File size: 10,906 Bytes
198ccb0 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 |
"""Error analysis utilities for multi-label classification."""
import logging
from typing import List, Dict, Tuple, Optional
import torch
import pandas as pd
import numpy as np
from collections import defaultdict, Counter
from evaluation.metrics import per_class_metrics, confusion_matrix_per_class
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class ErrorAnalyzer:
"""
Analyze classification errors for multi-label classification.
Identifies common misclassification patterns, false positives/negatives,
and provides insights for model improvement.
"""
def __init__(self):
"""Initialize error analyzer."""
pass
def analyze_false_positives(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None
) -> Dict[str, List[int]]:
"""
Identify false positive predictions per class.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
Returns:
Dictionary mapping class name to list of sample indices with false positives
Example:
>>> analyzer = ErrorAnalyzer()
>>> target = torch.tensor([[0, 1], [1, 0]])
>>> pred = torch.tensor([[1, 1], [1, 0]])
>>> fps = analyzer.analyze_false_positives(target, pred)
>>> fps["class_0"]
[0]
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
false_positives = {name: [] for name in class_names}
for i in range(num_classes):
class_target = target[:, i]
class_pred = y_pred[:, i]
# False positives: predicted but not in target
fp_mask = (class_pred == 1) & (class_target == 0)
fp_indices = torch.where(fp_mask)[0].tolist()
false_positives[class_names[i]] = fp_indices
return false_positives
def analyze_false_negatives(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None
) -> Dict[str, List[int]]:
"""
Identify false negative predictions per class.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
Returns:
Dictionary mapping class name to list of sample indices with false negatives
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
false_negatives = {name: [] for name in class_names}
for i in range(num_classes):
class_target = target[:, i]
class_pred = y_pred[:, i]
# False negatives: in target but not predicted
fn_mask = (class_pred == 0) & (class_target == 1)
fn_indices = torch.where(fn_mask)[0].tolist()
false_negatives[class_names[i]] = fn_indices
return false_negatives
def find_common_misclassification_patterns(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None,
top_k: int = 10
) -> List[Tuple[Tuple[str, ...], Tuple[str, ...], int]]:
"""
Find common patterns of misclassification.
Identifies frequently co-occurring classes that are misclassified together.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
top_k: Number of top patterns to return
Returns:
List of tuples: (predicted_classes, actual_classes, count)
Sorted by frequency (descending)
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
patterns = Counter()
for sample_idx in range(target.shape[0]):
# Get predicted and actual classes
pred_classes = tuple(sorted([
class_names[i] for i in range(num_classes) if y_pred[sample_idx, i] == 1
]))
actual_classes = tuple(sorted([
class_names[i] for i in range(num_classes) if target[sample_idx, i] == 1
]))
# Only count if there's a mismatch
if pred_classes != actual_classes:
patterns[(pred_classes, actual_classes)] += 1
# Return top K patterns
return patterns.most_common(top_k)
def analyze_class_confusion(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None
) -> pd.DataFrame:
"""
Analyze confusion between classes.
Creates a confusion matrix showing which classes are frequently
confused with each other.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
Returns:
DataFrame with confusion analysis
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
# Count confusions: when class A is predicted but class B is actual
confusion_counts = defaultdict(int)
for sample_idx in range(target.shape[0]):
pred_indices = set(i for i in range(num_classes) if y_pred[sample_idx, i] == 1)
actual_indices = set(i for i in range(num_classes) if target[sample_idx, i] == 1)
# False positives: predicted but not actual
for pred_idx in pred_indices - actual_indices:
for actual_idx in actual_indices:
confusion_counts[(class_names[pred_idx], class_names[actual_idx])] += 1
# Create DataFrame
if confusion_counts:
data = [
{"predicted": pred, "actual": actual, "count": count}
for (pred, actual), count in confusion_counts.items()
]
df = pd.DataFrame(data)
df = df.sort_values("count", ascending=False)
else:
df = pd.DataFrame(columns=["predicted", "actual", "count"])
return df
def get_error_summary(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None
) -> Dict:
"""
Get comprehensive error summary.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
Returns:
Dictionary with error statistics
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
# Get per-class metrics
per_class = per_class_metrics(target, y_pred, class_names)
# Calculate totals
total_fp = sum(metrics["fp"] for metrics in per_class.values())
total_fn = sum(metrics["fn"] for metrics in per_class.values())
total_tp = sum(metrics["tp"] for metrics in per_class.values())
total_tn = sum(metrics["tn"] for metrics in per_class.values())
# Find classes with most errors
classes_by_fp = sorted(
per_class.items(),
key=lambda x: x[1]["fp"],
reverse=True
)[:10]
classes_by_fn = sorted(
per_class.items(),
key=lambda x: x[1]["fn"],
reverse=True
)[:10]
return {
"total_samples": target.shape[0],
"total_classes": num_classes,
"total_false_positives": total_fp,
"total_false_negatives": total_fn,
"total_true_positives": total_tp,
"total_true_negatives": total_tn,
"fp_rate": total_fp / (total_fp + total_tn + 1e-5),
"fn_rate": total_fn / (total_fn + total_tp + 1e-5),
"top_fp_classes": [
{"class": name, "count": metrics["fp"]}
for name, metrics in classes_by_fp
],
"top_fn_classes": [
{"class": name, "count": metrics["fn"]}
for name, metrics in classes_by_fn
],
"per_class_metrics": per_class
}
def visualize_errors(
self,
target: torch.Tensor,
y_pred: torch.Tensor,
class_names: Optional[List[str]] = None
) -> Dict[str, pd.DataFrame]:
"""
Create visualizations-ready DataFrames for error analysis.
Args:
target: Ground truth binary matrix [batch_size, num_classes]
y_pred: Predicted binary matrix [batch_size, num_classes]
class_names: Optional list of class names
Returns:
Dictionary with DataFrames for visualization
"""
num_classes = target.shape[1]
if class_names is None:
class_names = [f"class_{i}" for i in range(num_classes)]
# Per-class metrics DataFrame
per_class = per_class_metrics(target, y_pred, class_names)
metrics_df = pd.DataFrame(per_class).T
# Confusion analysis DataFrame
confusion_df = self.analyze_class_confusion(target, y_pred, class_names)
# Error counts per class
error_counts = []
for name, metrics in per_class.items():
error_counts.append({
"class": name,
"false_positives": metrics["fp"],
"false_negatives": metrics["fn"],
"true_positives": metrics["tp"],
"true_negatives": metrics["tn"]
})
error_df = pd.DataFrame(error_counts)
return {
"per_class_metrics": metrics_df,
"confusion_analysis": confusion_df,
"error_counts": error_df
}
|