app.py

import gradio as gr
import pandas as pd
import json

from huggingface_hub import hf_hub_url, cached_download, hf_hub_download
from gensim.models import KeyedVectors

# Setup model
json_file = hf_hub_download(repo_id="map222/recipe-spice-model", filename="ingredient_count.json")
with open(json_file, 'r') as json_stream:
    ingredient_count = json.load(json_stream )

w2v_file = hf_hub_download(repo_id="map222/recipe-spice-model", filename="recipe_w2v.gensim")
recipe_w2v = KeyedVectors.load(w2v_file)
print('Loaded w2v')
recipe_tsv = hf_hub_download(repo_id="map222/recipe-spice-model", filename="recipe_NER.tsv")
print('downloaded recipes')
recipe_NER = pd.read_csv(recipe_tsv, sep='\t' )

def calc_cooccurrence(ingredient: str,
                      candidates,
                      recipes):
  ''' Calc how often the top ingredient co-occurs with the candidates
    - also removes candidates that are re-phrase of ingredient (e.g. "beef" and "ground beef")
    ingredient: str name of an ingredient ("apple")
    candidates: potential other ingredients ("orange")
    recipes: iterable of possible ingredients
  '''


  co_count = {}
  for candidate in candidates:
    co_count[candidate] = sum([candidate in recipe and ingredient in recipe for recipe in recipes])
  return co_count

def get_fusion_ingredients(ingredient: str,
                           recipe_model, #gensim model
                           recipes, #iterable of recipes
                           ingredient_count: dict,
                           max_candidates = 20,
                           min_occurence_factor = 100 # minimum number of recipes an ingredient has to be in
                           ):
    print(recipes.head() )
    print(recipes.info() )
    print(recipes.apply(lambda row: ingredient in row).head() )
    ingredient_recipes = recipes.loc[recipes.apply(lambda row: ingredient in row)]
  
    ingredient_candidates = recipe_model.wv.most_similar(ingredient, topn=50) # get top similar ingredients
    candidate_names = list(zip(*ingredient_candidates))[0]
    pruned_candidates = [candidate for candidate in candidate_names if ingredient not in candidate][:max_candidates] # clean up candidates to remove duplicates (e.g. "gala apple")
    cooccurrence_counts = calc_cooccurrence(ingredient, candidate_names, ingredient_recipes) # get counts for normalization
    # final score for sorting: similarity / how often co-occur / total occurences
    min_occurences = max(cooccurrence_counts.values()) / min_occurence_factor
    freq_norm_candidates = {candidate[0]: candidate[1] / (cooccurrence_counts[candidate[0]]+1) / ingredient_count[candidate[0]] for candidate in ingredient_candidates if candidate[0] in pruned_candidates and cooccurrence_counts[candidate[0]] > min_occurences}
    top_candidates = sorted([(k,v) for k,v in freq_norm_candidates.items()], key=lambda x: x[1])[-5:]
    return top_candidates, cooccurrence_counts, pruned_candidates # return multiple for debugging

def helper_func(text):
    spicy_candidates, cooccurrence_counts, most_similar = get_fusion_ingredients(text, recipe_w2v, recipe_NER['NER'], ingredient_count )
    print('ran similarity')
    spicy_df = pd.DataFrame(spicy_candidates, columns = ["Spicy ingredient", "spiciness"]).iloc[::-1]
    print('in between')
    top_df = pd.DataFrame(most_similar, columns = ["Top ingredient"])
    print('made dataframes')
    return spicy_df, top_df
    
app = gr.Blocks()

with app:
    gr.Markdown("# Recipe Spice")
    gr.Markdown(
        """
        This model uses Word2Vec trained on a 200k recipe corpus to generate ingredient similarities. Then it finds the "spiciest"
        ingredient by finding other ingredients that are similar, but don't occur together often. Enter an ingredient below;
        for output, the left column shows the "spiciest" ingredients, and the right column shows the closest.
        """
        )
    with gr.Row():
        text_in = gr.Textbox(lines=1, placeholder="Ingredient", label="Find something fun!")

    with gr.Row():
        with gr.Column():
            gr.Markdown("Spicy ingredients")
            spicy_df = gr.Dataframe(interactive=False)
        with gr.Column():
            gr.Markdown("Top ingredients")
            top_df = gr.Dataframe(interactive=False)
    text_in.submit(helper_func, inputs=[text_in], outputs=[spicy_df, top_df])

app.launch()