torchsx 0.2.2__tar.gz → 0.2.4__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.
- torchsx-0.2.4/PKG-INFO +239 -0
- torchsx-0.2.4/README.md +212 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/pyproject.toml +6 -6
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_amortized_training.py +14 -3
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_average_last.py +2 -2
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_cli_pipeline.py +172 -42
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_cli_production.py +91 -11
- torchsx-0.2.4/tests/test_completeness.py +84 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_fold_transformer.py +100 -1
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_grid_tooling.py +10 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_large_batch_diagnostics.py +0 -2
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_phase_classes.py +3 -2
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_posterior_prior.py +8 -14
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_prepared_store.py +53 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_tether.py +11 -5
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/__init__.py +1 -1
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/_pipeline.py +195 -131
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/merge_nn.py +3 -3
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/train.py +61 -9
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/times.py +7 -2
- torchsx-0.2.4/torchsx/merging/hkl/stats.py +108 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/prepare_store.py +11 -1
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scale/amortized.py +11 -3
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/store.py +27 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/fold_transformer.py +79 -11
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/posterior_target.py +10 -13
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/train_amortized.py +76 -23
- torchsx-0.2.4/torchsx.egg-info/PKG-INFO +239 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx.egg-info/SOURCES.txt +1 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx.egg-info/requires.txt +2 -2
- torchsx-0.2.2/PKG-INFO +0 -224
- torchsx-0.2.2/README.md +0 -197
- torchsx-0.2.2/torchsx/merging/hkl/stats.py +0 -78
- torchsx-0.2.2/torchsx.egg-info/PKG-INFO +0 -224
- {torchsx-0.2.2 → torchsx-0.2.4}/LICENSE +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/setup.cfg +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_amortized_scale.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_asym_scatter.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_bg_mean_offset.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_core.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_duckdb_load.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_ewald_offset.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_filter.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_geometry.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_geometry_reciprocal.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_hkl.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_io.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_lazy_imports.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_levels.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_measurement_sources.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_merge_stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_merging_layers.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_metadata.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_native_index.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_native_parse.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_observation_filter.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_peer_features.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_pooling.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_posterior_nodes.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_posterior_smooth_sigma.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_pseudo_translation.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_reciprocal_basis.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_scatter_mixture.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_split_gaussian_mean.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_split_log_grid.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_split_mixture.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_stateless_merge.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_stream.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_structured_prior.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_train_schedule.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_wilson_surface.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_writeback.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/tests/test_writeback_native.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/_categorical.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/_core.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/merge_stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/cli/stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/geometry.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_crystfel.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_duckdb.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_metadata.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_native.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_parser.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_safetensors.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/_writeback.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/io/sources.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/levels.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/measurement.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merge_stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/config.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/hkl/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/hkl/filter.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/hkl/hkl.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/hkl/noise.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/likelihood/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/likelihood/base.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/likelihood/detector_noise.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/likelihood/poisson.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/observations.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/priors/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/priors/modulation.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/priors/reciprocal_basis.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/_posterior_nodes_triton.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/_split_log_grid_triton.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/_split_mixture_triton.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/base.