards):
137
+ path = self.root / shard["path"]
138
+ if path.stat().st_size != shard["bytes"] or digest(path) != shard["sha256"]:
139
+ raise ValueError("Archive checksum mismatch")
140
+ count = 0
141
+ with tarfile.open(path, "r|") as archive:
142
+ for member in archive:
143
+ if not member.isfile() or member.name not in self.entries or member.name in seen:
144
+ raise ValueError("Unexpected or duplicate member")
145
+ row = self.entries[member.name]
146
+ if row["shard"] != shard["path"] or member.size != row["bytes"]:
147
+ raise ValueError("Wrong member placement or size")
148
+ with archive.extractfile(member) as handle:
149
+ data = handle.read()
150
+ if hashlib.sha256(data).hexdigest() != row["sha256"]:
151
+ raise ValueError("Member checksum mismatch")
152
+ detail = validate_npz(data, context="sample_id" not in row)
153
+ if "sample_id" in row:
154
+ boundary = detail["boundary_sha256"]
155
+ if boundaries.setdefault(row["context_id"], boundary) != boundary:
156
+ raise ValueError("Inconsistent branch boundary")
157
+ elif detail["state_length"] != row["state_length"] or row["state_length"] != layouts[row["task_id"]]:
158
+ raise ValueError("Wrong physics layout")
159
+ seen.add(member.name)
160
+ count += 1
161
+ if count != shard["members"]:
162
+ raise ValueError("Wrong archive member count")
163
+ print(json.dumps({"verified_shards": index + 1, "verified_members": len(seen)}), flush=True)
164
+ if seen != self.entries.keys() or set(boundaries) != set(range(10000)):
165
+ raise ValueError("Incomplete downloaded data")
166
+ return {"status": "verified", "samples": len(self.samples), "contexts": len(self.contexts),
167
+ "shards": len(self.shards), "dataset_sha256": digest(self.root / "dataset.json")}
168
+
169
+
170
+ if __name__ == "__main__":
171
+ parser = argparse.ArgumentParser(description=__doc__)
172
+ parser.add_argument("root", type=Path)
173
+ parser.add_argument("--receipt", type=Path)
174
+ args = parser.parse_args()
175
+ report = Dataset(args.root).verify()
176
+ if args.receipt:
177
+ args.receipt.write_text(json.dumps(report, indent=2) + "\n")
178
+ print(json.dumps(report), flush=True)