app.py

from enum import Enum
import os
import re
import aiohttp
import requests
import json
import subprocess
import asyncio
from io import BytesIO
import uuid

from math import ceil
from tqdm import tqdm
from pathlib import Path
from huggingface_hub import Repository
from PIL import Image, ImageOps
from fastapi import FastAPI, BackgroundTasks
from fastapi_utils.tasks import repeat_every
from fastapi.middleware.cors import CORSMiddleware
import boto3

from db import Database

AWS_ACCESS_KEY_ID = os.getenv('MY_AWS_ACCESS_KEY_ID')
AWS_SECRET_KEY = os.getenv('MY_AWS_SECRET_KEY')
AWS_S3_BUCKET_NAME = os.getenv('MY_AWS_S3_BUCKET_NAME')


HF_TOKEN = os.environ.get("HF_TOKEN")

S3_DATA_FOLDER = Path("sd-multiplayer-data")

DB_FOLDER = Path("diffusers-gallery-data")

CLASSIFIER_URL = "https://radames-aesthetic-style-nsfw-classifier.hf.space/run/inference"
ASSETS_URL = "https://d26smi9133w0oo.cloudfront.net/diffusers-gallery/"


s3 = boto3.client(service_name='s3',
                  aws_access_key_id=AWS_ACCESS_KEY_ID,
                  aws_secret_access_key=AWS_SECRET_KEY)


repo = Repository(
    local_dir=DB_FOLDER,
    repo_type="dataset",
    clone_from="huggingface-projects/diffusers-gallery-data",
    use_auth_token=True,
)
repo.git_pull()

database = Database(DB_FOLDER)


async def upload_resize_image_url(session, image_url):
    print(f"Uploading image {image_url}")
    async with session.get(image_url) as response:
        if response.status == 200 and response.headers['content-type'].startswith('image'):
            image = Image.open(BytesIO(await response.read())).convert('RGB')
            # resize image proportional
            image = ImageOps.fit(image, (400, 400), Image.LANCZOS)
            image_bytes = BytesIO()
            image.save(image_bytes, format="JPEG")
            image_bytes.seek(0)
            fname = f'{uuid.uuid4()}.jpg'
            s3.upload_fileobj(Fileobj=image_bytes, Bucket=AWS_S3_BUCKET_NAME, Key="diffusers-gallery/" + fname,
                              ExtraArgs={"ContentType": "image/jpeg", "CacheControl": "max-age=31536000"})
            return fname
    return None


def fetch_models(page=0):
    response = requests.get(
        f'https://huggingface.co/models-json?pipeline_tag=text-to-image&p={page}')
    data = response.json()
    return {
        "models": [model for model in data['models'] if not model['private']],
        "numItemsPerPage": data['numItemsPerPage'],
        "numTotalItems": data['numTotalItems'],
        "pageIndex": data['pageIndex']
    }


def fetch_model_card(model_id):
    response = requests.get(
        f'https://huggingface.co/{model_id}/raw/main/README.md')
    return response.text


async def find_image_in_model_card(text):
    image_regex = re.compile(r'https?://\S+(?:png|jpg|jpeg|webp)')
    urls = re.findall(image_regex, text)
    if not urls:
        return []

    async with aiohttp.ClientSession() as session:
        tasks = [asyncio.ensure_future(upload_resize_image_url(
            session, image_url)) for image_url in urls[0:3]]
        return await asyncio.gather(*tasks)


def run_classifier(images):
    images = [i for i in images if i is not None]
    if len(images) > 0:
        # classifying only the first image
        images_urls = [ASSETS_URL + images[0]]
        response = requests.post(CLASSIFIER_URL, json={"data": [
            {"urls": images_urls},  # json urls: list of images urls
            False,  # enable/disable gallery image output
            None,  # single image input
            None,  # files input
        ]}).json()

        # data response is array data:[[{img0}, {img1}, {img2}...], Label, Gallery],
        class_data = response['data'][0][0]
        print(class_data)
        class_data_parsed = {row['label']: round(
            row['score'], 3) for row in class_data}