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/quadrature/split_log_grid.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scale/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scale/pseudo_translation.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scatter/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scatter/base.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scatter/gengauss.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scatter/mixture.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/scatter/predicted.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/tether/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/tether/batch.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/tether/kmean.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/tether/terms.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/merging/tether/tether.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/__init__.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/heads.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/merge_target.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/metrics.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/objective.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/obs_features.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/nn/pooling.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/stats.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx/stream.py +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx.egg-info/dependency_links.txt +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx.egg-info/entry_points.txt +0 -0
- {torchsx-0.2.2 → torchsx-0.2.4}/torchsx.egg-info/top_level.txt +0 -0
torchsx-0.2.4/PKG-INFO
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: torchsx
|
|
3
|
+
Version: 0.2.4
|
|
4
|
+
Summary: A PyTorch-based scaling and merging library for serial crystallography
|
|
5
|
+
Author: Hans Peter Seidel
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Project-URL: Repository, https://github.com/HatPdotS/TorchSX
|
|
8
|
+
Classifier: Programming Language :: Python :: 3
|
|
9
|
+
Classifier: Programming Language :: Python :: 3 :: Only
|
|
10
|
+
Classifier: Intended Audience :: Science/Research
|
|
11
|
+
Classifier: Topic :: Scientific/Engineering :: Physics
|
|
12
|
+
Requires-Python: >=3.10
|
|
13
|
+
Description-Content-Type: text/markdown
|
|
14
|
+
License-File: LICENSE
|
|
15
|
+
Requires-Dist: torchref<0.9,>=0.8.1
|
|
16
|
+
Requires-Dist: torch<2.15,>=2.7
|
|
17
|
+
Requires-Dist: numpy>=2
|
|
18
|
+
Requires-Dist: safetensors>=0.4
|
|
19
|
+
Requires-Dist: matplotlib>=3.7
|
|
20
|
+
Requires-Dist: duckdb>=1.0
|
|
21
|
+
Requires-Dist: torchsx-native==0.1.*
|
|
22
|
+
Provides-Extra: dev
|
|
23
|
+
Requires-Dist: pytest>=6.0; extra == "dev"
|
|
24
|
+
Requires-Dist: black>=21.5; extra == "dev"
|
|
25
|
+
Requires-Dist: isort>=5.9; extra == "dev"
|
|
26
|
+
Dynamic: license-file
|
|
27
|
+
|
|
28
|
+
<p align="center">
|
|
29
|
+
<picture>
|
|
30
|
+
<source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/torchsx-banner-dark.png">
|
|
31
|
+
<img src="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/torchsx-banner-light.png" alt="TorchSX" width="100%">
|
|
32
|
+
</picture>
|
|
33
|
+
</p>
|
|
34
|
+
|
|
35
|
+
<h1 align="center">A Py<ins>Torch</ins>-based scaling and merging library for serial crystallography (<ins>SX</ins>)</h1>
|
|
36
|
+
|
|
37
|
+
<div align="center">
|
|
38
|
+
|
|
39
|
+
[](https://www.python.org/downloads/)
|
|
40
|
+
[](https://pytorch.org/)
|
|
41
|
+
[](https://github.com/HatPdotS/TorchRef)
|
|
42
|
+
[](https://developer.nvidia.com/cuda-zone)
|
|
43
|
+
[](https://opensource.org/licenses/MIT)
|
|
44
|
+
|
|
45
|
+
</div>
|
|
46
|
+
|
|
47
|
+
TorchSX merges serial crystallography data with an error model learned for
|
|
48
|
+
every single observation instead of a few parameters per crystal. On the
|
|
49
|
+
time-resolved tubulin data below, its difference maps from 3,875 images per
|
|
50
|
+
state are as good as Partialator's from 31,000.
|
|
51
|
+
|
|
52
|
+
<p align="center">
|
|
53
|
+
<img src="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/readme-tubulin-overview.webp" alt="TorchSX overview: a fold transformer predicts a scale and a scatter law for every observation; tubulin azo-combretastatin difference maps from TorchSX and Partialator against the number of images per state" width="100%">
|
|
54
|
+
</p>
|
|
55
|
+
|
|
56
|
+
**Top:** a fold transformer reads one crystal's observations together:
|
|
57
|
+
geometry, I, σ and I/Ī for most of them, geometry only for the fold whose
|
|
58
|
+
observations it is predicting, and the crystal's cell, energy and resolution.
|
|
59
|
+
For every observation it predicts a scale K and a scatter law, a mixture of
|
|
60
|
+
Gaussians in log intensity whose components cover pooled, partial, outlier
|
|
61
|
+
and sharp observations. **Bottom:** light-minus-dark difference maps of
|
|
62
|
+
azo-combretastatin unbinding from tubulin, 1 ns after light activation
|
|
63
|
+
(Wranik et al.<sup>[1](#references)</sup>), against the number of images per
|
|
64
|
+
state in the merge. The correlation with the noise-free map is estimated as
|
|
65
|
+
√CC, where CC is the correlation between the maps of two independent merges,
|
|
66
|
+
A and B, of disjoint images; open symbols are the full merge (A + B). Shaded: TorchSX with
|
|
67
|
+
3,875 images per state and Partialator with 31,000 (8×), the closest pair in
|
|
68
|
+
that correlation.
|
|
69
|
+
|
|
70
|
+
## What it does
|
|
71
|
+
|
|
72
|
+
- **Learns an error model per observation.** One network per dataset predicts,
|
|
73
|
+
for every observation, a scale K (partiality, pulse intensity, crystal
|
|
74
|
+
volume) and how far the measurement may plausibly fall from it.
|
|
75
|
+
- **Never uses an observation to predict itself.** Each crystal's observations
|
|
76
|
+
are split into five folds. The network reads the whole crystal once per
|
|
77
|
+
fold, with that fold's measurements hidden, and predicts the hidden ones
|
|
78
|
+
from their geometry and the rest of the crystal. On the way it corrects the
|
|
79
|
+
crystal's orientation and photon energy.
|
|
80
|
+
- **Merges by posterior, not by averaging.** Each reflection's intensity is
|
|
81
|
+
integrated out under a Wilson prior. `F-obs` and `SIGF-obs` are the mean and
|
|
82
|
+
standard deviation of the posterior of |F| given all its observations.
|
|
83
|
+
- **Cannot see light or dark.** The scale network is never given a state,
|
|
84
|
+
delay or timepoint, so it cannot absorb a difference signal. One network
|
|
85
|
+
serves all states, and each state is merged on its own.
|
|
86
|
+
- **Holds the merge at the sum of the spots.** A tether keeps the mean scale
|
|
87
|
+
free of trends in resolution and direction, so partiality stays in K and the
|
|
88
|
+
fall-off of the integrated intensities stays in the merged data.
|
|
89
|
+
- **Handles pseudo-translation.** Translational NCS, lattice translocation
|
|
90
|
+
defects and stacking faults are found in the data and modelled in the prior
|
|
91
|
+
and the scale model.
