File size: 13,603 Bytes
7111e1a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 | """
IAM_train.py
============
Fine-tune the CRNN model using the IAM Handwriting Word Database.
Builds on top of EMNIST-trained model (best_model_emnist.pth).
FIXES vs old version:
- IMG_WIDTH 400 -> 512 (must match pipeline)
- Added log_softmax before CTCLoss (was missing β caused catastrophic forgetting)
- Phase 1: CNN FROZEN β only RNN+FC trained
- Phase 2: Full model at very low LR
- Loads from best_model_emnist.pth, falls back to best_model.pth
- Uses get_crnn_model() with correct architecture from checkpoint config
DATASET:
Download from: https://www.kaggle.com/datasets/nibinv23/iam-handwriting-word-database
Expected structure:
data/IAM/iam_words/
words/ <- word image folders (a01, a02, ...)
words.txt <- annotation file
USAGE:
python IAM_train.py --prepare # convert IAM -> annotation JSON
python IAM_train.py --train # fine-tune model
python IAM_train.py --prepare --train # do both
"""
import os
import sys
import json
import argparse
import random
from pathlib import Path
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader, ConcatDataset
sys.path.append('.')
from crnn_model import get_crnn_model
from dataset import CivilRegistryDataset, collate_fn
# βββββββββββββββββββββββββββββββββββββββββββββ
# CONFIG
# βββββββββββββββββββββββββββββββββββββββββββββ
IAM_ROOT = "data/IAM/iam_words"
IAM_WORDS_TXT = f"{IAM_ROOT}/words.txt"
IAM_WORDS_DIR = f"{IAM_ROOT}/words"
TRAIN_ANN = "data/iam_train_annotations.json"
IAM_VAL_ANN = "data/iam_val_annotations.json" # written by --prepare (IAM word images)
SYNTH_VAL_ANN = "data/val_annotations.json" # real civil registry val set β never overwritten
TRAIN_IMG_DIR = "data/train/iam"
VAL_IMG_DIR = "data/val/iam"
IMG_HEIGHT = 64
IMG_WIDTH = 512 # FIXED: was 400 β must match pipeline
BATCH_SIZE = 32
VAL_SPLIT = 0.1
MAX_SAMPLES = 50000
# Load from EMNIST checkpoint, fall back to synthetic if not found
CHECKPOINT_IN = "checkpoints/best_model_emnist.pth"
CHECKPOINT_IN2 = "checkpoints/best_model.pth" # fallback
CHECKPOINT_OUT = "checkpoints/best_model_iam.pth"
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# βββββββββββββββββββββββββββββββββββββββββββββ
# STEP 1 β PREPARE
# βββββββββββββββββββββββββββββββββββββββββββββ
def prepare_iam():
from PIL import Image
print("\n" + "=" * 50)
print("STEP 1 β Preparing IAM dataset")
print("=" * 50)
if not os.path.exists(IAM_WORDS_TXT):
print(f"ERROR: {IAM_WORDS_TXT} not found!")
print("Download from: https://www.kaggle.com/datasets/nibinv23/iam-handwriting-word-database")
print("Expected structure:")
print(" data/IAM/iam_words/words.txt")
print(" data/IAM/iam_words/words/")
sys.exit(1)
os.makedirs(TRAIN_IMG_DIR, exist_ok=True)
os.makedirs(VAL_IMG_DIR, exist_ok=True)
entries = []
print(f" Reading {IAM_WORDS_TXT} ...")
with open(IAM_WORDS_TXT, "r") as f:
for line in f:
line = line.strip()
if not line or line.startswith("#"):
continue
parts = line.split(" ")
if len(parts) < 9:
continue
word_id = parts[0]
seg_result = parts[1]
text = parts[-1]
if seg_result != "ok":
continue
if len(text) < 1 or len(text) > 32:
continue
parts_id = word_id.split("-")
img_path = os.path.join(
IAM_WORDS_DIR,
parts_id[0],
f"{parts_id[0]}-{parts_id[1]}",
f"{word_id}.png"
)
if not os.path.exists(img_path):
continue
entries.append((img_path, text))
print(f" Found {len(entries)} valid word entries")
if MAX_SAMPLES and len(entries) > MAX_SAMPLES:
random.shuffle(entries)
entries = entries[:MAX_SAMPLES]
print(f" Limiting to {MAX_SAMPLES} samples")
random.shuffle(entries)
split_idx = int(len(entries) * (1 - VAL_SPLIT))
train_entries = entries[:split_idx]
val_entries = entries[split_idx:]
print(f" Train: {len(train_entries)} | Val: {len(val_entries)}")
print(" Copying and resizing images...")
