app.py

import torch
import json
import gradio as gr

TITLE = "Danboru Tag Similarity"
DESCRIPTION = """
与えられたダンボールタグの類似度を計算します。\n
対応するタグのリストはFilesからそれぞれのテキストファイルを参照してください。(Dartと同じです）。\n
Dartを参考に、isek-ai/danbooru-tags-2023データセットでタグをシャッフルして2エポック学習しました。\n
学習後のトークン埋め込みを元に計算しています。
"""

with open("num_to_token.json", "r") as f:
    num_to_token = json.load(f)
with open("popular.txt", "r") as f:
    populars = f.read().splitlines()
with open("character.txt", "r") as f:
    characters = f.read().splitlines()
    characters_populars = list(set(characters) & set(populars))
with open("copyright.txt", "r") as f:
    copyrights = f.read().splitlines()
    copyrights_populars = list(set(copyrights) & set(populars))
with open("general.txt", "r") as f:
    generals = f.read().splitlines()
    generals_populars = list(set(generals) & set(populars))

    
token_to_num = {v:k for k,v in num_to_token.items()}
token_embeddings = torch.load("token_embeddings.pt")

tags = sorted(list(num_to_token.values()))

def predict(target_tag, sort_by, category, popular):
    if sort_by == "descending":
        multiplier = 1
    else:
        multiplier = -1
    target_embedding = token_embeddings[int(token_to_num[target_tag])].unsqueeze(0)
    sims = torch.cosine_similarity(target_embedding, token_embeddings, dim=1)
    results = {num_to_token[str(i)]:sims[i].item() * multiplier for i in range(len(num_to_token))}
    if category == "general":
        tag_list = generals if popular == "all" else generals_populars
    elif category == "character":
        tag_list = characters if popular == "all" else characters_populars
    elif category == "copyright":
        tag_list = copyrights if popular == "all" else copyrights_populars
    else:
        tag_list = results.keys() if popular == "all" else populars
        
    return {k:results[k] for k in tag_list}

demo = gr.Interface(
        fn=predict, 
        inputs=[
            gr.Textbox(label="Target tag", value="otoko no ko"),
            gr.Radio(choices=["descending", "ascending"], label="Sort by", value="descending"),
            gr.Dropdown(choices=["all", "general", "character", "copyright"], value="all", label="category"),
            gr.Radio(choices=["all", "only_popular"], label="Only popular tag (count>=1000)", value="all"),
        ],
        outputs=gr.Label(num_top_classes=50),
        title=TITLE,
        description=DESCRIPTION
    )

demo.launch()