junidude14 commited on
Commit
80e4cdf
·
verified ·
1 Parent(s): 83a4c7c

Archive 4 checkpoints (1-4)

Browse files
paper-releases/pbisc-diffusion/updates/20261013T015549Z-composition-error-comparison/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/composition_error_20261013/REPORT.md ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Healthy composition: replace overlapping scatters with paired error comparisons
2
+
3
+ The user could not distinguish superiority in overlapping, similar-colour T-cell and monocyte scatters. The new slide retains original Arial styling but uses separated orange/blue error bars and donor win counts for all six cell types.
4
+
5
+ ## Metric and pairing
6
+
7
+ Input: the same saved ImMicro exact-five-donor comparison, five draws, no additional validation and 3,000-step LoRA. There are 510 matched donor-draw observations per method and 107 unique held-out donors. Observed fractions match between methods. Each donor appears in three to five draws (85 donors in five, 19 in four, three in three).
8
+
9
+ For each type, compute `100 * abs(generated_fraction - observed_fraction)` for each donor-draw. Average within donor, then weight donors equally. The two methods are compared on the same donor and draws. A donor win is lower mean absolute error, tolerance 1e-10 percentage points. The right-hand bar lengths show percentages; their numeric labels show counts out of 107.
10
+
11
+ Overall composition L1 is the sum of six fraction errors. It agrees with saved comp_l1 to floating-point precision. PBISC mean L1 is 0.175779; scSemiProfiler 0.232414. Relative reduction is 24.3683%; PBISC has lower overall error in 93/107 donors.
12
+
13
+ | Type | PBISC MAE (pp) | scSemi MAE (pp) | PBISC / scSemi donor wins |
14
+ |---|---:|---:|---:|
15
+ | T cells | 6.2194 | 6.8621 | 63 / 44 |
16
+ | B cells | 1.9524 | 3.1745 | 93 / 14 |
17
+ | Monocytes | 4.5507 | 5.3897 | 61 / 46 |
18
+ | NK cells | 3.6388 | 4.4150 | 78 / 29 |
19
+ | DC | 0.8118 | 3.0234 | 101 / 6 |
20
+ | Other | 0.4047 | 0.3768 | 36 / 71 |
21
+
22
+ PBISC has lower mean error in five of six types. T-cell and monocyte improvements are modest; Other slightly favors scSemiProfiler. B-cell and DC differences are larger. No p-values or general population superiority claims are made. Overlapping support draws are not independent training replicates.
23
+
24
+ ## Difference from the prior display
25
+
26
+ The old scatter averaged proportions across held-out draws before plotting. Its distance from the diagonal can cancel errors with opposite signs. The replacement measures each draw's absolute error first, then averages within donor. This is a new descriptive summary of the same outputs, not a change to model predictions or a rerun of experiments. The old scatter's sum-error means were 0.163944 versus 0.206845; those are a different estimator and are not substituted into the new panel.
27
+
28
+ The overall-L1 comparison, mean type errors and win fractions answer distinct descriptive questions. All types, including the one favoring scSemiProfiler, remain visible.
29
+
30
+ ## Validation
31
+
32
+ An independent calculation matched all 642 donor-by-type errors to at most 1.42e-14 pp and all type means and win counts. The primary per-draw L1 matches saved comp_l1. Non-target new figure PNGs are byte-identical to the prior deck. The deck retains 34 slides and original figure assets; whole-deck source hashes, text/shape bounds and PDF rendering are verified. No GPU job, training or new cell generation was launched.
paper-releases/pbisc-diffusion/updates/20261013T015549Z-composition-error-comparison/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/style_review_20261013/VERIFICATION.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "pbisc.deck-style-verification.v1",
3
+ "slides": 34,
4
+ "pdf_pages": 34,
5
+ "style_renderer": "original generate_deck.py",
6
+ "fonts": [
7
+ "Arial"
8
+ ],
9
+ "reused_original_figure_images_verified": 26,
10
+ "all_figures_uncropped": true,
11
+ "source_hashes_verified": 101,
12
+ "originals_preserved": 6,
13
+ "shape_and_pdf_text_bounds_pass": true,
14
+ "pdf_sha256": "eaa3fb51f6e4d7618f955d47b13f25f007984a2eeb447e17b2b2c6e2a54788c7",
15
+ "pptx_sha256": "c79955158717996ac5e14dc28afd4346a6b8599d22e9b6233bcc55fd336630c0",
16
+ "new_training": false,
17
+ "new_generation": false,
18
+ "underlying_experiment_outputs_changed": false,
19
+ "derived_metric_update": "Healthy composition: paired absolute errors before donor averaging; win counts added"
20
+ }
paper-releases/pbisc-diffusion/updates/20261013T015549Z-composition-error-comparison/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/style_review_20261013/slide_13.png ADDED