def process_entries(entry_list, out_dir, prefix):
annotations = []
for i, (src_path, text) in enumerate(entry_list):
try:
img = Image.open(src_path).convert("RGB")
img = img.resize((IMG_WIDTH, IMG_HEIGHT)) # FIXED: 512x64
fname = f"iam_{prefix}_{i:06d}.jpg"
out_path = os.path.join(out_dir, fname)
img.save(out_path, quality=90)
annotations.append({"image_path": f"iam/{fname}", "text": text})
except Exception:
continue
if i % 5000 == 0:
print(f" {i}/{len(entry_list)} processed...")
return annotations
train_ann = process_entries(train_entries, TRAIN_IMG_DIR, "train")
val_ann = process_entries(val_entries, VAL_IMG_DIR, "val")
with open(TRAIN_ANN, "w") as f:
json.dump(train_ann, f, indent=2)
with open(IAM_VAL_ANN, "w") as f:
json.dump(val_ann, f, indent=2)
print(f"\n Train annotations -> {TRAIN_ANN} ({len(train_ann)} entries)")
print(f" Val annotations -> {IAM_VAL_ANN} ({len(val_ann)} entries)")
print("\n Done! Now run: python IAM_train.py --train")
# βββββββββββββββββββββββββββββββββββββββββββββ
# STEP 2 β TRAIN
# βββββββββββββββββββββββββββββββββββββββββββββ
def train_iam():
print("\n" + "=" * 55)
print("STEP 2 β Fine-tuning CRNN with IAM dataset")
print("=" * 55)
print(f" Device : {DEVICE}")
for ann_file in [TRAIN_ANN, SYNTH_VAL_ANN]:
if not os.path.exists(ann_file):
print(f"ERROR: {ann_file} not found! Run --prepare first.")
sys.exit(1)
train_dataset = CivilRegistryDataset(
data_dir="data/train", annotations_file=TRAIN_ANN,
img_height=IMG_HEIGHT, img_width=IMG_WIDTH, augment=True
)
# FIXED: mix synthetic data in so the model never forgets Filipino multi-word sequences
synth_dataset = CivilRegistryDataset(
data_dir="data/train", annotations_file="data/train_annotations.json",
img_height=IMG_HEIGHT, img_width=IMG_WIDTH, augment=True
)
mixed_train = ConcatDataset([train_dataset, synth_dataset])
val_dataset = CivilRegistryDataset(
data_dir="data/val", annotations_file=SYNTH_VAL_ANN,
img_height=IMG_HEIGHT, img_width=IMG_WIDTH, augment=False
)
print(f" IAM train : {len(train_dataset)}")
print(f" Synthetic train: {len(synth_dataset)}")
print(f" Mixed train : {len(mixed_train)}")
print(f" Val : {len(val_dataset)}")
train_loader = DataLoader(mixed_train, batch_size=BATCH_SIZE,
shuffle=True, num_workers=0, collate_fn=collate_fn)
val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE,
shuffle=False, num_workers=0, collate_fn=collate_fn)
# ββ Load checkpoint (EMNIST preferred, synthetic fallback) ββ
ckpt_path = CHECKPOINT_IN if os.path.exists(CHECKPOINT_IN) else CHECKPOINT_IN2
if not os.path.exists(ckpt_path):
print(f"ERROR: No checkpoint found at {CHECKPOINT_IN} or {CHECKPOINT_IN2}")
print("Run: python train.py then python train_with_emnist.py")
sys.exit(1)
print(f" Loading: {ckpt_path}")
ckpt = torch.load(ckpt_path, map_location=DEVICE, weights_only=False)
config = ckpt.get('config', {})
model = get_crnn_model(
model_type = config.get('model_type', 'standard'),
img_height = config.get('img_height', 64),
num_chars = train_dataset.num_chars,
hidden_size = config.get('hidden_size', 128),
num_lstm_layers = config.get('num_lstm_layers', 1),
).to(DEVICE)
missing, _ = model.load_state_dict(ckpt['model_state_dict'], strict=False)
if missing:
print(f" Note: {len(missing)} layers re-initialized")
