Datasets:

Modalities:
Audio
Video
Size:
> 1T
Libraries:
Datasets
License:
File size: 12,702 Bytes
60be9be
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
"""Frame-independent log-u16, row delta, byte shuffle and Zstd depth codec.

The .dlog file concatenates independent Zstd frames. Its .index.json sidecar
contains byte ranges, dimensions and quantization parameters. Physical units
remain centimeters. Codes 1..65535 span [near_cm, far_cm]; 0 decodes to zero.
Nonzero values outside that range, signed zero and nonfinite values are stored
as exact float32 exceptions, so they are never silently clipped.
"""
from __future__ import annotations

import hashlib
import json
import math
from pathlib import Path

import numpy as np
import zstandard as zstd

MAGIC = b"HELIOS_DLOG_V1\n"
SCHEMA = "helios-depth-log-u16-v1"


def parameters(near_cm=10.0, far_cm=100_000_000.0):
    if not (math.isfinite(near_cm) and math.isfinite(far_cm) and 0 < near_cm < far_cm):
        raise ValueError("Require finite 0 < near_cm < far_cm")
    return {"schema": SCHEMA, "unit": "centimeter", "near_cm": float(near_cm),
            "far_cm": float(far_cm), "quantization": "log-nearest-u16",
            "valid_codes": [1, 65535], "zero_code": 0,
            "delta": "uint16-modular-horizontal-reset-every-row",
            "shuffle": "two-byte-planes-per-frame", "compression": "zstd",
            "compression_level": 3, "zstd_checksum": True,
            "exceptions": "uint32-pixel-indices-followed-by-float32le-values",
            "max_relative_quantization_error": math.expm1(math.log(far_cm/near_cm)/(2*65534))}


def encode_frame(depth, config, compressor=None):
    x = np.asarray(depth, dtype='<f4')
    if x.ndim != 2 or x.size > np.iinfo(np.uint32).max:
        raise ValueError("Expected a 2D frame with uint32-addressable pixels")
    lo, hi = config['near_cm'], config['far_cm']
    valid = np.isfinite(x) & (x >= lo) & (x <= hi)
    # Retain all exceptional nonzero values and the sign bit of negative zero.
    exceptional = ~valid & ((x != 0) | np.signbit(x))
    indices = np.flatnonzero(exceptional).astype('<u4')
    values = x.ravel()[indices].astype('<f4', copy=False)
    codes = np.zeros(x.shape, dtype='<u2')
    codes[valid] = (1 + np.rint(65534 * np.log(x[valid].astype(np.float64)/lo)
                              / math.log(hi/lo))).astype(np.uint16)
    delta = codes.copy()
    delta[:, 1:] = codes[:, 1:] - codes[:, :-1]
    shuffled = delta.view(np.uint8).reshape(-1, 2).T.copy().tobytes()
    raw = shuffled + indices.tobytes() + values.tobytes()
    compressor = compressor or zstd.ZstdCompressor(level=3, write_checksum=True)
    blob = compressor.compress(raw)
    entry = {"length": len(blob), "uncompressed_bytes": len(raw),
             "exception_count": len(indices)}
    return blob, entry, codes


def decode_codes(blob, entry, height, width):
    raw = zstd.ZstdDecompressor().decompress(blob, max_output_size=entry['uncompressed_bytes'])
    n, count = height*width, entry['exception_count']
    if len(raw) != 2*n + 8*count or len(raw) != entry['uncompressed_bytes']:
        raise ValueError("Invalid decompressed frame size")
    delta = np.frombuffer(raw, dtype=np.uint8, count=2*n).reshape(2, n).T.copy().view('<u2')
    codes = np.cumsum(delta.reshape(height, width), axis=1, dtype=np.uint16)
    indices = np.frombuffer(raw, dtype='<u4', count=count, offset=2*n)
    values = np.frombuffer(raw, dtype='<f4', count=count, offset=2*n+4*count)
    if count and (indices[-1] >= n or np.any(indices[1:] <= indices[:-1])):
        raise ValueError("Invalid exception indices")
    return codes, indices, values


def codes_to_depth(codes, indices, values, config):
    out = np.zeros(codes.shape, dtype='<f4')
    valid = codes > 0
    out[valid] = config['near_cm'] * np.exp((codes[valid].astype(np.float64)-1)
                                          * math.log(config['far_cm']/config['near_cm']) / 65534)
    out.ravel()[indices] = values
    return out