|
|
92
|
+
|
|
93
|
+
TorchSX reads integrated CrystFEL `.stream` files and probixi `.duckdb` runs
|
|
94
|
+
and writes merged MTZs. DIALS input is in progress. TorchSX is under active
|
|
95
|
+
development: if something behaves unexpectedly, please open an issue on GitHub.
|
|
96
|
+
|
|
97
|
+
The output works with any refinement or map program. Refinement and difference
|
|
98
|
+
maps in this repo were done with [TorchRef](https://github.com/HatPdotS/TorchRef)<sup>[2](#references)</sup>.
|
|
99
|
+
|
|
100
|
+
## Install
|
|
101
|
+
|
|
102
|
+
```bash
|
|
103
|
+
pip install torchsx
|
|
104
|
+
```
|
|
105
|
+
|
|
106
|
+
Python ≥ 3.10. This brings TorchRef 0.8, torch, triton, gemmi and
|
|
107
|
+
`torchsx-native`, the compiled `.stream` reader for your platform.
|
|
108
|
+
|
|
109
|
+
- **GPU.** TorchSX is meant to run on a GPU: its core algorithms are Triton
|
|
110
|
+
kernels, and CPU runtime, or whether it runs on a CPU at all, has not been
|
|
111
|
+
investigated.
|
|
112
|
+
- **GPU memory** should stay below 10 GB. It scales with the redundancy
|
|
113
|
+
requested for training (default 100) and the number of unique reflections,
|
|
114
|
+
not with the amount of data.
|
|
115
|
+
- **Host memory** does not scale with the amount of data either. The inputs
|
|
116
|
+
are prepared block by block into an on-disk store of 56 bytes per
|
|
117
|
+
observation, and training and merging read it one batch at a time, the next
|
|
118
|
+
batch while the current one is computed.
|
|
119
|
+
|
|
120
|
+
From a checkout, build the native package first (needs Rust, see
|
|
121
|
+
[native/README.md](https://github.com/HatPdotS/TorchSX/blob/master/native/README.md)):
|
|
122
|
+
`pip install ./native`, then `pip install -e .`.
|
|
123
|
+
|
|
124
|
+
## Quick start
|
|
125
|
+
|
|
126
|
+
```bash
|
|
127
|
+
torchsx.train indexed.stream --spacegroup 'P 63' --dmin 2.0 -o run/
|
|
128
|
+
```
|
|
129
|
+
|
|
130
|
+
This trains the network (400 updates by default) and merges the data with it.
|
|
131
|
+
`run/` then contains:
|
|
132
|
+
|
|
133
|
+
```
|
|
134
|
+
run/
|
|
135
|
+
├── merged.mtz merged amplitudes from all crystals
|
|
136
|
+
├── merged_half1.mtz the same from one half of the crystals
|
|
137
|
+
├── merged_half2.mtz ... and from the other half
|
|
138
|
+
├── merged.metrics.json in-run precision: CC1/2 and R_split between the two halves
|
|
139
|
+
├── best.pt the network at the lowest validation loss; it made the merges
|
|
140
|
+
├── last.pt the network at the end of training
|
|
141
|
+
├── prepared/ the prepared observations on disk, reused by later runs on the same input
|
|
142
|
+
├── run.json the resolved settings
|
|
143
|
+
├── training_summary.json updates, best step, run time; time and GPU memory by phase
|
|
144
|
+
└── metrics.jsonl training log: one line per update and per validation
|
|
145
|
+
```
|
|
146
|
+
|
|
147
|
+
Each MTZ holds:
|
|
148
|
+
|
|
149
|
+
| Column | MTZ type | Content |
|
|
150
|
+
|---|---|---|
|
|
151
|
+
| `H`, `K`, `L` | H | Miller index; Friedel mates merged |
|
|
152
|
+
| `F-obs` | F | posterior mean of the amplitude \|F\| |
|
|
153
|
+
| `SIGF-obs` | Q | posterior standard deviation of the amplitude |
|
|
154
|
+
| `R-free-flags` | I | refinement free set: 0 = free (5 %), 1 = work |
|
|
155
|
+
|
|
156
|
+
There are no intensity columns. `F-obs` is already the posterior estimate of
|
|
157
|
+
the amplitude, so it must not go through French–Wilson again; refinement
|
|
158
|
+
programs use it as it is. With `--no-merge-friedel` the mates stay apart and the
|
|
159
|
+
files carry `F-obs(+)`, `SIGF-obs(+)`, `F-obs(-)` and `SIGF-obs(-)`.
|
|
160
|
+
`--free-from earlier.mtz` takes the free set from an existing file.
|
|
161
|
+
|
|
162
|
+
The two halves split the crystals, not the training: both are merged with the
|
|
163
|
+
same network, so the CC1/2 and R_split in `merged.metrics.json` measure the
|
|
164
|
+
in-run precision, not the agreement of independent merges. For independent
|
|
165
|
+
halves, train two runs on disjoint sets of patterns.
