utils.py

import pandas as pd
import gradio as gr
import csv
import json
import os
import requests
import io
import shutil
from huggingface_hub import Repository

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

TASKS = ["Classification", "VQA", "Retrieval", "Grounding"]

MODEL_INFO = [
    "Rank", "Models", "Model Size(B)", "Data Source",
    "Overall",
    "Classification", "VQA", "Retrieval", "Grounding"
]

BASE_COLS = [col for col in MODEL_INFO if col not in TASKS]

DATA_TITLE_TYPE = ['number', 'markdown', 'str', 'markdown', 'number', 'number', 'number', 'number', 'number', 'number', 'number']

SUBMISSION_NAME = "MMEB"
SUBMISSION_URL = os.path.join("https://huggingface.co/spaces/TIGER-Lab/", SUBMISSION_NAME)
FILE_NAME = "results.csv"
CSV_DIR = "results.csv"

COLUMN_NAMES = MODEL_INFO

LEADERBOARD_INTRODUCTION = """
# MMEB Leaderboard

## Introduction
We introduce a novel benchmark, MMEB (Massive Multimodal Embedding Benchmark), 
which includes 36 datasets spanning four meta-task categories: classification, visual question answering, retrieval, and visual grounding. MMEB provides a comprehensive framework for training
and evaluating embedding models across various combinations of text and image modalities. 
All tasks are reformulated as ranking tasks, where the model follows instructions, processes a query, and selects the correct target from a set of candidates. The query and target can be an image, text,
or a combination of both. MMEB is divided into 20 in-distribution datasets, which can be used for
training, and 16 out-of-distribution datasets, reserved for evaluation.

The detailed explanation of the benchmark and datasets can be found in our paper: https://doi.org/10.48550/arXiv.2410.05160. \n
Github link: https://github.com/TIGER-AI-Lab/VLM2Vec. \n
Overview: https://tiger-ai-lab.github.io/VLM2Vec/. \n
"""

TABLE_INTRODUCTION = """"""

LEADERBOARD_INFO = """
## Dataset Summary
MMEB is organized into four primary meta-task categories:
- **Classification**: This category comprises 5 in-distribution and 5 out-of-distribution datasets. Queries
consist of instructions and images, optionally accompanied by related text. Targets are class labels,
and the number of class labels corresponds to the number of classes in the dataset. \n
        - IND: ImageNet-1k, N24News, HatefulMemes, VOC2007, SUN397 \n
        - OOD: Place365, ImageNet-A, ImageNet-R, ObjectNet, Country-211 \n
- **Visual Question Answering**: This category includes 6 in-distribution and 4 out-of-distribution
datasets. The query consists of an instruction, an image, and a piece of text as the question, while
the target is the answer. Each query has 1,000 target candidates: 1 ground truth and 999 distractors. \n
        - IND: OK-VQA, A-OKVQA, DocVQA, InfographicVQA, ChartQA, Visual7W \n
        - OOD: ScienceQA, VizWiz, GQA, TextVQA \n
- **Information Retrieval**: This category contains 8 in-distribution and 4 out-of-distribution datasets.
Both the query and target sides can involve a combination of text, images, and instructions. Similar
to the VQA task, each query has 1,000 candidates, with 1 ground truth and 999 distractors. \n
        - IND: VisDial, CIRR, VisualNews_t2i, VisualNews_i2t, MSCOCO_t2i, MSCOCO_i2t, NIGHTS, WebQA \n
        - OOD: OVEN, FashionIQ, EDIS, Wiki-SS-NQ \n
- **Visual Grounding**: This category includes 1 in-distribution and 3 out-of-distribution datasets, which are adapted from object detection tasks. Queries consist of an instruction, an image, and text referring to a specific region or object within the image. The target may include a cropped image of the object or text describing the same region. Each query includes 1,000 candidates: 1 ground truth and 999 distractors. These distractors may include hard negatives from the same object class, other objects in the image, or random objects from different images. \n
        - IND: MSCOCO \n
        - OOD: Visual7W-Pointing, RefCOCO, RefCOCO-Matching \n
"""

CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""@article{jiang2024vlm2vec,
  title={VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks},
  author={Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu},
  journal={arXiv preprint arXiv:2410.05160},
  year={2024}
}"""

SUBMIT_INTRODUCTION = """# Submit on MMEB Leaderboard Introduction

## ⚠ Please note that you need to submit the JSON file with the following format:
```json
[
    {
        "Model": "<Model Name>",
        <Optional>"URL": "<Model URL>",
        "Model Size(B)": 1000,
        "Data Source": Self-Reported,
        "Overall": 50.0,
        "Classification": 50.0,
        "VQA": 50.0,
        "Retrieval": 50.0,
        "Grounding": 50.0
    }, 
]
```
You may refer to the Github page for instructions about evaluating your model. 
Github link: https://github.com/TIGER-AI-Lab/VLM2Vec. \n
Please send us an email at [email protected], attaching the JSON file. We will review your submission and update the leaderboard accordingly.
"""

def create_hyperlinked_names(df):
    def convert_url(url, model_name):
        return f'<a href="{url}">{model_name}</a>' if url is not None else model_name

    def add_link_to_model_name(row):
        row['Models'] = convert_url(row['URL'], row['Models'])
        return row
    
    df = df.copy()
    df = df.apply(add_link_to_model_name)
    return df

def fetch_data(file: str) -> pd.DataFrame:
    # fetch the leaderboard data from remote
    if file is None:
        raise ValueError("URL Not Provided")
    url = f"https://huggingface.co/spaces/TIGER-Lab/MMEB/resolve/main/{file}"
    print(f"Fetching data from {url}")
    response = requests.get(url)
    if response.status_code != 200:
        raise requests.HTTPError(f"Failed to fetch data: HTTP status code {response.status_code}")
    return pd.read_csv(io.StringIO(response.text))

def get_urls(file: str='urls.csv') -> dict:
    urls = fetch_data(file)
    return dict(zip(urls['Models'], urls['URL']))

def get_df(file="results.jsonl"):
    df = fetch_data(file)
    df.to_json(file, orient='records', lines=True)
    df['Model Size(B)'] = df['Model Size(B)'].apply(process_model_size)
    df = df.sort_values(by=['Overall'], ascending=False)
    df = create_hyperlinked_names(df)
    df['Rank'] = range(1, len(df) + 1)
    return df

def refresh_data():
    df = get_df()
    return df[COLUMN_NAMES]

def search_and_filter_models(df, query, min_size, max_size):
    filtered_df = df.copy()
    
    if query:
        filtered_df = filtered_df[filtered_df['Models'].str.contains(query, case=False, na=False)]

    size_mask = filtered_df['Model Size(B)'].apply(lambda x: 
        (min_size <= 1000.0 <= max_size) if x == 'unknown' 
        else (min_size <= x <= max_size))
    
    filtered_df = filtered_df[size_mask]
    
    return filtered_df[COLUMN_NAMES]


def search_models(df, query):
    if query:
        return df[df['Models'].str.contains(query, case=False, na=False)]
    return df

def get_size_range(df):
    sizes = df['Model Size(B)'].apply(lambda x: 0.0 if x == 'unknown' else x)
    if (sizes == 0.0).all():
        return 0.0, 1000.0
    return float(sizes.min()), float(sizes.max())


def process_model_size(size):
    if pd.isna(size) or size == 'unk':
        return 'unknown'
    try:
        val = float(size)
        return val
    except (ValueError, TypeError):
        return 'unknown'


def filter_columns_by_tasks(df, selected_tasks=None):
    if selected_tasks is None or len(selected_tasks) == 0:
        return df[COLUMN_NAMES]
    
    base_columns = ['Models', 'Model Size(B)', 'Data Source', 'Overall']
    selected_columns = base_columns + selected_tasks
    
    available_columns = [col for col in selected_columns if col in df.columns]
    return df[available_columns]

def get_task_choices():
    return TASKS