def decode_frame(blob, entry, height, width, config):
    return codes_to_depth(*decode_codes(blob, entry, height, width), config)


def compress_hdf5(source, output, config, source_uri=None, source_info=None):
    import h5py
    source, output = Path(source), Path(output)
    output.parent.mkdir(parents=True, exist_ok=True)
    index = {"codec": config, "source_uri": source_uri or str(source),
             "source_info": source_info or {}, "source_bytes": source.stat().st_size,
             "source_sha256": file_sha256(source), "frames": []}
    compressor = zstd.ZstdCompressor(level=3, write_checksum=True)
    max_relative = 0.0
    special_count = 0
    with h5py.File(source, 'r') as f, output.open('wb') as out:
        d = f['depth']
        if d.ndim != 3 or d.dtype.kind != 'f' or d.dtype.itemsize != 4:
            raise ValueError(f"Expected (frames, height, width) float32, got {d.shape} {d.dtype}")
        encoding = d.attrs.get('depth_encoding', '')
        unit = d.attrs.get('unit', '')
        if isinstance(encoding, bytes): encoding = encoding.decode()
        if isinstance(unit, bytes): unit = unit.decode()
        if encoding != 'linear_scene_depth_cm' or unit != 'centimeter':
            raise ValueError(f"Unsupported source depth encoding/unit: {encoding!r}/{unit!r}")
        count, height, width = d.shape
        if count < 1: raise ValueError('Cannot recompress an empty depth dataset')
        index.update(frame_count=count, height=height, width=width,
                     original_attributes={k: str(v) for k,v in d.attrs.items()})
        out.write(MAGIC)
        for i in range(count):
            x = np.asarray(d[i], dtype='<f4')
            blob, entry, codes = encode_frame(x, config, compressor)
            # Verify every compressed frame, including exact delta inversion.
            recovered, indices, values = decode_codes(blob, entry, height, width)
            if not np.array_equal(codes, recovered):
                raise ValueError(f"Integer codec roundtrip failed at frame {i}")
            y = codes_to_depth(recovered, indices, values, config)
            if len(indices) and not np.array_equal(x.ravel()[indices].view('<u4'),
                                                  y.ravel()[indices].view('<u4')):
                raise ValueError(f"Exception preservation failed at frame {i}")
            valid = codes > 0
            if valid.any():
                err = float(np.max(np.abs(y[valid].astype(np.float64)/x[valid]-1)))
                if err > config['max_relative_quantization_error'] + 2e-7:
                    raise ValueError(f"Quantization bound exceeded at frame {i}: {err}")
                max_relative = max(max_relative, err)
            entry.update(frame_index=i, offset=out.tell())
            out.write(blob)
            index['frames'].append(entry)
            special_count += entry['exception_count']
    index.update(compressed_bytes=output.stat().st_size, compressed_sha256=file_sha256(output),
                 verified=True, max_observed_relative_error=max_relative,
                 exception_pixel_count=special_count,
                 original_raw_bytes=count*height*width*4)
    sidecar = Path(str(output)+'.index.json')
    sidecar.write_text(json.dumps(index, ensure_ascii=False, separators=(',', ':')))
    return index, sidecar


def file_sha256(path):
    h = hashlib.sha256()
    with Path(path).open('rb') as f:
        for block in iter(lambda: f.read(8*1024*1024), b''): h.update(block)
    return h.hexdigest()


def pack_into_hdf5(source, output, config, source_uri=None, source_info=None):
    """Keep a genuine .h5 container and original filename for in-place updates.

