app.py

import os
import platform
import re
from collections import defaultdict

import gradio as gr
from cachetools import TTLCache, cached
from cytoolz import groupby
from huggingface_hub import CollectionItem, get_collection, list_datasets, list_models
from tqdm.auto import tqdm
from apscheduler.schedulers.background import BackgroundScheduler
from apscheduler.triggers.cron import CronTrigger

os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
is_macos = platform.system() == "Darwin"
local = platform.system() == "Darwin"
LIMIT = 1000 if is_macos else None  # limit for local dev because slooow internet
CACHE_TIME = 60 * 15  # 15 minutes


@cached(cache=TTLCache(maxsize=100, ttl=CACHE_TIME))
def get_models():
    print("getting models...")
    return list(tqdm(iter(list_models(full=True, limit=LIMIT))))


@cached(cache=TTLCache(maxsize=100, ttl=CACHE_TIME))
def get_datasets():
    print("getting datasets...")
    return list(tqdm(iter(list_datasets(full=True, limit=LIMIT))))


get_models()  # warm up the cache
get_datasets()  # warm up the cache


def check_for_arxiv_id(model):
    return [tag for tag in model.tags if "arxiv" in tag] if model.tags else False


def extract_arxiv_id(input_string: str) -> str:
    pattern = re.compile(r"\barxiv:(\d+\.\d+)\b")
    match = pattern.search(input_string)
    return match[1] if match else None


@cached(cache=TTLCache(maxsize=100, ttl=CACHE_TIME))
def create_model_to_arxiv_id_dict():
    models = get_models()
    model_to_arxiv_id = {}
    for model in models:
        if arxiv_papers := check_for_arxiv_id(model):
            clean_arxiv_ids = []
            for paper in arxiv_papers:
                if arxiv_id := extract_arxiv_id(paper):
                    clean_arxiv_ids.append(arxiv_id)
            model_to_arxiv_id[model.modelId] = clean_arxiv_ids
    return model_to_arxiv_id


@cached(cache=TTLCache(maxsize=100, ttl=CACHE_TIME))
def create_dataset_to_arxiv_id_dict():
    datasets = get_datasets()
    dataset_to_arxiv_id = {}
    for dataset in datasets:
        if arxiv_papers := check_for_arxiv_id(dataset):
            clean_arxiv_ids = []
            for paper in arxiv_papers:
                if arxiv_id := extract_arxiv_id(paper):
                    clean_arxiv_ids.append(arxiv_id)
            dataset_to_arxiv_id[dataset.id] = clean_arxiv_ids
    return dataset_to_arxiv_id


def get_collection_type(collection_item: CollectionItem):
    try:
        return f"{collection_item.item_type}s"
    except AttributeError:
        return None


def group_collection_items(collection_slug: str):
    collection = get_collection(collection_slug)
    items = collection.items
    return groupby(get_collection_type, items)


@cached(cache=TTLCache(maxsize=500, ttl=CACHE_TIME))
def get_papers_for_collection(collection_slug: str):
    dataset_to_arxiv_id = create_dataset_to_arxiv_id_dict()
    models_to_arxiv_id = create_model_to_arxiv_id_dict()
    collection = group_collection_items(collection_slug)
    collection_datasets = collection.get("datasets", None)
    collection_models = collection.get("models", None)
    papers = collection.get("papers", None)
    dataset_papers = defaultdict(dict)
    model_papers = defaultdict(dict)
    collection_papers = defaultdict(dict)
    if collection_datasets is not None:
        for dataset in collection_datasets:
            if arxiv_ids := dataset_to_arxiv_id.get(dataset.item_id, None):
                data = {
                    "arxiv_ids": arxiv_ids,
                    "hub_paper_links": [
                        f"https://huggingface.co/papers/{arxiv_id}"
                        for arxiv_id in arxiv_ids
                    ],
                }
                dataset_papers[dataset.item_id] = data
    if collection_models is not None:
        for model in collection.get("models", []):
            if arxiv_ids := models_to_arxiv_id.get(model.item_id, None):
                data = {
                    "arxiv_ids": arxiv_ids,
                    "hub_paper_links": [
                        f"https://huggingface.co/papers/{arxiv_id}"
                        for arxiv_id in arxiv_ids
                    ],
                }
                model_papers[model.item_id] = data
    if papers is not None:
        for paper in papers:
            data = {
                "arxiv_ids": paper.item_id,
                "hub_paper_links": [f"https://huggingface.co/papers/{paper.item_id}"],
            }
            collection_papers[paper.item_id] = data
    if not dataset_papers:
        dataset_papers = None
    if not model_papers:
        model_papers = None
    if not collection_papers:
        collection_papers = None
    return {
        "dataset papers": dataset_papers,
        "model papers": model_papers,
        "papers": collection_papers,
    }


scheduler = BackgroundScheduler()
scheduler.add_job(get_datasets, "interval", minutes=15)
scheduler.add_job(get_models, "interval", minutes=15)
scheduler.start()

placeholder_url = "HF-IA-archiving/models-to-archive-65006a7fdadb8c628f33aac9"
slug_input = gr.Textbox(
    placeholder=placeholder_url, interactive=True, label="Collection slug", max_lines=1
)
description = (
    "Enter a Collection slug to get the arXiv IDs and Hugging Face Paper links for"
    " papers associated with models and datasets in the collection. If the collection"
    " includes papers the arXiv IDs and Hugging Face Paper links will be returned for"
    " those papers as well."
)

examples = [
    placeholder_url,
    "davanstrien/historic-language-modeling-64f99e243188ade79d7ad74b",
]


gr.Interface(
    get_papers_for_collection,
    slug_input,
    "json",
    title="📄🔗: Extract linked papers from a Hugging Face Collection",
    description=description,
    examples=examples,
    cache_examples=True,
).queue(concurrency_count=4).launch()