Datasets:
File size: 10,772 Bytes
4f3323a b6f22ac 4f3323a b6f22ac 4f3323a b6f22ac 4f3323a | 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 | #!/usr/bin/env python3
"""Create class-specific multiplicity figures from per-class-per-image AUPRO CSVs.
The script intersects M3DM/STFPM by (scene, split, image_id, class), so every
method comparison uses identical class-specific samples. Dataset-wide uses
sample counts only.
"""
import argparse
from pathlib import Path
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib.patches import Patch
CLASSES = ['residual_content', 'nuisance_sticker_inside', 'nuisance_gunk', 'nuisance_paper']
KS = list(range(1, 8))
def norm(df, fallback):
required = {'scene', 'image_id', 'class', 'multiplicity', 'aupro_30'}
missing = required - set(df.columns)
if missing:
raise SystemExit(f'missing columns: {sorted(missing)}')
d = df.copy()
if 'split' not in d:
d['split'] = 'anomaly'
if 'method' not in d:
d['method'] = fallback
d['scene'] = d['scene'].astype(str).str.zfill(6)
d['image_id'] = d['image_id'].astype(str).str.replace(r'\.0$', '', regex=True).str.zfill(5)
d['split'] = d['split'].astype(str).str.lower()
d['class'] = d['class'].astype(str)
d['multiplicity'] = pd.to_numeric(d['multiplicity']).astype(int)
d['aupro_30'] = pd.to_numeric(d['aupro_30'], errors='coerce')
return d
def aggregate(d):
return (d.groupby(['method', 'class', 'multiplicity'])['aupro_30']
.agg(Value='mean', n='count', std='std').reset_index())
def pooled_mean(agg):
x = agg.assign(total=agg.Value * agg.n)
return x.groupby('class').total.sum() / x.groupby('class').n.sum()
def class_means(ax, means, ylim, fmt='{:.2f}'):
for i, c in enumerate(CLASSES):
if c not in means:
continue
v = float(means[c])
ax.hlines(v, i - .4, i + .4, color='red', linestyle=':', linewidth=3,
label='Class Mean' if i == 0 else None)
ax.text(i + .3, v + .018 * (ylim[1] - ylim[0]), fmt.format(v), color='red', ha='left',
va='center', fontsize=9, fontweight='bold')
def finish(ax, ylim, ylabel, handles=None, labels=None, ncol=4):
ax.set_xlabel('')
ax.set_ylabel(ylabel)
ax.set_ylim(*ylim)
ax.tick_params(axis='x', rotation=15)
for t in ax.get_xticklabels():
t.set_ha('right')
if handles is None:
handles, labels = ax.get_legend_handles_labels()
ax.legend(handles, labels, loc='upper left', ncol=ncol, frameon=True, framealpha=.2,
facecolor='white', edgecolor='black')
def save(fig, stem):
fig.tight_layout()
fig.savefig(stem.with_suffix('.pdf'), bbox_inches='tight')
fig.savefig(stem.with_suffix('.png'), dpi=300, bbox_inches='tight')
plt.close(fig)
def grouped(agg, means, stem, ylim, method=None):
d = agg if method is None else agg[agg.method == method]
sns.set_theme(style='whitegrid', font_scale=1.1)
fig, ax = plt.subplots(figsize=(14, 5))
sns.barplot(data=d, x='class', y='Value', hue='multiplicity', order=CLASSES,
hue_order=KS, palette='viridis', ax=ax)
class_means(ax, means, ylim)
h, l = ax.get_legend_handles_labels()
l = [('1 type' if x == '1' else f'{x} types') if x.isdigit() else x for x in l]
finish(ax, ylim, 'Mean Pixel AUPRO Score', h, l, 4)
save(fig, stem)
def counts(common, stem):
unique = common[['scene', 'split', 'image_id', 'class', 'multiplicity']].drop_duplicates()
a = (unique.groupby(['class', 'multiplicity']).size().reset_index(name='Value'))
