| """Reader-facing examples from preserved comprehension records; no inference.""" | |
| from pathlib import Path | |
| import html | |
| import json | |
| import numpy as np | |
| import matplotlib.pyplot as plt | |
| from PIL import Image as PILImage, ImageDraw, ImageFont, features | |
| from IPython.display import display, HTML, Markdown | |
| from editorial_views import finish, panels, COLORS | |
| BASE=Path(__file__).resolve().parents[1] | |
| def _table(columns, rows, widths=None): | |
| """Native notebook HTML table with wrapped, explicitly separated fields.""" | |
| def value(text): | |
| return html.escape(str(text)).replace('\n','<br>') | |
| head=''.join('<th style="text-align:left;vertical-align:top;padding:10px;line-height:1.5">' | |
| +value(c)+'</th>' for c in columns) | |
| body=''.join('<tr>'+''.join( | |
| '<td style="text-align:left;vertical-align:top;white-space:normal;' | |
| 'padding:10px;line-height:1.5;overflow-wrap:anywhere;'+ | |
| (f'width:{widths[i]}%;' if widths else '')+'">'+value(v)+'</td>' | |
| for i,v in enumerate(row))+'</tr>' for row in rows) | |
| cols=('<colgroup>'+''.join(f'<col style="width:{w}%">' for w in widths)+'</colgroup>') if widths else '' | |
| display(HTML('<table style="width:100%;table-layout:fixed;text-align:left">'+cols+ | |
| '<thead><tr>'+head+'</tr></thead><tbody>'+body+'</tbody></table>')) | |
| def text_only_example(): | |
| """One identical item, three languages, with properly shaped Devanagari.""" | |
| records=[json.loads(line) for line in | |
| (BASE/'data/backbone/predictions.jsonl').read_text().splitlines()] | |
| order=[('english','English'),('hinglish','Roman Hinglish'),('hindi','Hindi')] | |
| assert features.check('raqm'), 'Hindi rendering requires Pillow RAQM' | |
| dpi=200 | |
| width,height=2520,1000 | |
| canvas=PILImage.new('RGB',(width,height),'white') | |
| draw=ImageDraw.Draw(canvas) | |
| fontpath=str(BASE/'notebook_backend/fonts/NotoSansDevanagari.ttf') | |
| def font(size,bold=False): | |
| f=ImageFont.truetype(fontpath,size,layout_engine=ImageFont.Layout.RAQM) | |
| if bold:f.set_variation_by_name('Bold') | |
| return f | |
| normal=font(33); bold=font(34,True); heading=font(42,True) | |
| def write(x,y,text,f,fill,maxwidth=730,lineheight=53): | |
| for rawline in text.split('\n'): | |
| line='' | |
| for word in rawline.split(' '): | |
| candidate=(line+' '+word).strip() | |
| if line and draw.textlength(candidate,font=f)>maxwidth: | |
| draw.text((x,y),line,font=f,fill=fill,anchor='lt');y+=lineheight | |
| line=word | |
| else:line=candidate | |
| draw.text((x,y),line,font=f,fill=fill,anchor='lt');y+=lineheight | |
| return y | |
| selected=[] | |
| for i,(language,label) in enumerate(order): | |
| r=next(r for r in records if r['id']=='spatial_1' and r['language']==language) | |
| assert r['answer']=='D' | |
| selected.append(r) | |
| left=18+i*840 | |
| color=COLORS[language] | |
| draw.rounded_rectangle((left,8,left+804,height-14),radius=12, | |
| fill='white',outline='#cbd5dc',width=2) | |
| draw.rectangle((left+2,10,left+802,99),fill=color) | |
| write(left+25,29,label,heading,'white') | |
| write(left+25,130,'Exact input',bold,'#233442') | |
| bottom=write(left+25,193,r['prompt'],normal,'#233442') | |
| assert bottom<700, (language,bottom) | |
| draw.line((left+25,721,left+780,721),fill='#cbd5dc',width=2) | |
| write(left+25,748,'Model output: '+r['response'],bold,color) | |
| answer={'english':'D. inside','hinglish':'D. andar','hindi':'D. अंदर'}[language] | |
| write(left+25,811,'Expected: '+answer,normal,'#233442') | |
| correct=r['response']==r['answer'] | |
| write(left+25,874,'Result: '+('Correct' if correct else 'Incorrect'),bold, | |
| '#327052' if correct else '#a14747') | |
| fig,ax=plt.subplots(figsize=(width/dpi,height/dpi),layout='constrained') | |
| ax.imshow(np.asarray(canvas));ax.axis('off') | |
| finish(fig,'libero_text_only_triplet') | |
| def outline_examples(): | |
| records=json.loads((BASE/'data/visual_exploration/selected_responses.json').read_text()) | |
| lookup={r['key']:r for r in records} | |
| def image_paths(key): | |
| return [str(BASE/'data/visual_exploration'/p) for p in lookup[key]['image_files']] | |
| def exact_prompt(prompt): | |
| display(HTML('<p><strong>Exact input prompt</strong></p><blockquote style="white-space:pre-wrap">' | |
| +html.escape(prompt)+'</blockquote>')) | |
| display(Markdown('##### Single-object identification\n\n' | |
| 'The same prompt was used with each of these three distinct annotated images. ' | |
| 'Result assesses the identity of the outlined object.')) | |
| keys=['stage4_agentview_plate','stage4_agentview_wine_bottle','stage4_wrist_wine_bottle'] | |
| prompt=lookup[keys[0]]['prompt'] | |
| assert all(lookup[k]['prompt']==prompt for k in keys) | |
| exact_prompt(prompt) | |
| titles=['Plate / external camera','Bottle / external camera','Bottle / wrist camera'] | |
| panels([image_paths(k)[0] for k in keys],titles, | |
| name='outline_object_inputs',width=12,image_height=3) | |
| _table(['Input image','Expected','Predicted (excerpt)','Result'],[ | |
| (titles[0],'Plate','“white, round plate”','Correct'), | |
| (titles[1],'Wine bottle','“dark-colored wine bottle”','Correct'), | |
| (titles[2],'Wine bottle','“ping pong ball”','Incorrect') | |
| ],[24,20,40,16]) | |
| display(Markdown('##### Category-cued recognition\n\n' | |
| 'Result assesses recognition of the food-package category and its position relative to the bowl. ' | |
| 'The prompt supplies the category names.')) | |
| exact_prompt(lookup['stage2_wrist']['prompt']) | |
| # Use only the supplied wrist image; no second camera was provided for this probe. | |
| panels(image_paths('stage2_wrist'),['Wrist-camera input'], | |
| name='category_cued_input',width=5,image_height=3.3) | |
| _table(['Expected','Predicted (excerpt)','Result'],[ | |
| ('Packaged food to the right of the bowl', | |
| '“rectangular blue package with a white label”; “to the right of the bowl”','Correct') | |
| ],[36,48,16]) | |
Xet Storage Details
- Size:
- 6.19 kB
- Xet hash:
- a994d6824aaa22d33a5ece80c5e7f3f235086c8b7a902a0aa615d63f7f569d03
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.