|
|
166
|
+
|
|
167
|
+
### Light and dark
|
|
168
|
+
|
|
169
|
+
```bash
|
|
170
|
+
torchsx.train --split dark=dark.stream --split light=light.stream \
|
|
171
|
+
--spacegroup 'P 63' --dmin 2.0 -o run/
|
|
172
|
+
```
|
|
173
|
+
|
|
174
|
+
One network is trained on both states, and each state is merged on its own:
|
|
175
|
+
`merged_dark.mtz` and `merged_light.mtz`, each with its two halves. A state
|
|
176
|
+
can come from several streams (repeat its name), and streams with a CrystFEL
|
|
177
|
+
custom-split file work too (`--split-file`); patterns the file assigns to no
|
|
178
|
+
state still train the network. See
|
|
179
|
+
[States](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md#states).
|
|
180
|
+
|
|
181
|
+
## The model
|
|
182
|
+
|
|
183
|
+
The signal that reached the detector aperture of observation *j* of
|
|
184
|
+
reflection *h* is
|
|
185
|
+
|
|
186
|
+
S_j = K_j · I_h · exp(ε_j)
|
|
187
|
+
|
|
188
|
+
where I_h is the reflection's true intensity, a latent variable with a Wilson
|
|
189
|
+
prior; K_j is the predicted scale of this observation; and ε_j is a random
|
|
190
|
+
multiplicative scatter, with E[exp(ε_j)] = 1, whose distribution is also
|
|
191
|
+
predicted for this observation: a mixture of four Gaussians in log space, each
|
|
192
|
+
of which may be wider below its centre than above it, as the scatter of
|
|
193
|
+
partial observations is. The detector then records Poisson counts of S_j on
|
|
194
|
+
top of the background.
|
|
195
|
+
|
|
196
|
+
The network is trained on random batches of whole crystals, sized to a given
|
|
197
|
+
multiplicity (`--redundancy`, default 100), by maximising the marginal
|
|
198
|
+
likelihood of their observations with every I_h integrated out. No point
|
|
199
|
+
estimate of the intensity enters the loss, so the quantity that is scored is
|
|
200
|
+
the quantity that is merged. Like a weather forecaster scored on its
|
|
201
|
+
forecasts, the network is rewarded for confidence where the data agree and
|
|
202
|
+
penalised for it where they do not; a wider scatter costs likelihood, so
|
|
203
|
+
agreement cannot be bought by widening it. The integrals over I_h and ε_j run
|
|
204
|
+
in Triton kernels that never hold the full integrand in memory.
|
|
205
|
+
|
|
206
|
+
The [command-line reference](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md#model)
|
|
207
|
+
describes the model, the tether and the pseudo-translation search in more
|
|
208
|
+
detail.
|
|
209
|
+
|
|
210
|
+
## Other commands
|
|
211
|
+
|
|
212
|
+
- `torchsx.merge run/best.pt more.stream -o merged.mtz` merges new or additional
|
|
213
|
+
data with a trained network.
|
|
214
|
+
- `torchsx.merge-stats A/merged.mtz B/merged.mtz -o stats/` compares merges:
|
|
215
|
+
completeness, CC1/2, R_split and, for light/dark pairs, the information in
|
|
216
|
+
ΔF ([docs/merge_stats.md](https://github.com/HatPdotS/TorchSX/blob/master/docs/merge_stats.md)).
|
|
217
|
+
`torchsx.merge-stats run/` finds the merges of each state in a run directory
|
|
218
|
+
and writes `run/merge_stats/`.
|
|
219
|
+
- `torchsx.observation-stats indexed.stream --spacegroup 'P 63' -o figures/`
|
|
220
|
+
plots the raw observations: intensity distributions, skewness, examples
|
|
221
|
+
([docs/stats.md](https://github.com/HatPdotS/TorchSX/blob/master/docs/stats.md)).
|
|
222
|
+
|
|
223
|
+
All options, gain and measurement settings, caching and reproducibility are in
|
|
224
|
+
the [command-line reference](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md).
|
|
225
|
+
|
|
226
|
+
In Python, `torchsx.SFXStream.read('indexed.stream', spacegroup='P 63')` reads a
|
|
227
|
+
stream with its geometry and measurement settings.
|
|
228
|
+
|
|
229
|
+
## References
|
|
230
|
+
|
|
231
|
+
1. Wranik, M. et al. Watching the release of a photopharmacological drug from
|
|
232
|
+
tubulin using time-resolved serial crystallography. *Nat. Commun.* **14**,
|
|
233
|
+
903 (2023). [doi:10.1038/s41467-023-36481-5](https://doi.org/10.1038/s41467-023-36481-5)
|
|
234
|
+
2. Seidel, H.-P., Standfuss, J. & Weinert, T. TorchRef: an open-source PyTorch
|
|
235
|
+
framework for crystallographic refinement. *J. Appl. Cryst.*, accepted (2026).
|
|
236
|
+
|
|
237
|
+
## License
|
|
238
|
+
|
|
239
|
+
MIT. See [LICENSE](https://github.com/HatPdotS/TorchSX/blob/master/LICENSE).
