לנפות 1 מ'

  • תיאור :

הטבעות מאומנות מראש לחיפוש משוער של השכנים הקרובים ביותר באמצעות המרחק האוקלידי. מערך נתונים זה מורכב משני פיצולים:

  1. 'בסיס נתונים': מורכב מ-1,000,000 נקודות נתונים, לכל אחת יש תכונות: 'הטמעה' (128 צפים), 'אינדקס' (int64), 'שכנים' (רשימה ריקה).
  2. 'מבחן': מורכב מ-10,000 נקודות נתונים, לכל אחת מהן תכונות: 'הטבעה' (128 צפים), 'אינדקס' (int64), 'שכנים' (רשימת 'אינדקס' ו'מרחק' של השכנים הקרובים ביותר במסד הנתונים. )
לְפַצֵל דוגמאות
'database' 1,000,000
'test' 10,000
  • מבנה תכונה :
FeaturesDict({
   
'embedding': Tensor(shape=(128,), dtype=float32),
   
'index': Scalar(shape=(), dtype=int64, description=Index within the split.),
   
'neighbors': Sequence({
       
'distance': Scalar(shape=(), dtype=float32, description=Neighbor distance.),
       
'index': Scalar(shape=(), dtype=int64, description=Neighbor index.),
   
}),
})
  • תיעוד תכונה :
תכונה מַחלָקָה צוּרָה Dtype תֵאוּר
FeaturesDict
הטבעה מוֹתֵחַ (128,) לצוף32
מַדָד סקלר int64 אינדקס בתוך הפיצול.
שכנים רֶצֶף השכנים המחושבים, שזמינים רק לפיצול המבחן.
שכנים/מרחק סקלר לצוף32 מרחק שכן.
שכנים/אינדקס סקלר int64 מדד השכנים.
  • ציטוט :
@article{jegou2010product,
  title
={Product quantization for nearest neighbor search},
  author
={Jegou, Herve and Douze, Matthijs and Schmid, Cordelia},
  journal
={IEEE transactions on pattern analysis and machine intelligence},
  volume
={33},
  number
={1},
  pages
={117--128},
  year
={2010},
  publisher
={IEEE}
}