File size: 6,741 Bytes
881cb3b
2ad85cb
881cb3b
 
 
2ad85cb
881cb3b
 
65e41ab
 
 
70a5538
 
881cb3b
2ad85cb
881cb3b
 
 
 
2ad85cb
881cb3b
 
 
 
 
d218f6e
881cb3b
 
2ad85cb
881cb3b
2ad85cb
881cb3b
 
 
 
 
d218f6e
881cb3b
70a5538
881cb3b
 
 
70a5538
e776436
881cb3b
 
 
 
70a5538
881cb3b
 
 
 
70a5538
881cb3b
70a5538
881cb3b
 
65e41ab
 
70a5538
 
881cb3b
 
 
 
 
 
 
2ad85cb
881cb3b
2ad85cb
 
65e41ab
 
881cb3b
65e41ab
881cb3b
2ad85cb
 
881cb3b
 
 
 
 
 
d218f6e
881cb3b
2ad85cb
881cb3b
 
 
 
 
d218f6e
881cb3b
 
 
 
 
65e41ab
 
881cb3b
 
 
e776436
 
 
881cb3b
 
65e41ab
2ad85cb
 
65e41ab
 
 
 
 
 
 
 
 
2ad85cb
 
65e41ab
 
 
881cb3b
 
 
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
# DRMHB conversion and evaluation examples

Use Python 3.12 and `vibframe-anndata==0.3.0`. Run from the repository root.
These examples explicitly request complete annotations; the package itself
keeps ground-truth import disabled by default.

```bash
python -m pip install -r examples/requirements.txt
python examples/01_create_featureless_anndata.py data/DRMHB-compact.vibframe.zip derived/DRMHB-compact.v0.3.0.raw.h5ad
python examples/02_inspect_anndata.py derived/DRMHB-compact.v0.3.0.raw.h5ad --waveform-details
python examples/03_add_features.py derived/DRMHB-compact.v0.3.0.raw.h5ad derived/DRMHB-compact.v0.3.0.features.h5ad
```

## 01 — Import once

Accepts a ZIP or extracted VibFrame. The base is featureless: `X` has zero
columns; raw waveforms, spectra, tachometers, identities and metadata remain
available for later features. The default is `float32`, 8 MiB raw blocks,
`scope=all`, strict missing truth and unchanged uncompressed raw HDF5 arrays.

`--ground-truth-scope snapshot` selects the legacy footprint; `none` (or
`--without-ground-truth`) excludes evaluation labels. `--max-sidecar-mib 512`
limits original encoded sidecars, not total memory. `--block-size-mib` and
`--dtype` are explicit. Existing destinations require `--overwrite`; inputs
cannot be used as outputs.

No feature is computed from `trends.parquet` or from construction/diagnostic
labels. The published dataset itself contains no scalar trends.

## 02 — Inspect without loading multi-GiB raw signals

The inspector uses `h5py` group metadata, small `obs`/`var`/package metadata and
public ground-truth accessors. It never calls `anndata.read_h5ad(backed="r")`
as a blanket out-of-core solution: that mode does not make all `obsm`/`uns`
objects lazy. `--waveform-details` expands annotations only. `--verify-hashes`
reads one original annotation file at a time and checks its SHA-256.

## 03 — Add the reference features

The thirteen feature requests remain the same as in the published EDA:
velocity/acceleration RMS, subsynchronous/1X/low-harmonic bands, 1X/2X/BPFO/BPFI
amplitudes, VPF-family RMS, and waveform RMS/crest factor/kurtosis.

For the current 2.6.0 fleet they materialized 468 machine/source-scoped columns.
Each machine populates its relevant descriptors; structural NaNs are not zeros.
The EDA constructs a common logical descriptor matrix without averaging different
machines. Preserve `var` units, processing mode, point and algorithm/options
when aligning columns. Do not match columns by short display name alone.

`--block-rows 256` controls feature block size. `--features-config FILE.yaml`
accepts an alternative package feature configuration. The original VibFrame
is not reopened. Dimension counts and original-annotation manifest identities
are compared before and after feature generation.