means = a.groupby('class').Value.mean()
top = max(a.Value.max(), means.max()) * 1.42
sns.set_theme(style='whitegrid', font_scale=1.1)
fig, ax = plt.subplots(figsize=(14, 5))
sns.barplot(data=a, x='class', y='Value', hue='multiplicity', order=CLASSES,
hue_order=KS, palette='viridis', ax=ax)
class_means(ax, means, (0, top), '{:.0f}')
h, l = ax.get_legend_handles_labels()
l = [('1 type' if x == '1' else f'{x} types') if x.isdigit() else x for x in l]
finish(ax, (0, top), 'Number of Images', h, l, 4)
save(fig, stem)
a.to_csv(stem.with_name(stem.name + '_values.csv'), index=False)
def compare_coloured(agg, means, stem, ylim, mode):
methods = list(agg.method.drop_duplicates())
d = agg.copy()
if mode == 'multiplicity':
d['group'] = d.multiplicity.astype(str) + ' ' + d.method
order = [f'{k} {m}' for k in KS for m in methods]
else:
d['group'] = d.method + ' k=' + d.multiplicity.astype(str)
order = [f'{m} k={k}' for m in methods for k in KS]
order = [x for x in order if x in set(d.group)]
pal = {}
for x in order:
k = int(x.split()[0]) if mode == 'multiplicity' else int(x.rsplit('=', 1)[1])
m = x.split(' ', 1)[1] if mode == 'multiplicity' else x.rsplit(' k=', 1)[0]
base = np.array(plt.cm.viridis((k - 1) / 6)[:3])
shade = 1 if m == methods[0] else .55
pal[x] = tuple(base * shade + (1 - shade) * np.ones(3))
sns.set_theme(style='whitegrid', font_scale=1.1)
fig, ax = plt.subplots(figsize=(14, 5))
sns.barplot(data=d, x='class', y='Value', hue='group', order=CLASSES,
hue_order=order, palette=pal, ax=ax)
class_means(ax, means, ylim)
finish(ax, ylim, 'Mean Pixel AUPRO Score', ncol=7)
save(fig, stem)
def geometry(groups, size, within, between, span=.8):
block = size + (size - 1) * within
total = groups * block + (groups - 1) * between
scale = span / total
start = -span / 2
out = []
for g in range(groups):
for b in range(size):
out.append(start + (g * (block + between) + b * (1 + within)) * scale + scale / 2)
return np.array(out), scale
def compare_spaced(agg, means, stem, ylim, mode):
methods = list(agg.method.drop_duplicates())
lookup = {(r['class'], r.multiplicity, r.method): r.Value for _, r in agg.iterrows()}
order = [(k, m) for k in KS for m in methods] if mode == 'pair' else [(k, m) for m in methods for k in KS]
off, w = geometry(7, 2, .18, .75) if mode == 'pair' else geometry(2, 7, .06, 1.6)
colors = {k: plt.cm.viridis((k - 1) / 6) for k in KS}
sns.set_theme(style='whitegrid', font_scale=1.1)
fig, ax = plt.subplots(figsize=(14, 5))
for ci, c in enumerate(CLASSES):
for oi, (k, m) in enumerate(order):
v = lookup.get((c, k, m))
if v is not None:
ax.bar([ci + off[oi]], [v], width=w, color=colors[k], edgecolor='black', linewidth=.4)
ax.set_xticks(range(4), CLASSES)
class_means(ax, means, ylim)
handles = [Patch(facecolor=colors[k], edgecolor='black', label='1 type' if k == 1 else f'{k} types') for k in KS]
handles.append(plt.Line2D([], [], color='red', linestyle=':', linewidth=3, label='Class Mean'))
finish(ax, ylim, 'Mean Pixel AUPRO Score', handles, [h.get_label() for h in handles], 4)
note = 'left to right within each pair: ' if mode == 'pair' else 'left to right, block by block: '
ax.text(.5, -.30, note + ' → '.join(methods), transform=ax.transAxes, ha='center', va='top', fontsize=9, color='.3')
save(fig, stem)
def find_file(name):
