bobleesj commited on
Commit
475260b
·
verified ·
1 Parent(s): 6bdff7c

Keep synthetic data card and checksums data-only

Browse files
denoise/simulated_au_nanobeam_pg_dose1/README.md CHANGED
@@ -1,81 +1,40 @@
1
- # Synthetic Au nanobeam: start with TV
2
-
3
- Download **one noiseless abTEM scan**, not a collection of generated outputs.
4
- The HDF5 master/data pair is about **121 MiB**: 32×32 scan positions,
5
- 192×192 detector pixels, dose-1 expected detected events in float32.
6
- `geometry.json` supplies sampling and beam center; `noise-recipe.json` supplies
7
- explicit synthetic parameters and random seeds. `SHA256SUMS` checks these inputs.
8
-
9
- The tutorial generates its noisy observation, independent dark/flat exposures,
10
- validity maps and `.pt` calibration artifacts **locally on CUDA**. Local
11
- preparation took **1.2 s** on one RTX PRO 6000. Nothing is rounded to integer
12
- counts before denoising. No measured experimental data or trained weights are
13
- included. abTEM is not required to consume this precomputed expectation.
14
-
15
- ## Which tutorial first?
16
-
17
- Use the matching Denova development checkout and its `docs/tutorials` directory:
18
-
19
- 1. **[01_tv.ipynb](../01_tv.ipynb) — recommended first.** About **36 s** for TV1 strength
20
- selection/reconstruction, **42 s** across the original full notebook before
21
- download. Local measurement preparation adds about 1.2 s. This measured
22
- 32×32 example ran under one minute; other GPUs, first-use compilation and
23
- contention may take longer.
24
- 2. **[02_shine.ipynb](../02_shine.ipynb) — experimental comparison.** About **271 s (4½ min)**
25
- for 64 epochs × 128 steps and reconstruction, **276 s** across all cells
26
- before download/preparation. No pre-trained model is needed.
27
- 3. **[03_generate_abtem.ipynb](../03_generate_abtem.ipynb) — optional specimen regeneration.** About **12 s**
28
- across code cells, including a **6.94 s** propagation/detector stage. This
29
- requires abTEM, ASE, CuPy and the separate abtem-cuda accelerator.
30
-
31
- Keep `synthetic_detector.py` beside the denoising notebooks. Install the
32
- matching development code with `pip install -e '.[fourdstem]'`; a released
33
- Denova version containing this revised API has not been verified. The tested
34
- installed versions were QuantEM.GPU 0.0.1rc15 and QuantEM.Widget 0.0.1rc41.
35
-
36
- ```python
37
- from quantem.widget.io import hub
38
- from pathlib import Path
39
- from synthetic_detector import prepare_measurements
40
-
41
- source = Path(hub.download("simulated_au_nanobeam_pg_dose1", repo="bobleesj/quantem-data"))
42
- data = prepare_measurements(source, Path("4dstem-results/inputs"))
43
- ```
44
-
45
- The notebooks load the generated acquisition through QuantEM.GPU, let you
46
- inspect a native scan rectangle in Show4DSTEM, independently fit detector
47
- noise with `PoissonGaussian.calibrate`, show its maps and variance chart,
48
- denoise, and compare truth/noisy/result at the same scan position. Truth and
49
- generator parameters do not enter denoiser training or result selection.
50
- Calibration `.pt` files and reconstruction outputs are saved locally.
51
-
52
- ## Measured quality: execution is not acceptance
53
-
54
- On this new example, TV1 reduced ADU RMSE from **2.45 to 1.67**, but the
55
- 0–5% radial detector band failed the declared residual criteria. SHINE-PG
56
- increased RMSE from **2.44 to 3.59**, with failures in both inner bands.
57
- Both selectors returned `no_feasible_candidate`. The methods evaluate their
58
- own processed masks; their RMSE rows are not a common-support ranking.
59
- Use TV first for exploration, retain the failed audits, and do not infer
60
- zero false peaks or scientific acceptance from a smooth image.
61
-
62
- ## What is and is not simulated?
63
-
64
- 300 kV, 0.622 mrad convergence, four static Au nanoparticles and a declared
65
- 25% diffuse Gaussian angular response (4 mrad SD). The background is
66
- illustrative, not atomistic water/window scattering. Static-lattice and
67
- infinite-projection assumptions require physical convergence studies.
68
- Four accelerated patterns were bit-identical to native GPU abTEM; that
69
- checks numerical execution, not experimental realism.
70
-
71
- The local recipe draws independent Poisson events plus additive Gaussian
72
- noise, with spatial gain, offset and four invalid detector pixels. Darks
73
- and multilevel flats use separate seeds from the science observation.
74
- This is a general positive-read-noise baseline, **not validated ARINA count-rate,
75
- charge-sharing or covariance physics**, nor a reproduction of the frozen paper.
76
- The older complete upload remains in version history; current downloads keep
77
- only the reusable simulation input and small metadata.
78
-
79
- Experimental inspiration: **Serin Lee and Andrew Barnum**,
80
  [Au liquid-cell dataset](https://doi.org/10.5281/zenodo.18167694), CC BY 4.0.
81
- Their experimental observations are not distributed here.
 
