Karthikraj Sivakumar
commited on
Commit
·
7e07095
1
Parent(s):
c9c30b5
fix biased uncertainty problem
Browse files
app.py
CHANGED
|
@@ -379,16 +379,35 @@ def predict_captcha(image):
|
|
| 379 |
prediction, confidence = ctc_decode_with_confidence(log_probs, idx_to_char)
|
| 380 |
confidence_pct = confidence * 100
|
| 381 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 382 |
# Format output
|
| 383 |
output = f"**Primary Prediction:** {prediction}\n\n"
|
| 384 |
|
| 385 |
-
# Add status
|
| 386 |
-
if
|
| 387 |
-
|
| 388 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 389 |
|
| 390 |
-
|
| 391 |
-
top_predictions = ctc_decode_top_k(log_probs, idx_to_char, k=3)
|
| 392 |
|
| 393 |
output += f"{status} — {confidence_pct:.1f}%\n"
|
| 394 |
output += f"{note}\n\n"
|
|
@@ -396,7 +415,8 @@ def predict_captcha(image):
|
|
| 396 |
|
| 397 |
for i, (text, conf) in enumerate(top_predictions, 1):
|
| 398 |
conf_pct = conf * 100
|
| 399 |
-
|
|
|
|
| 400 |
|
| 401 |
output += "\n💡 *Tip: Check which makes sense in context*"
|
| 402 |
|
|
|
|
| 379 |
prediction, confidence = ctc_decode_with_confidence(log_probs, idx_to_char)
|
| 380 |
confidence_pct = confidence * 100
|
| 381 |
|
| 382 |
+
# Get top-k predictions to check uncertainty
|
| 383 |
+
top_predictions = ctc_decode_top_k(log_probs, idx_to_char, k=3)
|
| 384 |
+
|
| 385 |
+
# Check if alternatives are close (uncertainty margin)
|
| 386 |
+
show_alternatives = False
|
| 387 |
+
if len(top_predictions) >= 2:
|
| 388 |
+
top1_conf = top_predictions[0][1]
|
| 389 |
+
top2_conf = top_predictions[1][1]
|
| 390 |
+
margin = top1_conf - top2_conf
|
| 391 |
+
|
| 392 |
+
# Show alternatives if:
|
| 393 |
+
# 1. Low confidence (< 70%), OR
|
| 394 |
+
# 2. Top 2 predictions are very close (margin < 0.1)
|
| 395 |
+
if confidence < 0.70 or margin < 0.1:
|
| 396 |
+
show_alternatives = True
|
| 397 |
+
|
| 398 |
# Format output
|
| 399 |
output = f"**Primary Prediction:** {prediction}\n\n"
|
| 400 |
|
| 401 |
+
# Add status and alternatives based on confidence and margin
|
| 402 |
+
if show_alternatives:
|
| 403 |
+
if confidence < 0.6:
|
| 404 |
+
status = "⚠️ Low Confidence"
|
| 405 |
+
elif confidence < 0.70:
|
| 406 |
+
status = "⚡ Medium Confidence"
|
| 407 |
+
else:
|
| 408 |
+
status = "⚠️ Uncertain" # High confidence but close alternatives
|
| 409 |
|
| 410 |
+
note = "Visual ambiguity detected (e.g., 0/o, i/1/l confusion)"
|
|
|
|
| 411 |
|
| 412 |
output += f"{status} — {confidence_pct:.1f}%\n"
|
| 413 |
output += f"{note}\n\n"
|
|
|
|
| 415 |
|
| 416 |
for i, (text, conf) in enumerate(top_predictions, 1):
|
| 417 |
conf_pct = conf * 100
|
| 418 |
+
marker = "→" if i == 1 else " "
|
| 419 |
+
output += f"{marker} {i}. `{text}` — {conf_pct:.1f}%\n"
|
| 420 |
|
| 421 |
output += "\n💡 *Tip: Check which makes sense in context*"
|
| 422 |
|