File size: 12,173 Bytes
091afb2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 | """
check_cer.py
============
Measures TRUE CER by actually running the model on images.
Usage:
python check_cer.py # live CER on val set
python check_cer.py --saved # old behavior (fast, unreliable)
python check_cer.py --images test_images/ # run on any image folder
"""
import os
import sys
import json
import random
import cv2
import numpy as np
import editdistance
from pathlib import Path
try:
import torch
except ImportError:
print("ERROR: torch not installed. Run: pip install torch")
exit(1)
USE_SAVED = '--saved' in sys.argv
IMAGE_DIR = None
for i, arg in enumerate(sys.argv[1:], 1):
if arg == '--images' and i < len(sys.argv) - 1:
IMAGE_DIR = sys.argv[i + 1]
elif arg.startswith('--images='):
IMAGE_DIR = arg.split('=', 1)[1]
CHECKPOINTS = [
'checkpoint_epoch_50.pth',
'checkpoint_epoch_60.pth',
'checkpoint_epoch_70.pth',
'checkpoint_epoch_80.pth',
'checkpoint_epoch_90.pth',
'checkpoint_epoch_100.pth',
]
CHECKPOINT_DIR = 'checkpoints'
VAL_DATA_DIR = 'data/val'
VAL_ANN_FILE = 'data/val_annotations.json'
class AdaptiveImageNormalizer:
def __init__(self, target_height=64, target_width=512):
self.H = target_height
self.W = target_width
def _crop_to_text(self, gray):
inv = cv2.bitwise_not(gray)
_, thresh = cv2.threshold(inv, 20, 255, cv2.THRESH_BINARY)
coords = np.column_stack(np.where(thresh > 0))
if len(coords) == 0:
return gray
y_min, x_min = coords.min(axis=0)
y_max, x_max = coords.max(axis=0)
pad = max(4, int((y_max - y_min) * 0.15))
y_min = max(0, y_min - pad)
x_min = max(0, x_min - pad)
y_max = min(gray.shape[0] - 1, y_max + pad)
x_max = min(gray.shape[1] - 1, x_max + pad)
return gray[y_min:y_max + 1, x_min:x_max + 1]
def _smart_resize_gray(self, gray):
h, w = gray.shape
if h == 0 or w == 0:
return np.ones((self.H, self.W), dtype=np.uint8) * 255
scale = self.H / h
new_w = int(w * scale)
new_h = self.H
if new_w > self.W:
scale = self.W / w
new_h = int(h * scale)
new_w = self.W
resized = cv2.resize(gray, (new_w, new_h), interpolation=cv2.INTER_LANCZOS4)
canvas = np.ones((self.H, self.W), dtype=np.uint8) * 255
y_off = (self.H - new_h) // 2
x_off = (self.W - new_w) // 2
canvas[y_off:y_off + new_h, x_off:x_off + new_w] = resized
return canvas
def _binarize(self, img):
_, otsu = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
white_ratio = np.mean(otsu == 255)
if white_ratio < 0.30 or white_ratio > 0.97:
return cv2.adaptiveThreshold(
img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, 11, 2)
return otsu
def normalize(self, img):
if len(img.shape) == 3:
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
else:
gray = img.copy()
gray = cv2.fastNlMeansDenoising(gray, None, 10, 7, 21)
gray = self._crop_to_text(gray)
gray = self._smart_resize_gray(gray)
return self._binarize(gray)
def to_tensor(self, img):
return torch.FloatTensor(
img.astype(np.float32) / 255.0
).unsqueeze(0).unsqueeze(0)
def greedy_decode(outputs, idx_to_char):
pred_indices = torch.argmax(outputs, dim=2).permute(1, 0)
results = []
for seq in pred_indices:
chars, prev = [], -1
for idx in seq:
idx = idx.item()
if idx != 0 and idx != prev and idx in idx_to_char:
chars.append(idx_to_char[idx])
prev = idx
results.append(''.join(chars))
return results
def measure_live_cer(model, idx_to_char, img_h, img_w,
ann_file, data_dir, device, max_samples=200):
if not os.path.exists(ann_file):
return None, 0, f"Annotation file not found: {ann_file}"
with open(ann_file, 'r', encoding='utf-8') as f:
annotations = json.load(f)
if len(annotations) > max_samples:
random.seed(42)
annotations = random.sample(annotations, max_samples)
normalizer = AdaptiveImageNormalizer(img_h, img_w)
model.eval()
total_char_dist = 0
total_chars = 0
total_word_dist = 0
total_words = 0
n_exact = 0
n_evaluated = 0
worst_errors = []
with torch.no_grad():
for ann in annotations:
img_path = os.path.join(data_dir, ann['image_path'])
gt = ann['text']
if not os.path.exists(img_path):
continue
try:
raw = cv2.imread(img_path)
if raw is None:
continue
norm = normalizer.normalize(raw)
tensor = normalizer.to_tensor(norm).to(device)
out = model(tensor)
pred = greedy_decode(out.cpu(), idx_to_char)[0]
cd = editdistance.eval(pred, gt)
wd = editdistance.eval(pred.split(), gt.split())
total_char_dist += cd
total_chars += len(gt)
total_word_dist += wd
total_words += len(gt.split())
if pred == gt:
n_exact += 1
if cd > 0:
worst_errors.append((gt, pred, cd))
n_evaluated += 1
except Exception:
continue
if n_evaluated == 0:
return None, 0, "No images could be evaluated"
cer = (total_char_dist / total_chars * 100) if total_chars > 0 else 0
wer = (total_word_dist / total_words * 100) if total_words > 0 else 0
acc = (n_exact / n_evaluated * 100)
return {
'cer': cer, 'wer': wer, 'exact_match': acc,
'n_evaluated': n_evaluated,
'errors': sorted(worst_errors, key=lambda x: x[2], reverse=True)[:5]
}, n_evaluated, None
def run_on_folder(model, idx_to_char, img_h, img_w, folder, device):
normalizer = AdaptiveImageNormalizer(img_h, img_w)
model.eval()
exts = {'.jpg', '.jpeg', '.png', '.bmp'}
paths = sorted(p for p in Path(folder).rglob('*') if p.suffix.lower() in exts)
results = []
with torch.no_grad():
for p in paths:
try:
raw = cv2.imread(str(p))
norm = normalizer.normalize(raw)
tensor = normalizer.to_tensor(norm).to(device)
pred = greedy_decode(model(tensor).cpu(), idx_to_char)[0]
results.append((p.name, pred))
except Exception as e:
results.append((p.name, f'ERROR: {e}'))
return results
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# MAIN
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if USE_SAVED:
print("=" * 65)
print(" SAVED CER (training-time value β may not reflect real accuracy)")
print(" Run without --saved for true live CER.")