 
1
+ # Synthetic Au nanobeam expectation, relative dose 1
2
+
3
+ Reusable input data for the calibrated 4D-STEM tutorials maintained in the
4
+ [denoise repository](https://github.com/bobleesj/denoise). Hugging Face hosts
5
+ only the data and descriptive metadata; scientific code, notebooks and result
6
+ comparisons remain in the software repository. Use its matching development
7
+ checkout; a released package containing the revised workflow is not qualified.
8
+
9
+ ## Contents and units
10
+
11
+ - `truth-events-dose-1.0_master.h5` and its linked `_data_000001.h5` shard:
12
+ noiseless abTEM expectation in detected-event units, lossless float32,
13
+ shaped `(scan_row, scan_col, detector_row, detector_col) = (32,32,192,192)`.
14
+ Keep the master and shard together.
15
+ - `geometry.json`: scan/detector sampling, beam center and simulation settings.
16
+ - `noise-recipe.json`: numeric synthetic detector parameters and random seeds.
17
+ - `meta.json`: dataset provenance.
18
+ - `SHA256SUMS`: integrity checks for the data and metadata, including this card.
19
+
20
+ The input is about 120 MiB. No noisy observations, dark/flat stacks, trained
21
+ weights, notebooks, Python helpers or `.pt` artifacts are distributed here.
22
+ Those reproducible observations and calibration files are generated locally
23
+ from this expectation and the numeric recipe.
24
+
25
+ ## Simulation scope
26
+
27
+ 300 kV, 0.622 mrad convergence, four differently oriented static Au particles,
28
+ and an assumed 25% diffuse angular response with 4 mrad Gaussian width.
29
+ The diffuse background is illustrative, not atomistic water/window scattering.
30
+ The stored expectation is not an independently measured clean reference.
31
+
32
+ The numeric recipe specifies independent Poisson arrivals plus additive
33
+ Gaussian noise, pixelwise gain/offset and four invalid detector pixels.
34
+ It is not a validated model of corrected ARINA count-rate response,
35
+ charge sharing or inter-pixel covariance, nor a reproduction of the frozen paper.
36
+
37
+ Experimental inspiration and data credit: **Serin Lee and Andrew Barnum**,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  [Au liquid-cell dataset](https://doi.org/10.5281/zenodo.18167694), CC BY 4.0.
39
+ Only synthetic data are distributed here. Earlier repository versions may
40
+ contain superseded artifacts; the current data tree follows the data-only rule.
denoise/simulated_au_nanobeam_pg_dose1/SHA256SUMS CHANGED
@@ -1,4 +1,4 @@
1
- 82ed1befb3f95a478a21b9a840ee0c21e45e0aba792a1e30ba1d40177f848d27 README.md
2
  a517d38291b837fad683051874e3ab8b7b82d09e1888e1f45ff08910f1d19e13 geometry.json
3
  35095b4ea696cd4051bdd5882c18029915a9315a3a6e36823699f983d49b843f meta.json
4
  6a355e059ab64cd3992b6fe998acc03f57c5a79e42b1915b497e75807ecd81ea noise-recipe.json
 
1
+ cd531c5fc9d7f0cab37c5b8b7cc6ec9e83bb6670c99bc0d4e513ffc635367297 README.md
2
  a517d38291b837fad683051874e3ab8b7b82d09e1888e1f45ff08910f1d19e13 geometry.json
3
  35095b4ea696cd4051bdd5882c18029915a9315a3a6e36823699f983d49b843f meta.json
4
  6a355e059ab64cd3992b6fe998acc03f57c5a79e42b1915b497e75807ecd81ea noise-recipe.json