candidates = [
Path(f"checkpoint_data/{name}"),
Path(f"../checkpoint_data/{name}"),
Path(__file__).resolve().parent.parent / "checkpoint_data" / name,
Path(__file__).resolve().parent / name,
Path(name),
]
for c in candidates:
if c.is_file():
return c
return None
def default_out_dir():
if Path("checkpoint_data").is_dir():
return Path("checkpoint_data/figures_per_class_multiplicity")
if (Path(__file__).resolve().parent.parent / "checkpoint_data").is_dir():
return Path(__file__).resolve().parent.parent / "checkpoint_data/figures_per_class_multiplicity"
return Path("./figures_per_class_multiplicity")
def main():
p = argparse.ArgumentParser(description="Create class-specific multiplicity figures from per-class CSVs")
p.add_argument('--m3dm', default=None, help="Path to per_class_image_aupro.csv (auto-detected if omitted)")
p.add_argument('--stfpm', default=None, help="Path to per_class_image_aupro_stfpm.csv (auto-detected if omitted)")
p.add_argument('--out', default=None, help="Output directory (default: checkpoint_data/figures_per_class_multiplicity)")
a = p.parse_args()
m3dm_path = Path(a.m3dm) if a.m3dm else find_file("per_class_image_aupro.csv")
stfpm_path = Path(a.stfpm) if a.stfpm else find_file("per_class_image_aupro_stfpm.csv")
if not m3dm_path or not m3dm_path.is_file():
raise SystemExit(f"M3DM per-class CSV not found: {m3dm_path}")
if not stfpm_path or not stfpm_path.is_file():
raise SystemExit(f"STFPM per-class CSV not found: {stfpm_path}")
m = norm(pd.read_csv(m3dm_path), 'M3DM')
s = norm(pd.read_csv(stfpm_path), 'STFPM')
key = ['scene', 'split', 'image_id', 'class']
m = m[m['class'].isin(CLASSES)]
s = s[s['class'].isin(CLASSES)]
common = m[key].drop_duplicates().merge(s[key].drop_duplicates(), on=key)
m = m.merge(common, on=key)
s = s.merge(common, on=key)
chk = m[key + ['multiplicity']].merge(s[key + ['multiplicity']], on=key, suffixes=('_m', '_s'))
if not (chk.multiplicity_m == chk.multiplicity_s).all():
raise SystemExit('multiplicity mismatch between CSVs')
d = pd.concat([m, s], ignore_index=True)
agg = aggregate(d)
means = pooled_mean(agg)
out = Path(a.out) if a.out is not None else default_out_dir()
out.mkdir(parents=True, exist_ok=True)
top = max(agg.Value.max(), means.max()) * 1.42
ylim = (0, top)
counts(m, out / '01_multiplicity_dataset_wide')
grouped(agg, agg[agg.method == 'M3DM'].groupby('class')[['Value', 'n']].apply(lambda x: np.average(x.Value, weights=x.n)), out / '02_multiplicity_m3dm', ylim, 'M3DM')
grouped(agg, agg[agg.method == 'STFPM'].groupby('class')[['Value', 'n']].apply(lambda x: np.average(x.Value, weights=x.n)), out / '03_multiplicity_stfpm', ylim, 'STFPM')
compare_coloured(agg, means, out / '04_multiplicity_compare_by_multiplicity', ylim, 'multiplicity')
compare_coloured(agg, means, out / '05_multiplicity_compare_by_method', ylim, 'method')
compare_spaced(agg, means, out / '06_multiplicity_paired_spaced', ylim, 'pair')
compare_spaced(agg, means, out / '07_multiplicity_blocked_spaced', ylim, 'block')
agg.to_csv(out / 'per_class_multiplicity_aggregates.csv', index=False)
print(f'M3DM selected rows: {len(m)}')
print(f'STFPM selected rows: {len(s)}')
print(f'Common class-specific rows per method: {len(common)}')
print(f'Wrote 7 figures to {out}')
if __name__ == '__main__':
main()
|