probability-flow 1.4.4__tar.gz → 1.5.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {probability_flow-1.4.4 → probability_flow-1.5.0}/CHANGELOG.md +105 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/PKG-INFO +1 -1
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/__init__.py +1 -1
- probability_flow-1.5.0/probability_flow/generation.py +1637 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/serialize.py +54 -7
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/variable.py +42 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/viz.py +113 -2
- {probability_flow-1.4.4 → probability_flow-1.5.0}/pyproject.toml +1 -1
- probability_flow-1.4.4/probability_flow/generation.py +0 -906
- {probability_flow-1.4.4 → probability_flow-1.5.0}/.gitignore +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/LICENSE +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/README.md +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/cpd.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/decompose.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/edges.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/eliminate.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/engines.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/evidence.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/factors.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/graph.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/independent.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/inference.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/junction.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/logical.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/noisy.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/ordering.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/py.typed +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/query.py +0 -0
- {probability_flow-1.4.4 → probability_flow-1.5.0}/probability_flow/reveal.py +0 -0
|
@@ -1,5 +1,110 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## 1.5.0
|
|
4
|
+
|
|
5
|
+
### Changed, and this BREAKS SEED REPRODUCTION
|
|
6
|
+
|
|
7
|
+
- **A gated weight is now drawn SPIKE-AND-SLAB, and the default makes every gate a
|
|
8
|
+
true undercut.** `p_pure` (default **1.0**) is the spike: with that probability
|
|
9
|
+
`weight_when_gated` is exactly 0.0, so the premise stops carrying.
|
|
10
|
+
Otherwise the slab, `w * U(*slab_fraction_range)` with
|
|
11
|
+
`slab_fraction_range` defaulting to `(0.0, 1.0)`, an attenuation by a uniform
|
|
12
|
+
fraction of **the modulated edge's own weight**.
|
|
13
|
+
|
|
14
|
+
- **WHY THIS IS A MINOR BUMP AND NOT A PATCH.** At `p_pure = 1.0` the slab draw is
|
|
15
|
+
not taken, which shifts every subsequent draw in the stream. **A seed no longer
|
|
16
|
+
reproduces the graph it produced under 1.4.x.** The graphs are not wrong and
|
|
17
|
+
nothing silently changes meaning, but any artifact that records a seed under the
|
|
18
|
+
old draw must be regenerated or marked as pre-break. The previous release shipped
|
|
19
|
+
a breaking change in a patch and should not have. This one does not repeat it.
|
|
20
|
+
|
|
21
|
+
- **The slab is RELATIVE to the edge's weight, which is why
|
|
22
|
+
`gate_weight_bits_range` is removed rather than kept.** An absolute range made
|
|
23
|
+
the gated weight independent of the edge it modulated, so one setting was
|
|
24
|
+
near-total defeat of a weak edge and a scratch on a strong one. A gate attenuates
|
|
25
|
+
a specific edge, and "a fraction of whatever this edge was worth" is the claim.
|
|
26
|
+
|
|
27
|
+
- **SIGN REVERSAL IS NO LONGER REACHABLE BY DEFAULT, deliberately and reversibly.**
|
|
28
|
+
A same-sign slab cannot cross zero. Under the retired absolute draw, 27.8% of
|
|
29
|
+
gates reversed the sign of their edge, which was a property of that draw rather
|
|
30
|
+
than a modelling choice. Widening the slab below zero , `slab_fraction_range =
|
|
31
|
+
(-0.5, 1.0)`, makes reversal reachable again with no code change.
|
|
32
|
+
|
|
33
|
+
### Changed, and this also changes every generated graph
|
|
34
|
+
|
|
35
|
+
- **`GraphSpec.evidence_observed` now defaults to `True` rather than `"world"`.**
|
|
36
|
+
Every virtual-evidence node in a generated graph is observed TRUE, instead of
|
|
37
|
+
being observed according to the world the generator sampled. A reader of a
|
|
38
|
+
generated document therefore starts with every instrument having reported,
|
|
39
|
+
rather than with the subset the sampled world happened to trigger.
|
|
40
|
+
|
|
41
|
+
This is a generation default, not a format change: `observed` was already a
|
|
42
|
+
document field and still is. But it moves the numbers in every graph drawn
|
|
43
|
+
after it, so a set built before and a set built after are not comparable even
|
|
44
|
+
at the same seed.
