squirelmail/model-BotDetect-CAPTCHA-Generator

Model AI For Solve BotDetect-CAPTCHA-Generator Gov ID Captcha

🧠 CRNN+CTC Checkpoints

This directory contains Keras 3 save_weights-style checkpoints produced during training of a CRNN + CTC model for 5-char uppercase/digit CAPTCHA (image size H=50, W=250, grayscale).

---

📁 Contents

• captcha_best.weights.h5 — best validation loss (auto-updated during training).
• captcha_epNNN.weights.h5 — per-epoch snapshots (e.g., captcha_ep001.weights.h5 … captcha_ep022.weights.h5).

All files are weights only; they must be loaded into the same model architecture used in training (the tester builds that architecture for you).

---

✅ Model Result captcha_ep022.weights.h5 => 90.91% Accuracy

(venv) root@prod-exploit-sa-all-01:/home/infra# date && python3 cek_model_v6.py --weights captcha_ep022.weights.h5 --data-root ./dataset_1000_rand --sample 24000 && date
Thu Oct 30 01:12:49 WITA 2025
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1761761571.108235 2264160 cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
I0000 00:00:1761761571.304280 2264160 cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1761761575.452128 2264160 cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
Found weights: captcha_ep022.weights.h5 | size: 27757.0 KB | mtime: Thu Oct 30 01:02:51 2025
E0000 00:00:1761761576.513960 2264160 cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303)
TF GPUs: []
OK: weights loaded.
Base output shape: (None, 31, 37)
Testing on 24000 samples from ./dataset_1000_rand ...
W0000 00:00:1761761611.159498 2264160 cpu_allocator_impl.cc:84] Allocation of 1200000000 exceeds 10% of free system memory.
00  GT: 976VF   |   Pred: 976VF
01  GT: 7W20H   |   Pred: 7W20H
02  GT: UUU24   |   Pred: UUU24
03  GT: 1EMVZ   |   Pred: 1EMVZ
04  GT: WY4RD   |   Pred: WY4RD
05  GT: 0GNKE   |   Pred: 0GNKE
06  GT: 7Y5TY   |   Pred: 7Y5TY
07  GT: OC8C1   |   Pred: OC8C1
08  GT: 5ZIDQ   |   Pred: 5ZIDQ
09  GT: LP8IP   |   Pred: LP8IP
10  GT: AKQ7G   |   Pred: AKQ7G
11  GT: X23QD   |   Pred: X23QD

Exact match: 90.91%  |  Mean CER: 0.0194

Total images tested: 24000

Thu Oct 30 01:18:07 WITA 2025

---

📦 Requirements

Install from the pinned list in the repo root:

# (recommended) fresh virtualenv
python3 -m venv venv
source venv/bin/activate

# install exact deps
pip install -r captcha_requirements.txt

Important: Keras/TensorFlow versions should match what was used during training. If you trained with TF/Keras nightly or dev builds, test in the same environment to avoid weight-loading shape/key mismatches.

---

🧪 How to Test

The tester script re-creates the training graph (CRNN+CTC), loads the selected checkpoint, and runs inference with the base (CTC-free) submodel.

1) Single image

python3 check_model.py \
  --weights /workspace/captcha_final.weights.h5 \
  --image   /workspace/dataset_500/style7/K9NO2.png

Optional ground truth override:

python3 check_model.py \
  --weights /workspace/captcha_final.weights.h5 \
  --image   /workspace/dataset_500/style7/K9NO2.png \
  --gt      K9NO2

2) Batch from a dataset

python3 check_model.py \
  --weights /home/infra/models/captcha_ep002.weights.h5 \
  --data-root /datasets/dataset_500 \
  --samples 64

Expected directory layout for --data-root:

/datasets/dataset_500/
├── style0/
│   ├── A1B2C.png
│   └── ...
├── style1/
│   └── ...
└── ...
    └── style59/

Image format: grayscale PNG, resized to 50x250 in the script.
Labels: derived from filename (regex ^[A-Z0-9]{5}$).

---

🧩 Model Details (for reference)

• Backbone: 3× (Conv2D + BN + MaxPool), then reshape to time-steps.
• RNN head: 2× BiLSTM(128), return_sequences=True.
• Classifier: Dense(num_classes = 36 + 1) with softmax; +1 is the CTC blank.
• Time steps: width is downsampled by 8 ⇒ 250/8 = 31 time steps.

The tester script internally builds both: model_with_ctc (training graph) and base_model (inference). It loads weights into the training graph and then uses base_model for predictions.

---

🎛️ CLI Options

--weights <path>   : required, *.weights.h5 (same architecture)
--image <path>     : test a single image
--gt <text>        : ground truth for --image (default: file name)
--data-root <dir>  : style0..style59 folders for batch testing
--samples N        : max number of images for batch test (default 64)
--height H         : input height (default 50)
--width W          : input width  (default 250)
--ext png|jpg      : image extension for batch (default png)
--show K           : print K sample predictions (default 12)

---

📊 Output

• Per-sample preview lines: GT: ABC12 | Pred: ABC12
• Aggregate metrics:
  • Exact match (% of predictions exactly equal to GT)
  • Mean CER (character error rate)

---

🧯 Troubleshooting

• “A total of 1 objects could not be loaded… ”
  Mismatch between Keras/TF versions or model definition. Use the same environment and architecture as training.

• GPU not used
  Ensure a CUDA-enabled TF build and matching drivers. For server-side issues, test with:

  import tensorflow as tf
  print(tf.config.list_physical_devices('GPU'))
  

• NaN loss during training
  Check: label regex filtering, correct input_length=31, use int32 for CTC inputs, disable LSTM dropouts when using cuDNN (set to 0.0).

---

🔐 Notes

• CTC blank ID = 36 (since charset is 36 chars: 0-9 + A-Z).

• All checkpoints here are weights only; to export a full model, save the base model as .keras after loading weights in the same environment:

  model_with_ctc, base_model = build_models(...)
  model_with_ctc.load_weights("captcha_epXXX.weights.h5")
  base_model.save("captcha_epXXX_base.keras")