Git LFS Details

  • SHA256: 20eb139277e3b214c3309d71b9a13655475164e5ce9f48fe774a6b4c4c48fc8b
  • Pointer size: 131 Bytes
  • Size of remote file: 197 kB
paper-releases/pbisc-diffusion/updates/20261013T015549Z-composition-error-comparison/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/style_review_20261013/verify.py ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from pathlib import Path
2
+ import hashlib,json,io
3
+ import fitz
4
+ from pptx import Presentation
5
+ from PIL import Image,ImageDraw
6
+ D=Path(__file__).resolve().parents[2];A=Path(__file__).resolve().parent
7
+ stem=D/'PBISC-Diffusion_results_review_20261013'
8
+ def sha(p):
9
+ with Path(p).open('rb') as f:return hashlib.file_digest(f,'sha256').hexdigest()
10
+ originals=json.loads((A/'originals_before.json').read_text())
11
+ for p,h in originals.items():assert sha(D/p)==h,p
12
+ spec=json.loads((D/'data/results_review_20261013_spec.json').read_text())['slides']
13
+ manifest=json.loads((D/'data/results_review_20261013_provenance.json').read_text())
14
+ for p,r in manifest['sources'].items():assert sha(p)==r['sha256'],p
15
+ prs=Presentation(stem.with_suffix('.pptx'));pdf=fitz.open(stem.with_suffix('.pdf'))
16
+ assert len(prs.slides)==len(pdf)==len(spec)==34
17
+ assert 'Historical paired-data learning curve' in pdf[8].get_text()
18
+ assert 'CV-HVG reuse' in pdf[8].get_text()
19
+ assert 'Identity screening where available' in pdf[15].get_text()
20
+ reused=0;fonts=set()
21
+ for i,(slide,s) in enumerate(zip(prs.slides,spec)):
22
+ for sh in slide.shapes:
23
+ assert sh.left>=0 and sh.top>=0 and sh.left+sh.width<=prs.slide_width+2 and sh.top+sh.height<=prs.slide_height+2
24
+ if sh.has_text_frame:
25
+ for p in sh.text_frame.paragraphs:
26
+ for r in p.runs:
27
+ if r.text.strip():
28
+ fonts.add(r.font.name);assert r.font.name=='Arial',(i,r.text,r.font.name)
29
+ pictures=[sh for sh in slide.shapes if sh.shape_type==13]
30
+ if s.get('figures'):
31
+ assert len(pictures)==len(s['figures'])
32
+ for sh,f in zip(pictures,s['figures']):
33
+ assert hashlib.sha256(sh.image.blob).hexdigest()==sha(D/f)
34
+ assert (sh.crop_left,sh.crop_top,sh.crop_right,sh.crop_bottom)==(0,0,0,0)
35
+ if s.get('original_slide'):reused+=len(pictures)
36
+ for block in pdf[i].get_text('dict')['blocks']:
37
+ if block['type']!=0:continue
38
+ for line in block['lines']:
39
+ for span in line['spans']:
40
+ assert fitz.Rect(span['bbox']).x1<=pdf[i].rect.width+.5,(i,span['text'])
41
+ assert fitz.Rect(span['bbox']).y1<=pdf[i].rect.height+.5,(i,span['text'])
42
+ for idx in [2,3,5,7,11,12,21,22,23,24]:
43
+ pdf[idx].get_pixmap(matrix=fitz.Matrix(1.5,1.5),alpha=False).save(A/f'slide_{idx+1:02d}.png')
44
+ for group in range(4):
45
+ canvas=Image.new('RGB',(1440,852),'#dddddd');draw=ImageDraw.Draw(canvas)
46
+ for cell,idx in enumerate(range(group*9,min(group*9+9,len(pdf)))):
47
+ pix=pdf[idx].get_pixmap(matrix=fitz.Matrix(.5,.5),alpha=False)
48
+ im=Image.open(io.BytesIO(pix.tobytes('png'))).convert('RGB');im.thumbnail((476,268))
49
+ x=(cell%3)*480;y=(cell//3)*284
50
+ canvas.paste(im,(x,y+16));draw.text((x+5,y+2),f'{idx+1:02d}',fill='black')
51
+ canvas.save(A/f'contact_{group+1}.png')
52
+ manifest['pdf_sha256']=sha(stem.with_suffix('.pdf'));manifest['pdf_status']='PPTX export verified: 34 pages, text bounds checked'
53
+ manifest['originals_preserved']=originals
54
+ (D/'data/results_review_20261013_provenance.json').write_text(json.dumps(manifest,indent=2)+'\n')
55
+ report={'schema':'pbisc.deck-style-verification.v1','slides':34,'pdf_pages':34,
56
+ 'style_renderer':'original generate_deck.py','fonts':sorted(fonts),
57
+ 'reused_original_figure_images_verified':reused,'all_figures_uncropped':True,
58
+ 'source_hashes_verified':len(manifest['sources']),'originals_preserved':len(originals),
59
+ 'shape_and_pdf_text_bounds_pass':True,'pdf_sha256':sha(stem.with_suffix('.pdf')),
60
+ 'pptx_sha256':sha(stem.with_suffix('.pptx')),
61
+ 'new_training':False,'new_generation':False,
62
+ 'underlying_experiment_outputs_changed':False,
63
+ 'derived_metric_update':'Healthy composition: paired absolute errors before donor averaging; win counts added'}
64
+ (A/'VERIFICATION.json').write_text(json.dumps(report,indent=2)+'\n')
65
+ print(json.dumps(report,indent=2))