gen_table.py

import copy as cp
import json
from collections import defaultdict
from urllib.request import urlopen

import gradio as gr
import numpy as np
import pandas as pd

from meta_data import META_FIELDS, NON_MCQ_DATASETS, URL, DATASETS_ALL, DATASETS_ESS


def listinstr(lst, s):
    assert isinstance(lst, list)
    for item in lst:
        if item in s:
            return True
    return False


def upper_key(k):
    if k == 'ocr':
        return 'OCR'
    elif '_' in k:
        k = k.split('_')
        k = [x[0].upper() + x[1:] for x in k]
        k = ' '.join(k)
        return k
    else:
        return k
    

PREV_LOADING_TIME = None
PREV_RESULTS = None

def load_results():
    import time
    global PREV_LOADING_TIME, PREV_RESULTS
    if PREV_LOADING_TIME is not None and time.time() - PREV_LOADING_TIME < 600:
        return PREV_RESULTS
    else:
        data = json.loads(urlopen(URL).read())
        PREV_LOADING_TIME = time.time()
        PREV_RESULTS = data
        return data


def nth_large(val, vals):
    return sum([1 for v in vals if v > val]) + 1


def model_size_flag(sz, FIELDS):
    if (pd.isna(sz) or sz == '') and 'Unknown' in FIELDS:
        return True
    if (pd.isna(sz) or sz == ''):
        return False
    sz = int(sz)
    if '<4B' in FIELDS and sz < 4:
        return True
    if '4B-10B' in FIELDS and sz >= 4 and sz < 10:
        return True
    if '10B-20B' in FIELDS and sz >= 10 and sz < 20:
        return True
    if '20B-40B' in FIELDS and sz >= 20 and sz < 40:
        return True
    if '>40B' in FIELDS and sz >= 40:
        return True
    return False


def model_type_flag(line, FIELDS):
    if 'OpenSource' in FIELDS and line['OpenSource'] == 'Yes':
        return True
    if 'API' in FIELDS and line['OpenSource'] == 'No':  
        return True 
    return False


def BUILD_L1_DF(results, circular=False):
    check_box = {}
    check_box['essential'] = ['Method', 'Org', 'Param (B)', 'Language Model', 'Vision Model']
    # revise there to set default dataset
    check_box['required'] = ['Overall'] + DATASETS_ESS
    check_box['all'] = ['Overall'] + DATASETS_ALL 
    type_map = defaultdict(lambda: 'number')
    type_map['Method'] = 'html'
    type_map['Language Model'] = type_map['Vision Model'] = type_map['Org'] = 'html'
    type_map['OpenSource'] = type_map['Verified'] = 'str'
    check_box['type_map'] = type_map
    df = generate_table(results, circular=circular)
    return df, check_box


def BUILD_L2_DF(results, dataset, circular=False):
    res = defaultdict(list)
    sub = [v for v in results.values() if dataset in v]
    assert len(sub), dataset
    fields = list(sub[0][dataset].keys())

    non_overall_fields = [x for x in fields if 'Overall' not in x]
    overall_fields = [x for x in fields if 'Overall' in x]

    for m in results:
        item = results[m]
        if dataset not in item:
            continue
        for k in META_FIELDS:
            if k == 'Param (B)':
                param = item['META']['Parameters']
                res[k].append(float(param.replace('B', '')) if param != '' else '')
            elif k == 'Method':
                name, url = item['META']['Method']
                res[k].append(f'<a href="{url}">{name}</a>')
            else:
                s = item['META'][k].replace('\n', '<br>')
                s = s.replace(' & ', '<br>')
                res[k].append(s)

        db_name = dataset + '_circular' if circular else dataset
        for d in overall_fields:
            res[d].append(float(item[db_name][d]))
        for d in non_overall_fields:
            res[d].append(float(item[db_name][d]))
        
    df = pd.DataFrame(res)
    all_fields = overall_fields + non_overall_fields
    # Use the first 5 non-overall fields as required fields
    required_fields = overall_fields if len(overall_fields) else non_overall_fields[:5]

    df = df.sort_values('Overall')
    df = df.iloc[::-1]

    check_box = {}
    check_box['essential'] = ['Method', 'Org', 'Param (B)', 'Language Model', 'Vision Model']
    check_box['required'] = required_fields
    check_box['all'] = all_fields
    type_map = defaultdict(lambda: 'number')
    type_map['Method'] = 'html'
    type_map['Language Model'] = type_map['Vision Model'] = type_map['Org'] = 'html'
    type_map['OpenSource'] = type_map['Verified'] = 'str'
    check_box['type_map'] = type_map
    return df, check_box


def generate_table(results, circular=False):

    res = defaultdict(list)
    for i, m in enumerate(results):
        item = results[m]
        avg = 0
        for k in META_FIELDS:
            if k == 'Param (B)':
                param = item['META']['Parameters']
                res[k].append(float(param.replace('B', '')) if param != '' else '')
            elif k == 'Method':
                name, url = item['META']['Method']
                res[k].append(f'<a href="{url}">{name}</a>')
            else:
                s = item['META'][k].replace('\n', '<br>')
                s = s.replace(' & ', '<br>')
                res[k].append(s)

        num_dataset = 0
        for d in DATASETS_ALL:
            key_name = 'Overall'
            db_name = d + '_circular' if circular else d
            if circular and d in NON_MCQ_DATASETS:
                continue

            if db_name in item:
                val = float(item[db_name][key_name])
                val = float(f'{val:.1f}')
                res[d].append(val)
            else:
                res[d].append(None)
            if d in DATASETS_ESS:
                if d in item and avg is not None:
                    avg += res[d][-1]
                    num_dataset += 1
                else:
                    avg = None

        if avg is not None:
            avg = float(f'{avg / num_dataset:.1f}')
            
        res['Overall'].append(avg)

    df = pd.DataFrame(res)
    overall_isna = df[pd.isna(df['Overall'])]
    overall_notna = df[~pd.isna(df['Overall'])]
    overall_notna = overall_notna.sort_values('Overall')
    overall_notna = overall_notna.iloc[::-1]
    overall_isna = overall_isna.sort_values('3DSRBench')
    overall_isna = overall_isna.iloc[::-1]
    df = pd.concat([overall_notna, overall_isna])   
    return df