print("=" * 65)
print("{:<8} {:<12} {:<12} {}".format("Epoch", "CER(%)", "WER(%)", "File"))
print("-" * 65)
best_cer, best_cp = float('inf'), None
for cp in CHECKPOINTS:
path = os.path.join(CHECKPOINT_DIR, cp)
if not os.path.exists(path):
continue
try:
c = torch.load(path, weights_only=False)
cer = c.get('val_cer', c.get('val_loss', 0))
epoch = c['epoch']
history = c.get('history', {})
wer_list = history.get('val_wer', [])
wer = wer_list[epoch - 1] if wer_list and epoch <= len(wer_list) else None
wer_s = f"{wer:.4f}%" if wer else 'N/A'
marker = ' <-- BEST' if cer < best_cer else ''
print("{:<8} {:<12} {:<12} {}{}".format(
epoch, f"{cer:.4f}%", wer_s, cp, marker))
if cer < best_cer:
best_cer, best_cp = cer, cp
except Exception as e:
print(f" Could not load {cp}: {e}")
print("=" * 65)
print(f"\nBEST: {best_cp} CER={best_cer:.4f}%")
else:
print("=" * 78)
print(" LIVE CER β model actually runs on images (true accuracy)")
print("=" * 78)
print("{:<8} {:<10} {:<10} {:<12} {:<8} {}".format(
"Epoch", "CER(%)", "WER(%)", "ExactMatch", "N", "File"))
print("-" * 78)
best_cer, best_cp, best_metrics = float('inf'), None, None
for cp in CHECKPOINTS:
cp_path = os.path.join(CHECKPOINT_DIR, cp)
if not os.path.exists(cp_path):
print(f" (skipping {cp} β not found)")
continue
try:
from crnn_model import get_crnn_model
c = torch.load(cp_path, map_location=device, weights_only=False)
epoch = c['epoch']
idx_to_char = c['idx_to_char']
config = c.get('config', {})
img_h = config.get('img_height', 64)
img_w = config.get('img_width', 512)
saved_cer = c.get('val_cer', c.get('val_loss', None))
model = get_crnn_model(
model_type=config.get('model_type', 'standard'),
img_height=img_h,
num_chars=c['model_state_dict']['fc.weight'].shape[0],
hidden_size=config.get('hidden_size', 128), # FIXED: was 256
num_lstm_layers=config.get('num_lstm_layers', 1) # FIXED: was 2
).to(device)
model.load_state_dict(c['model_state_dict'])
if IMAGE_DIR:
print(f"\nPredictions from {cp}:")
for fname, pred in run_on_folder(
model, idx_to_char, img_h, img_w, IMAGE_DIR, device):
print(f" {fname:<35} -> {pred}")
continue
metrics, n, err = measure_live_cer(
model, idx_to_char, img_h, img_w,
VAL_ANN_FILE, VAL_DATA_DIR, device)
if metrics is None:
print(f" Epoch {epoch} SKIP: {err}")
continue
cer = metrics['cer']
marker = ' <-- BEST' if cer < best_cer else ''
print("{:<8} {:<10} {:<10} {:<12} {:<8} {}{}".format(
epoch,
f"{cer:.2f}%",
f"{metrics['wer']:.2f}%",
f"{metrics['exact_match']:.1f}%",
n, cp, marker))
if saved_cer and abs(cer - saved_cer) > 2.0:
print(f" ^ MISMATCH: saved={saved_cer:.2f}% live={cer:.2f}%"
f" diff={abs(cer - saved_cer):.2f}%")
print(f" Cause: model trained on clean synthetic only.")
print(f" Fix: regenerate data with fix_data.py + retrain.")
if cer < best_cer:
best_cer, best_cp, best_metrics = cer, cp, metrics
except Exception as e:
print(f" Could not evaluate {cp}: {e}")
if not IMAGE_DIR:
print("=" * 78)
print(f"\nBEST CHECKPOINT : {best_cp}")
print(f"BEST LIVE CER : {best_cer:.4f}%")
if best_metrics and best_metrics['errors']:
print(f"\nWorst predictions (GT -> Predicted):")
for gt, pred, dist in best_metrics['errors']:
print(f" [{dist:2d}] '{gt}'")
print(f" '{pred}'")
print(f"\nTo use best model:")
print(f" import shutil")
print(f" shutil.copy('checkpoints/{best_cp}', 'checkpoints/best_model.pth')")
print("=" * 78) |