grocery-next-model / example.py
barrydeen's picture
Publish Grocery Next synthetic dataset and model
fa572b1 verified
Raw History Blame Contribute Delete
1.4 kB
#!/usr/bin/env python3
"""Example for the Grocery Next model repository.
The model ranks complementary grocery items by smoothed co-occurrence.
Without a popularity penalty the ubiquitous staples (onions, garlic,
tomatoes, ...) dominate every answer, so the recommended default is
``popularity_penalty=8``. Pass 0 to see the raw association estimate.
"""
from grocery import GroceryModel
PENALTY = 8.0 # 0 = raw estimate, 8 = CLI default, higher = emphasize rarity
model = GroceryModel.load("model.json")
# Recommended: results track the selected items.
for selected in (
["tomatoes", "onions"],
["jasmine rice", "green curry paste"],
["injera", "berbere"],
):
print(f"\nSelected: {', '.join(selected)} (popularity_penalty={PENALTY})")
for rank, prediction in enumerate(
model.predict(selected, top_k=5, popularity_penalty=PENALTY), start=1
):
print(f"{rank}. {prediction.item} ({prediction.score:.1%})")
# Why the penalty exists: the same input, raw vs. penalized.
selected = ["injera", "berbere"]
print(f"\nSame input, raw (popularity_penalty=0) vs. penalized ({PENALTY}):")
raw = model.predict(selected, top_k=5, popularity_penalty=0.0)
pen = model.predict(selected, top_k=5, popularity_penalty=PENALTY)
print(" raw: " + ", ".join(f"{p.item} {p.score:.1%}" for p in raw))
print(" penalized: " + ", ".join(f"{p.item} {p.score:.1%}" for p in pen))