|
torchsx-0.2.4/README.md
ADDED
|
@@ -0,0 +1,212 @@
|
|
|
1
|
+
<p align="center">
|
|
2
|
+
<picture>
|
|
3
|
+
<source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/torchsx-banner-dark.png">
|
|
4
|
+
<img src="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/torchsx-banner-light.png" alt="TorchSX" width="100%">
|
|
5
|
+
</picture>
|
|
6
|
+
</p>
|
|
7
|
+
|
|
8
|
+
<h1 align="center">A Py<ins>Torch</ins>-based scaling and merging library for serial crystallography (<ins>SX</ins>)</h1>
|
|
9
|
+
|
|
10
|
+
<div align="center">
|
|
11
|
+
|
|
12
|
+
[](https://www.python.org/downloads/)
|
|
13
|
+
[](https://pytorch.org/)
|
|
14
|
+
[](https://github.com/HatPdotS/TorchRef)
|
|
15
|
+
[](https://developer.nvidia.com/cuda-zone)
|
|
16
|
+
[](https://opensource.org/licenses/MIT)
|
|
17
|
+
|
|
18
|
+
</div>
|
|
19
|
+
|
|
20
|
+
TorchSX merges serial crystallography data with an error model learned for
|
|
21
|
+
every single observation instead of a few parameters per crystal. On the
|
|
22
|
+
time-resolved tubulin data below, its difference maps from 3,875 images per
|
|
23
|
+
state are as good as Partialator's from 31,000.
|
|
24
|
+
|
|
25
|
+
<p align="center">
|
|
26
|
+
<img src="https://raw.githubusercontent.com/HatPdotS/TorchSX/master/assets/readme-tubulin-overview.webp" alt="TorchSX overview: a fold transformer predicts a scale and a scatter law for every observation; tubulin azo-combretastatin difference maps from TorchSX and Partialator against the number of images per state" width="100%">
|
|
27
|
+
</p>
|
|
28
|
+
|
|
29
|
+
**Top:** a fold transformer reads one crystal's observations together:
|
|
30
|
+
geometry, I, σ and I/Ī for most of them, geometry only for the fold whose
|
|
31
|
+
observations it is predicting, and the crystal's cell, energy and resolution.
|
|
32
|
+
For every observation it predicts a scale K and a scatter law, a mixture of
|
|
33
|
+
Gaussians in log intensity whose components cover pooled, partial, outlier
|
|
34
|
+
and sharp observations. **Bottom:** light-minus-dark difference maps of
|
|
35
|
+
azo-combretastatin unbinding from tubulin, 1 ns after light activation
|
|
36
|
+
(Wranik et al.<sup>[1](#references)</sup>), against the number of images per
|
|
37
|
+
state in the merge. The correlation with the noise-free map is estimated as
|
|
38
|
+
√CC, where CC is the correlation between the maps of two independent merges,
|
|
39
|
+
A and B, of disjoint images; open symbols are the full merge (A + B). Shaded: TorchSX with
|
|
40
|
+
3,875 images per state and Partialator with 31,000 (8×), the closest pair in
|
|
41
|
+
that correlation.
|
|
42
|
+
|
|
43
|
+
## What it does
|
|
44
|
+
|
|
45
|
+
- **Learns an error model per observation.** One network per dataset predicts,
|
|
46
|
+
for every observation, a scale K (partiality, pulse intensity, crystal
|
|
47
|
+
volume) and how far the measurement may plausibly fall from it.
|
|
48
|
+
- **Never uses an observation to predict itself.** Each crystal's observations
|
|
49
|
+
are split into five folds. The network reads the whole crystal once per
|
|
50
|
+
fold, with that fold's measurements hidden, and predicts the hidden ones
|
|
51
|
+
from their geometry and the rest of the crystal. On the way it corrects the
|
|
52
|
+
crystal's orientation and photon energy.
|
|
53
|
+
- **Merges by posterior, not by averaging.** Each reflection's intensity is
|
|
54
|
+
integrated out under a Wilson prior. `F-obs` and `SIGF-obs` are the mean and
|
|
55
|
+
standard deviation of the posterior of |F| given all its observations.
|
|
56
|
+
- **Cannot see light or dark.** The scale network is never given a state,
|
|
57
|
+
delay or timepoint, so it cannot absorb a difference signal. One network
|
|
58
|
+
serves all states, and each state is merged on its own.
|
|
59
|
+
- **Holds the merge at the sum of the spots.** A tether keeps the mean scale
|
|
60
|
+
free of trends in resolution and direction, so partiality stays in K and the
|
|
61
|
+
fall-off of the integrated intensities stays in the merged data.
|
|
62
|
+
- **Handles pseudo-translation.** Translational NCS, lattice translocation
|
|
63
|
+
defects and stacking faults are found in the data and modelled in the prior
|
|
64
|
+
and the scale model.