print(f" Loaded epoch {ckpt.get('epoch', 'N/A')} "
f"val_loss={ckpt.get('val_loss', ckpt.get('val_cer', 0)):.4f}")
criterion = torch.nn.CTCLoss(blank=0, reduction='mean', zero_infinity=True)
os.makedirs("checkpoints", exist_ok=True)
def run_epoch(loader, training, optimizer=None):
model.train() if training else model.eval()
total, n = 0, 0
ctx = torch.enable_grad() if training else torch.no_grad()
with ctx:
for images, targets, target_lengths, _ in loader:
images = images.to(DEVICE)
batch_size = images.size(0)
if training:
optimizer.zero_grad()
# CRITICAL: log_softmax before CTCLoss
outputs = F.log_softmax(model(images), dim=2)
seq_len = outputs.size(0)
input_lengths = torch.full((batch_size,), seq_len, dtype=torch.long)
loss = criterion(outputs, targets, input_lengths, target_lengths)
if not torch.isnan(loss) and not torch.isinf(loss):
if training:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
optimizer.step()
total += loss.item()
n += 1
return total / max(n, 1)
def run_phase(num, epochs, lr, freeze_cnn, patience):
print(f"\n{'='*55}")
print(f" PHASE {num} β "
f"{'CNN FROZEN (RNN+FC only)' if freeze_cnn else 'FULL MODEL (all layers)'}"
f" LR={lr}")
print(f"{'='*55}")
for name, param in model.named_parameters():
param.requires_grad = not (freeze_cnn and 'cnn' in name)
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f" Trainable params : {trainable:,}")
opt = optim.Adam(
filter(lambda p: p.requires_grad, model.parameters()), lr=lr)
sched = optim.lr_scheduler.ReduceLROnPlateau(opt, patience=3, factor=0.5)
best = float('inf')
counter = 0
for epoch in range(1, epochs + 1):
tr = run_epoch(train_loader, True, opt)
vl = run_epoch(val_loader, False, None)
sched.step(vl)
if vl < best:
best = vl
counter = 0
torch.save({
'model_state_dict': model.state_dict(),
'config': config,
'char_to_idx': train_dataset.char_to_idx,
'idx_to_char': train_dataset.idx_to_char,
'epoch': epoch,
'val_loss': vl, # FIXED: renamed from val_cer β this is val loss, not CER%
}, CHECKPOINT_OUT)
print(f" Epoch {epoch:02d}/{epochs} "
f"Train={tr:.4f} Val={vl:.4f} <- saved")
else:
counter += 1
print(f" Epoch {epoch:02d}/{epochs} "
f"Train={tr:.4f} Val={vl:.4f} "
f"(patience {counter}/{patience})")
if counter >= patience:
print(f" Early stopping at epoch {epoch}.")
break
return best
# Phase 1: Freeze CNN
p1 = run_phase(1, epochs=30, lr=1e-4, freeze_cnn=True, patience=7)
# Phase 2: Full model, very low LR
p2 = run_phase(2, epochs=20, lr=1e-6, freeze_cnn=False, patience=5)
print(f"\n{'='*55}")
print(f"IAM fine-tuning complete!")
print(f" Phase 1 best val loss : {p1:.4f}")
print(f" Phase 2 best val loss : {p2:.4f}")
print(f" Saved : {CHECKPOINT_OUT}")
print(f"\nNext step: collect physical certificate scans")
# βββββββββββββββββββββββββββββββββββββββββββββ
# MAIN
# βββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--prepare", action="store_true")
parser.add_argument("--train", action="store_true")
args = parser.parse_args()
if not args.prepare and not args.train:
print("Usage:")
print(" python IAM_train.py --prepare # prepare dataset")
print(" python IAM_train.py --train # train model")
print(" python IAM_train.py --prepare --train # do both")
sys.exit(0)
if args.prepare:
prepare_iam()
if args.train:
train_iam() |