|
|
45
|
+
|
|
46
|
+
### The SCHEMA version deliberately stays at 1.4.5
|
|
47
|
+
|
|
48
|
+
- **`serialize.VERSION` is the DOCUMENT FORMAT version and no document field
|
|
49
|
+
changed**, so it does not move. The package version goes to 1.5.0 because the
|
|
50
|
+
generator's DRAW changed. The format did not.
|
|
51
|
+
- I bumped both at first and it **stopped another seat within minutes.** The shared
|
|
52
|
+
editable install means this working tree is what every interpreter on the machine
|
|
53
|
+
imports, the reader's series check then refused all 1.4.x documents, and six of
|
|
54
|
+
six committed benchmark graphs plus every run record stopped loading. Reverted
|
|
55
|
+
after verifying all six load again.
|
|
56
|
+
- **A package version and a schema version answer different questions.** "Did
|
|
57
|
+
behaviour change" and "did the file format change" are not the same question, and
|
|
58
|
+
only the second one may refuse to read a file. A breaking change to a DRAW cannot
|
|
59
|
+
justify refusing to read a document.
|
|
60
|
+
|
|
61
|
+
### Removed
|
|
62
|
+
|
|
63
|
+
- **`GraphSpec.p_hard_undercut`**, superseded by `p_pure`. It named one case of
|
|
64
|
+
what is now a distribution, and the spike-and-slab form says what the draw is.
|
|
65
|
+
- **`GraphSpec.gate_weight_bits_range`**, superseded by `slab_fraction_range`.
|
|
66
|
+
Passing either now raises.
|
|
67
|
+
|
|
68
|
+
## 1.4.5
|
|
69
|
+
|
|
70
|
+
### Removed, and this is a BREAKING CHANGE IN A PATCH RELEASE
|
|
71
|
+
|
|
72
|
+
- **`GraphSpec` no longer accepts `max_attempts` or `gate_size_range`.** Both
|
|
73
|
+
were declared and never read, so passing either did nothing. Passing one now
|
|
74
|
+
raises:
|
|
75
|
+
|
|
76
|
+
GraphSpec(max_attempts=5)
|
|
77
|
+
TypeError: GraphSpec.__init__() got an unexpected keyword argument 'max_attempts'
|
|
78
|
+
|
|
79
|
+
Raising is the better behaviour and the removal is right. **Shipping it in a
|
|
80
|
+
patch was not**, and the fault is in how this release was assembled rather
|
|
81
|
+
than in the change: the version level was chosen by looking at one commit's
|
|
82
|
+
diff instead of the diff between the tags, which is what a release is. Twelve
|
|
83
|
+
commits shipped and the note described one of them.
|
|
84
|
+
|
|
85
|
+
If you are pinned inside 1.4.x and construct a `GraphSpec` with either name,
|
|
86
|
+
delete the argument. It was never doing anything.
|
|
87
|
+
|
|
88
|
+
### Added
|
|
89
|
+
|
|
90
|
+
- **`evidence_depth_tilt` on `GraphSpec`**, which makes the debate shape
|
|
91
|
+
reachable rather than only the layout.
|
|
92
|
+
- **`variable.gates` and `variable.gate_sources`.** A gate source is a parent,
|
|
93
|
+
so it appears in `parents`, and nothing in that list said which of them were
|
|
94
|
+
there to gate. The only route was `cpd.gates`, which exists on the additive
|
|
95
|
+
family and raises on the other three and on a node with no CPD yet, so
|
|
96
|
+
walking a graph for its gates meant a try/except at every node. Both new
|
|
97
|
+
properties are `()` on anything without gates.
|
|
98
|
+
|
|
99
|
+
### Fixed
|
|
100
|
+
|
|
101
|
+
- **`AtMost` could draw an illegal threshold** in generation.
|
|
102
|
+
- **`evidence_weight_bits_range` was not the range it produced**, because the
|
|
103
|
+
weight is drawn first.
|
|
104
|
+
- **Provenance is now absent rather than wrong** where it cannot be determined,
|
|
105
|
+
and records the spec alongside the seed. A seed without a spec does not say
|
|
106
|
+
what was generated.
|
|
107
|
+
|
|
3
108
|
## 1.4.4
|
|
4
109
|
|
|
5
110
|
### Added
|