|
|
65
|
+
|
|
66
|
+
TorchSX reads integrated CrystFEL `.stream` files and probixi `.duckdb` runs
|
|
67
|
+
and writes merged MTZs. DIALS input is in progress. TorchSX is under active
|
|
68
|
+
development: if something behaves unexpectedly, please open an issue on GitHub.
|
|
69
|
+
|
|
70
|
+
The output works with any refinement or map program. Refinement and difference
|
|
71
|
+
maps in this repo were done with [TorchRef](https://github.com/HatPdotS/TorchRef)<sup>[2](#references)</sup>.
|
|
72
|
+
|
|
73
|
+
## Install
|
|
74
|
+
|
|
75
|
+
```bash
|
|
76
|
+
pip install torchsx
|
|
77
|
+
```
|
|
78
|
+
|
|
79
|
+
Python ≥ 3.10. This brings TorchRef 0.8, torch, triton, gemmi and
|
|
80
|
+
`torchsx-native`, the compiled `.stream` reader for your platform.
|
|
81
|
+
|
|
82
|
+
- **GPU.** TorchSX is meant to run on a GPU: its core algorithms are Triton
|
|
83
|
+
kernels, and CPU runtime, or whether it runs on a CPU at all, has not been
|
|
84
|
+
investigated.
|
|
85
|
+
- **GPU memory** should stay below 10 GB. It scales with the redundancy
|
|
86
|
+
requested for training (default 100) and the number of unique reflections,
|
|
87
|
+
not with the amount of data.
|
|
88
|
+
- **Host memory** does not scale with the amount of data either. The inputs
|
|
89
|
+
are prepared block by block into an on-disk store of 56 bytes per
|
|
90
|
+
observation, and training and merging read it one batch at a time, the next
|
|
91
|
+
batch while the current one is computed.
|
|
92
|
+
|
|
93
|
+
From a checkout, build the native package first (needs Rust, see
|
|
94
|
+
[native/README.md](https://github.com/HatPdotS/TorchSX/blob/master/native/README.md)):
|
|
95
|
+
`pip install ./native`, then `pip install -e .`.
|
|
96
|
+
|
|
97
|
+
## Quick start
|
|
98
|
+
|
|
99
|
+
```bash
|
|
100
|
+
torchsx.train indexed.stream --spacegroup 'P 63' --dmin 2.0 -o run/
|
|
101
|
+
```
|
|
102
|
+
|
|
103
|
+
This trains the network (400 updates by default) and merges the data with it.
|
|
104
|
+
`run/` then contains:
|
|
105
|
+
|
|
106
|
+
```
|
|
107
|
+
run/
|
|
108
|
+
├── merged.mtz merged amplitudes from all crystals
|
|
109
|
+
├── merged_half1.mtz the same from one half of the crystals
|
|
110
|
+
├── merged_half2.mtz ... and from the other half
|
|
111
|
+
├── merged.metrics.json in-run precision: CC1/2 and R_split between the two halves
|
|
112
|
+
├── best.pt the network at the lowest validation loss; it made the merges
|
|
113
|
+
├── last.pt the network at the end of training
|
|
114
|
+
├── prepared/ the prepared observations on disk, reused by later runs on the same input
|
|
115
|
+
├── run.json the resolved settings
|
|
116
|
+
├── training_summary.json updates, best step, run time; time and GPU memory by phase
|
|
117
|
+
└── metrics.jsonl training log: one line per update and per validation
|
|
118
|
+
```
|
|
119
|
+
|
|
120
|
+
Each MTZ holds:
|
|
121
|
+
|
|
122
|
+
| Column | MTZ type | Content |
|
|
123
|
+
|---|---|---|
|
|
124
|
+
| `H`, `K`, `L` | H | Miller index; Friedel mates merged |
|
|
125
|
+
| `F-obs` | F | posterior mean of the amplitude \|F\| |
|
|
126
|
+
| `SIGF-obs` | Q | posterior standard deviation of the amplitude |
|
|
127
|
+
| `R-free-flags` | I | refinement free set: 0 = free (5 %), 1 = work |
|
|
128
|
+
|
|
129
|
+
There are no intensity columns. `F-obs` is already the posterior estimate of
|
|
130
|
+
the amplitude, so it must not go through French–Wilson again; refinement
|
|
131
|
+
programs use it as it is. With `--no-merge-friedel` the mates stay apart and the
|
|
132
|
+
files carry `F-obs(+)`, `SIGF-obs(+)`, `F-obs(-)` and `SIGF-obs(-)`.
|
|
133
|
+
`--free-from earlier.mtz` takes the free set from an existing file.
|
|
134
|
+
|
|
135
|
+
The two halves split the crystals, not the training: both are merged with the
|
|
136
|
+
same network, so the CC1/2 and R_split in `merged.metrics.json` measure the
|
|
137
|
+
in-run precision, not the agreement of independent merges. For independent
|
|
138
|
+
halves, train two runs on disjoint sets of patterns.
