app.py

import gradio as gr
from faiss import IndexFlatIP
import pandas as pd
import numpy as np
from transformers import AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased")
input_embeddings = np.load("bert_input_embeddings.npy")
index = IndexFlatIP(input_embeddings.shape[-1])  
index.add(input_embeddings)
vocab = {v:k for k,v in tokenizer.vocab.items()}
lookup_table = pd.Series(vocab).sort_index()

def get_first_subword(word):
    try:
        return tokenizer.vocab[word]
    except:
        return tokenizer(word, add_special_tokens=False)['input_ids'][0]

def search(token_to_lookup, num_neighbors=50):
    i = get_first_subword(token_to_lookup)
    _ , I = index.search(input_embeddings[i:i+1], num_neighbors)
    hits = lookup_table.take(I[0])
    results = hits.values[1:]
    return [r for r in results if not "##" in r], [[r for r in results if "##" in r]]


iface = gr.Interface(
    fn=search,
    
    #inputs=[gr.Textbox(lines=1, label="Vocabulary Token", placeholder="Enter token..."), gr.Number(value=50, label="number of neighbors")],
    inputs=gr.Textbox(lines=1, label="Vocabulary Token", placeholder="Enter token..."),
    outputs=[gr.Textbox(label="Nearest tokens"), gr.Textbox(label="Nearest subwords")],
    examples=[
        ["##logy"],
        ["##ness"],
        ["##nity"],
        ["responded"],
        ["queen"],
        ["king"],
        ["hospital"],
        ["disease"],
        ["grammar"],
        ["philosophy"],
        ["aristotle"],
        ["##ting"],
        ["woman"],
        ["man"]
    ],
)
iface.launch(enable_queue=True, debug=True, show_error=True)