File size: 8,672 Bytes
0f5b104 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 |
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 DEFAULT_BENCH, META_FIELDS, URL
def listinstr(lst, s):
assert isinstance(lst, list)
for item in lst:
if item in s:
return True
return False
def load_results():
data = json.loads(urlopen(URL).read())
return data
def nth_large(val, vals):
return sum([1 for v in vals if v > val]) + 1
def format_timestamp(timestamp):
date = timestamp[:-6]
time = timestamp[-6:]
date = date[:-4] + '.' + date[-4:-2] + '.' + date[-2:]
time = time[:-4] + ':' + time[-4:-2] + ':' + time[-2:]
return date + ' ' + time
def model_size_flag(sz, FIELDS):
if pd.isna(sz) and 'Unknown' in FIELDS:
return True
if pd.isna(sz):
return False
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, fields):
check_box = {}
check_box['essential'] = ['Method', 'Param (B)', 'Language Model', 'Vision Model', 'Eval Date']
# revise there to set default dataset
check_box['required'] = ['Avg Score', 'Avg Rank'] + DEFAULT_BENCH
check_box['avg'] = ['Avg Score', 'Avg Rank']
check_box['all'] = check_box['avg'] + fields
type_map = defaultdict(lambda: 'number')
type_map['Method'] = 'html'
type_map['Language Model'] = type_map['Vision Model'] = 'html'
type_map['OpenSource'] = 'str'
type_map['Eval Date'] = 'str'
check_box['type_map'] = type_map
df = generate_table(results, fields)
return df, check_box
def BUILD_L2_DF(results, dataset):
res = defaultdict(list)
sub = [v for v in results.values() if dataset in v]
assert len(sub)
fields = list(sub[0][dataset].keys())
if dataset == 'Creation_MMBench':
reward_field = [f"{x}:reward" for x in fields]
vfs_field = [f"{x}:vfs" for x in fields]
fields = reward_field + vfs_field
raw_fields = list(sub[0][dataset].keys())
print(f'fields: {fields}')
print(res)
non_overall_fields = [x for x in fields if 'Overall' not in x]
overall_fields = [x for x in fields if 'Overall' in x]
# if dataset == 'MME':
# non_overall_fields = [x for x in non_overall_fields if not listinstr(['Perception', 'Cognition'], x)]
# overall_fields = overall_fields + ['Perception', 'Cognition']
# if dataset == 'OCRBench':
# non_overall_fields = [x for x in non_overall_fields if not listinstr(['Final Score'], x)]
# overall_fields = ['Final Score']
for m in results:
item = results[m]
if dataset not in item:
continue
meta = item['META']
for k in META_FIELDS:
if k == 'Param (B)':
param = meta['Parameters']
res[k].append(float(param.replace('B', '')) if param != '' else None)
elif k == 'Method':
name, url = meta['Method']
res[k].append(f'<a href="{url}">{name}</a>')
elif k == 'Eval Date':
eval_date = meta['Time'].split('/')
assert len(eval_date) == 3
eval_date = [x if len(x) > 1 else '0' + x for x in eval_date]
eval_date = '/'.join(eval_date)
res[k].append(eval_date)
else:
res[k].append(meta[k])
fields = [x for x in fields]
if dataset == 'Creation_MMBench':
for d in non_overall_fields:
original_d, data_type = d.split(':')[0], d.split(':')[-1]
res[d].append(item[dataset][original_d][data_type])
for d in overall_fields:
original_d, data_type = d.split(':')[0], d.split(':')[-1]
res[d].append(item[dataset][original_d][data_type])
# res[d].append(f"{item[dataset][d]['reward']}/{item[dataset][d]['vfs']}")
elif dataset in ['MMAlignBench', 'WildVision']:
for d in non_overall_fields:
res[d].append(item[dataset][d]['reward'])
for d in overall_fields:
res[d].append(item[dataset][d]['reward'])
else:
for d in non_overall_fields:
res[d].append(item[dataset][d])
for d in overall_fields:
res[d].append(item[dataset][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]
# if dataset == 'OCRBench':
# df = df.sort_values('Final Score')
# elif dataset == 'COCO_VAL':
# df = df.sort_values('CIDEr')
# elif dataset == 'VCR':
# df = df.sort_values('Overall-Jaccard')
# else:
# df = df.sort_values('Overall')
df = df.iloc[::-1]
check_box = {}
check_box['essential'] = ['Method', 'Param (B)', 'Language Model', 'Vision Model', 'Eval Date']
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'] = 'html'
type_map['OpenSource'] = 'str'
type_map['Eval Date'] = 'str'
check_box['type_map'] = type_map
return df, check_box
def generate_table(results, fields):
# def get_mmbench_v11(item):
# assert 'MMBench_TEST_CN_V11' in item and 'MMBench_TEST_EN_V11' in item
# val = (item['MMBench_TEST_CN_V11']['Overall'] + item['MMBench_TEST_EN_V11']['Overall']) / 2
# val = float(f'{val:.1f}')
# return val
res = defaultdict(list)
## item is each model's meta and results
for i, m in enumerate(results):
item = results[m]
meta = item['META']
for k in META_FIELDS:
if k == 'Param (B)':
param = meta['Parameters']
res[k].append(float(param.replace('B', '')) if param != '' else None)
elif k == 'Method':
name, url = meta['Method']
res[k].append(f'<a href="{url}">{name}</a>')
res['name'].append(name)
elif k == 'Eval Date':
eval_date = meta['Time'].split('/')
assert len(eval_date) == 3
eval_date = [x if len(x) > 1 else '0' + x for x in eval_date]
eval_date = '/'.join(eval_date)
res[k].append(eval_date)
else:
res[k].append(meta[k])
scores, ranks = [], []
# fields is dataset
for d in fields:
key_name = 'Overall'
if d in item:
if d in ['Creation_MMBench','MMAlignBench', 'WildVision']:
if d == 'Creation_MMBench':
res[d].append(f"{item[d][key_name]['reward']}/{item[d][key_name]['vfs']}") # need improve?
else:
res[d].append(item[d][key_name]['reward'])
scores.append((item[d][key_name]['reward']+100)/2)
ranks.append(nth_large(item[d][key_name]['reward'], [x[d][key_name]['reward'] for x in results.values() if d in x]))
else:
res[d].append(item[d][key_name])
scores.append(item[d][key_name])
ranks.append(nth_large(item[d][key_name], [x[d][key_name] for x in results.values() if d in x]))
else:
res[d].append(None)
scores.append(None)
ranks.append(None)
res['Avg Score'].append(round(np.mean(scores), 1) if None not in scores else None)
res['Avg Rank'].append(round(np.mean(ranks), 2) if None not in ranks else None)
df = pd.DataFrame(res)
valid, missing = df[~pd.isna(df['Avg Score'])], df[pd.isna(df['Avg Score'])]
valid = valid.sort_values('Avg Score')
valid = valid.iloc[::-1]
if len(fields):
missing = missing.sort_values(fields[0])
missing = missing.iloc[::-1]
df = pd.concat([valid, missing])
return df
|