Spaces:
Sleeping
Sleeping
File size: 6,790 Bytes
8d5a97a 75b2724 8d5a97a 77c903f 8d5a97a 75b2724 8d5a97a 75b2724 8d5a97a 7e85849 8d5a97a 7e85849 8d5a97a d4dddf1 8d5a97a 5aea366 8d5a97a 75b2724 8d5a97a 7e85849 8d5a97a 75b2724 8d5a97a 75b2724 8d5a97a 7e85849 8d5a97a 7e85849 8d5a97a 6821614 77c903f 8d5a97a 77c903f 8d5a97a 77c903f 8d5a97a 7ee6ea8 8d5a97a 7e85849 8d5a97a fdccb4e 8d5a97a 75b2724 8d5a97a 77c903f 75b2724 77c903f 8d5a97a 75b2724 8d5a97a fdccb4e 8d5a97a 6eea004 8d5a97a d4dddf1 8d5a97a 6eea004 8d5a97a 75b2724 8d5a97a fdccb4e 8d5a97a d4dddf1 fdccb4e 8d5a97a 1900305 fdccb4e 8d5a97a 6eea004 8d5a97a d4dddf1 8d5a97a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 |
import math
import operator as op
import itertools as it
import functools as ft
import collections as cl
from pathlib import Path
from dataclasses import fields, asdict
import pandas as pd
import gradio as gr
import seaborn as sns
import matplotlib.pyplot as plt
from datasets import load_dataset
from scipy.special import expit
from hdinterval import HDI, HDInterval
TabGroup = cl.namedtuple('TabGroup', 'name, docs, dataset')
#
#
#
def load(repo):
parameter = 'parameter'
model = 'model'
items = [
'chain',
'sample',
parameter,
model,
'value',
]
dataset = load_dataset(str(repo))
return (dataset
.get('train')
.to_pandas()
.rename(columns={'element': model})
.filter(items=items)
.query(f'{parameter} == "alpha"')
.drop(columns=parameter))
def summarize(df, ci=0.95):
def _aggregate(i, g):
values = g['value']
hdi = HDInterval(values)
interval = hdi(ci)
agg = {
'model': i,
'ability': values.median(),
'uncertainty': interval.width(),
}
agg.update(asdict(interval))
return agg
groups = df.groupby('model', sort=False)
records = it.starmap(_aggregate, groups)
return pd.DataFrame.from_records(records)
def rank(df, ascending, name='rank'):
df = (df
.sort_values(by=['ability', 'uncertainty'],
ascending=[ascending, not ascending])
.drop(columns='uncertainty')
.reset_index(drop=True))
df.index += 1
return df.reset_index(names=name)
def compare(df, model_1, model_2):
mcol = 'model'
models = [
model_1,
model_2,
]
view = (df
.query(f'{mcol} in @models')
.pivot(index=['chain', 'sample'],
columns=mcol,
values='value'))
return expit(view[model_1] - view[model_2])
#
#
#
class DataPlotter:
def __init__(self, df):
self.df = df
def plot(self):
fig = plt.figure(dpi=200)
ax = fig.gca()
self.draw(ax)
ax.grid(visible=True,
axis='both',
alpha=0.25,
linestyle='dotted')
fig.tight_layout()
return fig
def draw(self, ax):
raise NotImplementedError()
class RankPlotter(DataPlotter):
_y = 'y'
@ft.cached_property
def y(self):
return self.df[self._y]
def __init__(self, df, top=10):
view = rank(summarize(df), True, self._y)
view = (view
.tail(top)
.sort_values(by=self._y, ascending=False))
super().__init__(view)
def draw(self, ax):
self.df.plot.scatter('ability', self._y, ax=ax)
ax.hlines(self.y,
xmin=self.df['lower'],
xmax=self.df['upper'],
alpha=0.5)
ax.set_xlabel(ax.get_xlabel().title())
ax.set_ylabel('')
ax.set_yticks(self.y, self.df['model'])
class ComparisonPlotter(DataPlotter):
def __init__(self, df, model_1, model_2, ci):
super().__init__(compare(df, model_1, model_2))
self.hdi = HDInterval(self.df)
self.ci = ci
def draw(self, ax):
interval = self.hdi(self.ci)
sns.ecdfplot(self.df, ax=ax)
(_, color, *_) = sns.color_palette()
ax.axvline(x=self.df.median(),
color=color,
linestyle='dashed')
ax.axvspan(xmin=interval.lower,
xmax=interval.upper,
alpha=0.15,
color=color)
ax.set_xlabel('Pr(M$_{1}$ \u003E M$_{2}$)')
try:
ci_mid = self.hdi.at(0.5)
ax.text(x=0.01,
y=0.99,
s=f'0.5-min HDI: {ci_mid:.0%}',
horizontalalignment='left',
verticalalignment='top',
transform=ax.transAxes)
except ArithmeticError:
pass
def cplot(df):
def _plot(model_1, model_2, ci):
cp = ComparisonPlotter(df, model_1, model_2, ci)
return cp.plot()
return _plot
#
#
#
class DocumentationReader:
_suffix = '.md'
def __init__(self, root):
self.root = root
def __getitem__(self, item):
return (self
.root
.joinpath(item)
.with_suffix(self._suffix)
.read_text())
#
#
#
def layout(tab):
df = load(Path('jerome-white', tab.dataset))
docs = DocumentationReader(Path('docs', t.docs))
with gr.Row():
with gr.Column():
gr.Markdown(docs['readme'])
with gr.Column():
plotter = RankPlotter(df)
gr.Plot(plotter.plot())
with gr.Row():
view = rank(summarize(df), False)
columns = { x.name: f'HDI {x.name}' for x in fields(HDI) }
for i in view.columns:
columns.setdefault(i, i.title())
view = (view
.rename(columns=columns)
.style.format(precision=4))
gr.Dataframe(view)
with gr.Row():
with gr.Column(scale=3):
display = gr.Plot()
with gr.Row():
with gr.Column():
gr.Markdown('''
Probability that Model 1 is preferred to Model 2. The
solid blue curve is a CDF of that distribution;
formally the inverse logit of the difference in model
abilities. The dashed orange vertical line is the
median, while the band surrounding it is the [highest
density
interval](https://cran.r-project.org/package=HDInterval)
of your choice (default 95%).
''')
with gr.Column():
models = df['model'].unique()
choices = sorted(models, key=lambda x: x.lower())
partial = ft.partial(gr.Dropdown, choices=choices)
drops = (partial(label=f'Model {x}') for x in range(1, 3))
ci = gr.Number(value=0.95, minimum=0, maximum=1, step=1e-2)
inputs = [
*drops,
ci,
]
button = gr.Button(value='Compare!')
button.click(cplot(df), inputs=inputs, outputs=[display])
with gr.Accordion('Disclaimer', open=False):
gr.Markdown(docs['disclaimer'])
#
#
#
with gr.Blocks() as demo:
tabs = it.starmap(TabGroup, (
('Alpaca', 'alpaca', 'alpaca-bt-stan'),
('Chatbot Arena', 'arena', 'arena-bt-stan'),
))
for t in tabs:
with gr.Tab(t.name):
layout(t)
demo.launch()
|