        # update row data with classificator data
        return class_data_parsed
    else:
        return {}


async def get_all_new_models():
    initial = fetch_models(0)
    num_pages = ceil(initial['numTotalItems'] / initial['numItemsPerPage'])

    print(
        f"Total items: {initial['numTotalItems']} - Items per page: {initial['numItemsPerPage']}")
    print(f"Found {num_pages} pages")

    # fetch all models
    new_models = []
    for page in tqdm(range(0, num_pages)):
        print(f"Fetching page {page} of {num_pages}")
        page_models = fetch_models(page)
        new_models += page_models['models']
    return new_models


async def sync_data():
    print("Fetching models")
    new_models = await get_all_new_models()
    print(f"Found {len(new_models)} models")
    # save list of all models for ids
    with open(DB_FOLDER / "models.json", "w") as f:
        json.dump(new_models, f)
    # with open(DB_FOLDER / "models.json", "r") as f:
    #     new_models = json.load(f)

    new_models_ids = [model['id'] for model in new_models]

    # get existing models
    with database.get_db() as db:
        cursor = db.cursor()
        cursor.execute("SELECT id FROM models")
        existing_models = [row['id'] for row in cursor.fetchall()]
    models_ids_to_add = list(set(new_models_ids) - set(existing_models))
    # find all models id to add from new_models
    models = [model for model in new_models if model['id'] in models_ids_to_add]

    print(f"Found {len(models)} new models")
    for model in tqdm(models):
        model_id = model['id']
        model_card = fetch_model_card(model_id)
        images = await find_image_in_model_card(model_card)
        classifier = run_classifier(images)
        # update model row with image and classifier data
        with database.get_db() as db:
            cursor = db.cursor()
            cursor.execute("INSERT INTO models(id, data) VALUES (?, ?)",
                           [model_id, json.dumps({
                               **model,
                               "images": images,
                               "class": classifier
                           })])
            db.commit()

    if (len(models) > 0):
        print("Updating repository")
        subprocess.Popen(
            "git add . && git commit --amend -m 'update' && git push --force", cwd=DB_FOLDER, shell=True)
    else:
        print("No new models found")

app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


# @ app.get("/sync")
# async def sync(background_tasks: BackgroundTasks):
#     await sync_data()
#     return "Synced data to huggingface datasets"


MAX_PAGE_SIZE = 30


class Sort(str, Enum):
    trending = "trending"
    recent = "recent"
    likes = "likes"


@ app.get("/api/models")
def get_page(page: int = 1, sort: Sort = Sort.trending):
    page = page if page > 0 else 1
    if sort == Sort.trending:
        sort_query = "json_extract(data, '$.likes') / POW((JULIANDAY('now') - JULIANDAY(datetime(json_extract(data, '$.lastModified')))) + 2, 2) DESC"
    elif sort == Sort.recent:
        sort_query = "datetime(json_extract(data, '$.lastModified')) DESC"
    elif sort == Sort.likes:
        sort_query = "json_extract(data, '$.likes') DESC"

    with database.get_db() as db:
        cursor = db.cursor()
        cursor.execute(f"""
            SELECT *, COUNT(*) OVER() AS total
            FROM models
            WHERE json_extract(data, '$.likes') > 4
            ORDER BY {sort_query}
            LIMIT {MAX_PAGE_SIZE} OFFSET {(page - 1) * MAX_PAGE_SIZE}
        """)
        results = cursor.fetchall()
        total = results[0]['total'] if results else 0
        total_pages = (total + MAX_PAGE_SIZE - 1) // MAX_PAGE_SIZE

    return {
        "models": [json.loads(result['data']) for result in results],
        "totalPages": total_pages
    }


@app.get("/")
def read_root():
    return "Just a bot to sync data from diffusers gallery"


# @app.on_event("startup")
# @repeat_every(seconds=60 * 60 * 24, wait_first=True)
# async def repeat_sync():
#     await sync_data()
#     return "Synced data to huggingface datasets"