|
|
139
|
+
|
|
140
|
+
### Light and dark
|
|
141
|
+
|
|
142
|
+
```bash
|
|
143
|
+
torchsx.train --split dark=dark.stream --split light=light.stream \
|
|
144
|
+
--spacegroup 'P 63' --dmin 2.0 -o run/
|
|
145
|
+
```
|
|
146
|
+
|
|
147
|
+
One network is trained on both states, and each state is merged on its own:
|
|
148
|
+
`merged_dark.mtz` and `merged_light.mtz`, each with its two halves. A state
|
|
149
|
+
can come from several streams (repeat its name), and streams with a CrystFEL
|
|
150
|
+
custom-split file work too (`--split-file`); patterns the file assigns to no
|
|
151
|
+
state still train the network. See
|
|
152
|
+
[States](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md#states).
|
|
153
|
+
|
|
154
|
+
## The model
|
|
155
|
+
|
|
156
|
+
The signal that reached the detector aperture of observation *j* of
|
|
157
|
+
reflection *h* is
|
|
158
|
+
|
|
159
|
+
S_j = K_j · I_h · exp(ε_j)
|
|
160
|
+
|
|
161
|
+
where I_h is the reflection's true intensity, a latent variable with a Wilson
|
|
162
|
+
prior; K_j is the predicted scale of this observation; and ε_j is a random
|
|
163
|
+
multiplicative scatter, with E[exp(ε_j)] = 1, whose distribution is also
|
|
164
|
+
predicted for this observation: a mixture of four Gaussians in log space, each
|
|
165
|
+
of which may be wider below its centre than above it, as the scatter of
|
|
166
|
+
partial observations is. The detector then records Poisson counts of S_j on
|
|
167
|
+
top of the background.
|
|
168
|
+
|
|
169
|
+
The network is trained on random batches of whole crystals, sized to a given
|
|
170
|
+
multiplicity (`--redundancy`, default 100), by maximising the marginal
|
|
171
|
+
likelihood of their observations with every I_h integrated out. No point
|
|
172
|
+
estimate of the intensity enters the loss, so the quantity that is scored is
|
|
173
|
+
the quantity that is merged. Like a weather forecaster scored on its
|
|
174
|
+
forecasts, the network is rewarded for confidence where the data agree and
|
|
175
|
+
penalised for it where they do not; a wider scatter costs likelihood, so
|
|
176
|
+
agreement cannot be bought by widening it. The integrals over I_h and ε_j run
|
|
177
|
+
in Triton kernels that never hold the full integrand in memory.
|
|
178
|
+
|
|
179
|
+
The [command-line reference](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md#model)
|
|
180
|
+
describes the model, the tether and the pseudo-translation search in more
|
|
181
|
+
detail.
|
|
182
|
+
|
|
183
|
+
## Other commands
|
|
184
|
+
|
|
185
|
+
- `torchsx.merge run/best.pt more.stream -o merged.mtz` merges new or additional
|
|
186
|
+
data with a trained network.
|
|
187
|
+
- `torchsx.merge-stats A/merged.mtz B/merged.mtz -o stats/` compares merges:
|
|
188
|
+
completeness, CC1/2, R_split and, for light/dark pairs, the information in
|
|
189
|
+
ΔF ([docs/merge_stats.md](https://github.com/HatPdotS/TorchSX/blob/master/docs/merge_stats.md)).
|
|
190
|
+
`torchsx.merge-stats run/` finds the merges of each state in a run directory
|
|
191
|
+
and writes `run/merge_stats/`.
|
|
192
|
+
- `torchsx.observation-stats indexed.stream --spacegroup 'P 63' -o figures/`
|
|
193
|
+
plots the raw observations: intensity distributions, skewness, examples
|
|
194
|
+
([docs/stats.md](https://github.com/HatPdotS/TorchSX/blob/master/docs/stats.md)).
|
|
195
|
+
|
|
196
|
+
All options, gain and measurement settings, caching and reproducibility are in
|
|
197
|
+
the [command-line reference](https://github.com/HatPdotS/TorchSX/blob/master/docs/cli.md).
|
|
198
|
+
|
|
199
|
+
In Python, `torchsx.SFXStream.read('indexed.stream', spacegroup='P 63')` reads a
|
|
200
|
+
stream with its geometry and measurement settings.
|
|
201
|
+
|
|
202
|
+
## References
|
|
203
|
+
|
|
204
|
+
1. Wranik, M. et al. Watching the release of a photopharmacological drug from
|
|
205
|
+
tubulin using time-resolved serial crystallography. *Nat. Commun.* **14**,
|
|
206
|
+
903 (2023). [doi:10.1038/s41467-023-36481-5](https://doi.org/10.1038/s41467-023-36481-5)
|
|
207
|
+
2. Seidel, H.-P., Standfuss, J. & Weinert, T. TorchRef: an open-source PyTorch
|
|
208
|
+
framework for crystallographic refinement. *J. Appl. Cryst.*, accepted (2026).
|
|
209
|
+
|
|
210
|
+
## License
|
|
211
|
+
|
|
212
|
+
MIT. See [LICENSE](https://github.com/HatPdotS/TorchSX/blob/master/LICENSE).