    A contiguous byte dataset carries independent Zstd frames. The embedded
    index makes the file self describing; the sidecar enables remote ranges.
    Direct f['depth'] float32 access must be replaced with DepthLogReader.
    """
    import h5py
    import tempfile
    output = Path(output)
    output.parent.mkdir(parents=True,exist_ok=True)
    with tempfile.TemporaryDirectory(prefix='depth-pack-', dir=output.parent) as tmp:
        packed = Path(tmp)/'depth.dlog'
        index, _ = compress_hdf5(source, packed, config, source_uri, source_info)
        index.pop('compressed_sha256')
        index.pop('compressed_bytes')
        index['container'] = 'hdf5'
        index['payload_dataset'] = 'depth_payload_zstd'
        with h5py.File(output, 'w', libver='earliest') as f:
            payload = f.create_dataset('depth_payload_zstd', shape=(packed.stat().st_size,),
                                       dtype=np.uint8, chunks=None)
            with packed.open('rb') as raw:
                offset = 0
                for chunk in iter(lambda: raw.read(8*1024*1024), b''):
                    payload[offset:offset+len(chunk)] = np.frombuffer(chunk, dtype=np.uint8)
                    offset += len(chunk)
            f.flush()
            base = payload.id.get_offset()
            if base is None: raise ValueError('Payload dataset is not contiguous')
            index['payload_file_offset'] = base
            for frame in index['frames']: frame['offset'] += base
            f.attrs['depth_codec_schema'] = SCHEMA
            f.attrs['depth_encoding'] = 'log_u16'
            f.attrs['decoded_depth_encoding'] = 'linear_scene_depth_cm'
            f.attrs['unit'] = 'centimeter'
            f.attrs['depth_log_index_json'] = json.dumps(index, ensure_ascii=False, separators=(',', ':'))
            f.attrs['reader'] = 'tools.offload_data.depth_log_codec.DepthLogReader'
        # Re-read every frame through the final HDF5 byte ranges before publishing.
        reader = DepthLogReader(output)
        with packed.open('rb') as original:
            for i, entry in enumerate(index['frames']):
                original.seek(entry['offset']-base)
                expected = decode_codes(original.read(entry['length']),entry,index['height'],index['width'])
                actual = reader.read_codes(i)
                for a,b in zip(actual,expected):
                    if not np.array_equal(a.view(np.uint8),b.view(np.uint8)):
                        raise ValueError(f'HDF5 container roundtrip failed at frame {i}')
    index['compressed_bytes'] = output.stat().st_size
    index['compressed_sha256'] = file_sha256(output)
    sidecar = Path(str(output)+'.index.json')
    sidecar.write_text(json.dumps(index,ensure_ascii=False,separators=(',',':')))
    return index,sidecar


def embedded_hdf5_index(path):
    import h5py
    with h5py.File(path,'r') as f:
        if f.attrs.get('depth_codec_schema') != SCHEMA: return None
        return json.loads(f.attrs['depth_log_index_json'])


class DepthLogReader:
    """Read a local packed frame, or use fetch_range(offset, length) remotely.

    read_codes() retains uint16 for a compact cache. read() converts one frame
    to float32 centimeters. RGB, poses and actions remain at source_uri.
    """
    def __init__(self, path=None, index=None, fetch_range=None):
        self.path = Path(path) if path is not None else None
        self.index = index
        if self.index is None:
            with self.path.open('rb') as f: signature = f.read(8)
            if signature == b'\x89HDF\r\n\x1a\n':
                self.index = embedded_hdf5_index(self.path)
                if self.index is None: raise ValueError('HDF5 does not use the log depth codec')
            else:
                self.index = json.loads(Path(str(path)+'.index.json').read_text())
        if self.index['codec']['schema'] != SCHEMA: raise ValueError("Unsupported codec")
        self.fetch_range = fetch_range

    def read_codes(self, frame_index):
        if not 0 <= frame_index < self.index['frame_count']: raise IndexError(frame_index)
        entry = self.index['frames'][frame_index]
        if self.fetch_range is not None:
            blob = self.fetch_range(entry['offset'], entry['length'])
        else:
            with self.path.open('rb') as f:
                f.seek(self.index.get('payload_file_offset',0))
                if f.read(len(MAGIC)) != MAGIC: raise ValueError("Invalid depth file header")
                f.seek(entry['offset'])
                blob = f.read(entry['length'])
        if len(blob) != entry['length']: raise ValueError("Truncated depth frame")
        return decode_codes(blob, entry, self.index['height'], self.index['width'])

    def read(self, frame_index):
        return codes_to_depth(*self.read_codes(frame_index), self.index['codec'])


def read_depth_frame(path, frame_index):
    """Read either an original float32 HDF5 or a recompressed HDF5 during migration."""
    import h5py
    with h5py.File(path,'r') as f:
        if f.attrs.get('depth_codec_schema') != SCHEMA:
            return f['depth'][frame_index]
    return DepthLogReader(path).read(frame_index)