## 04 — Upgrade existing H5ADs

```bash
python examples/04_add_complete_ground_truth.py \
  derived/DRMHB-compact.v0.3.0.no-truth.h5ad data/DRMHB-compact.vibframe.zip \
  derived/DRMHB-compact.v0.3.0.enriched.h5ad
```

Uses `add_ground_truth_to_h5ad`. It preserves signals and features and creates a
transactional disk copy. Supply the matching source set and retain `obs.source`,
`obs.machine` and integer `obs.snap_t`. Roughly one extra H5AD of free disk is
needed temporarily. Source-name/path collisions must not be repaired by guessing.
Legacy capture-time companions are not populated retrospectively: reimport raw
signals to obtain those arrays. Per-waveform crop offsets remain available from
original and aligned waveform truth.

## 05 — Read originals and DiagGT

```bash
python examples/05_read_evaluation.py derived/DRMHB-compact.v0.3.0.features.h5ad
python examples/05_read_evaluation.py derived/DRMHB-compact.v0.3.0.features.h5ad \
  --path ground-truth/observations.parquet
python examples/05_read_evaluation.py derived/DRMHB-compact.v0.3.0.features.h5ad \
  --export derived/recovered-evaluation
```

Use the exact path from the printed inventory; the layout is not assumed.
Use `--source SOURCE_ID` when a relative path occurs in several sources.
`read_evaluation_table()` returns Arrow types; `read_evaluation_json()` returns
parsed JSON; `read_evaluation_file()` returns the exact original bytes.
Export refuses overwrites and retains dataset-wide originals, even for a sliced
AnnData. No implicit DiagGT interval-to-snapshot labelling is performed.

## Complete information and model input

`obsm['ground_truth']` and `obsm['waveform_ground_truth']` follow `obs_names`.
The byte archive in `uns['vibframe_evaluation']` is dataset-wide. `snap_t` is the
snapshot key in integer microseconds UTC; raw capture `t` may include a crop
offset. Waveform bindings include source/machine/point and supplied mode identity.
Unknown sidecars/fields are retained instead of silently discarded.

For normality learning, fit scaling, missing-value rules and models only on
calibration. Set thresholds on normal holdout. Evaluate on the future stream
using timestamps. Faults, severities, OOD, nuisance and latent physical state
are targets/audit fields, not automatic inputs. Declare explicitly any use of
regime or RPM. Keep a separately generated controls dataset in a separate H5AD;
controls are in `data/DRMHB-compact-controls.vibframe.zip`, separate from the
current longitudinal ZIP. Use a distinct output basename for their H5AD.

## Resources and verification

No raw compression settings or VibFrame bytes are changed. The current complete 0.3.0 base and feature H5ADs measure 8.169 and
8.265 GiB respectively; both retain 97 original files (26,874,052 bytes).
Verify local derivatives with `python tools/audit_dataset.py --deep`. Metadata decoding also consumes RAM, so the raw block size and
512 MiB encoded-sidecar budget are not process-memory caps.

For a quick end-to-end check using the smaller controls ZIP:

```bash
python examples/01_create_featureless_anndata.py \
  data/DRMHB-compact-controls.vibframe.zip derived/DRMHB-compact-controls.v0.3.0.raw.h5ad
python examples/02_inspect_anndata.py \
  derived/DRMHB-compact-controls.v0.3.0.raw.h5ad --verify-hashes
python examples/03_add_features.py \
  derived/DRMHB-compact-controls.v0.3.0.raw.h5ad \
  derived/DRMHB-compact-controls.v0.3.0.features.h5ad
python examples/05_read_evaluation.py \
  derived/DRMHB-compact-controls.v0.3.0.features.h5ad
```

To exercise example 04, first create an H5AD with `--without-ground-truth`
from that same controls ZIP, then pass it and the ZIP to example 04. These
commands validate the conversion workflow without regenerating the campaign.

Package documentation: https://413hq.github.io/vibframe-anndata-docs/0.3.0/
Ground truth guide: https://413hq.github.io/vibframe-anndata-docs/0.3.0/guides/ground-truth/