|
|
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
|
|
|
4
4
|
|
|
5
5
|
[project]
|
|
6
6
|
name = "torchsx"
|
|
7
|
-
version = "0.2.
|
|
7
|
+
version = "0.2.4"
|
|
8
8
|
description = "A PyTorch-based scaling and merging library for serial crystallography"
|
|
9
9
|
readme = "README.md"
|
|
10
10
|
license = "MIT"
|
|
@@ -20,11 +20,11 @@ classifiers = [
|
|
|
20
20
|
dependencies = [
|
|
21
21
|
# Brings torch, triton and gemmi. TorchRef's minor releases change its API,
|
|
22
22
|
# so the bound follows the minor version torchsx is tested against.
|
|
23
|
-
"torchref>=0.
|
|
24
|
-
# The
|
|
25
|
-
# on
|
|
26
|
-
#
|
|
27
|
-
"torch
|
|
23
|
+
"torchref>=0.8.1,<0.9",
|
|
24
|
+
# The fold encoder compiles with flex attention and dynamic shapes from 2.7
|
|
25
|
+
# on (compile_options sets inductor per version); where torch cannot compile
|
|
26
|
+
# it, it runs uncompiled.
|
|
27
|
+
"torch>=2.7,<2.15",
|
|
28
28
|
"numpy>=2", # the Python writer's float32 arithmetic (NEP 50)
|
|
29
29
|
"safetensors>=0.4", # the .sfx working format
|
|
30
30
|
"matplotlib>=3.7", # the figures of torchsx.observation-stats and torchsx.merge-stats
|
|
@@ -292,8 +292,18 @@ def test_per_condition_targets_keep_the_difference(tmp_path):
|
|
|
292
292
|
tr.close()
|
|
293
293
|
|
|
294
294
|
|
|
295
|
-
|
|
296
|
-
|
|
295
|
+
def test_flat_prior_is_rejected(tmp_path):
|
|
296
|
+
"""The flat prior was removed (#24): its posterior runs off the top of the grid."""
|
|
297
|
+
prep, params, *_ = _synthetic(n_hkl=12, n_xtal=24, per=12)
|
|
298
|
+
scale = AmortizedScale.from_prep(prep, d_model=8, context_folds=2)
|
|
299
|
+
with pytest.raises(ValueError, match="removed"):
|
|
300
|
+
AmortizedTrainer(
|
|
301
|
+
prep, scale, MarginalNLL(SplitLogGrid(n_points=96), Poisson()), params,
|
|
302
|
+
config=TrainConfig(device="cpu", outdir=str(tmp_path), posterior_prior="none"),
|
|
303
|
+
)
|
|
304
|
+
|
|
305
|
+
|
|
306
|
+
def test_evaluation_changes_no_state(tmp_path):
|
|
297
307
|
"""Scoring held-out data must not feed its intensities into later training."""
|
|
298
308
|
prep, params, *_ = _synthetic(n_hkl=12, n_xtal=24, per=12, n_cond=2)
|
|
299
309
|
scale = AmortizedScale.from_prep(
|
|
@@ -303,7 +313,7 @@ def test_evaluation_changes_no_state(tmp_path, prior):
|
|
|
303
313
|
prep, scale, MarginalNLL(SplitLogGrid(n_points=96), Poisson()), params,
|
|
304
314
|
config=TrainConfig(
|
|
305
315
|
max_obs_per_batch=24, heldout_frac=0.25, device="cpu", outdir=str(tmp_path),
|
|
306
|
-
posterior_prior=
|
|
316
|
+
posterior_prior="wilson", posterior_k_coarse=5, posterior_m_fine=17,
|
|
307
317
|
),
|
|
308
318
|
)
|
|
309
319
|
|
|
@@ -454,6 +464,7 @@ def test_deferred_encoder_backward_is_the_one_pass_gradient(tmp_path):
|
|
|
454
464
|
torch.testing.assert_close(grads[1][n], grads[0][n], rtol=1e-5, atol=1e-7, msg=n)
|
|
455
465
|
|
|
456
466
|
|
|
467
|
+
@pytest.mark.compile
|
|
457
468
|
@pytest.mark.skipif(not torch.cuda.is_available(), reason="the compiled fold path runs on CUDA only")
|
|
458
469
|
def test_fold_predictions_without_autograd_run_flex_and_match_the_masked_kernel():
|
|
459
470
|
"""Without autograd the fold path runs each block under a checkpoint with autograd
|
|
@@ -115,8 +115,8 @@ def test_without_averaging_the_latest_update_stays(tmp_path):
|
|
|
115
115
|
def _recorded_run(tr, oscillating, steps=6):
|
|
116
116
|
after, step = [], tr.step
|
|
117
117
|
|
|
118
|
-
def recorded(rows):
|
|
119
|
-
rec = step(rows)
|
|
118
|
+
def recorded(rows, **kwargs):
|
|
119
|
+
rec = step(rows, **kwargs)
|
|
120
120
|
after.append(_params(tr))
|
|
121
121
|
return rec
|
|
122
122
|
|