| --- |
| license: apache-2.0 |
| language: |
| - en |
| --- |
| # Model Quantization Notebook |
|
|
| This notebook converts a pre-trained Keras violence detection model into TensorFlow Lite (TFLite) format using three different quantization strategies, making it suitable for deployment on edge/mobile devices. |
|
|
| --- |
|
|
| ## Overview |
|
|
| | Property | Details | |
| |---|---| |
| | **Framework** | TensorFlow / TFLite | |
| | **Base Model** | `modelv2.keras` — a Keras video violence detection model | |
| | **Input Shape** | `(1, 16, 224, 224, 3)` — batch × frames × height × width × channels | |
| | **Architecture** | CNN + LSTM (contains dynamic LSTM loops) | |
| | **Platform** | Kaggle (GPU hidden to avoid CuDNN conflicts) | |
|
|
| --- |
|
|
| ## Quantization Methods |
|
|
| ### A — Dynamic Range Quantization |
| - **Output file:** `model_dynamic_quant.tflite` |
| - Quantizes weights from float32 to int8 at **conversion time**. |
| - Activations are quantized **dynamically** at inference time. |
| - Fastest to convert; no calibration data required. |
| - Good balance between size reduction and accuracy. |
|
|
| ### B — Float16 Quantization |
| - **Output file:** `model_fp16_quant.tflite` |
| - Reduces weight precision from float32 to **float16**. |
| - Ideal for GPU-accelerated edge devices that support fp16 natively. |
| - Smaller model size with minimal accuracy loss. |
|
|
| ### C — Full Integer (INT8) Quantization |
| - **Output file:** `model_full_int8.tflite` |
| - Quantizes **both weights and activations** to int8. |
| - Requires a **representative dataset** for calibration (currently uses random dummy data — replace with real video samples for best results). |
| - Input and output tensors are also forced to int8. |
| - Smallest model size; best suited for CPU-only or microcontroller deployment. |
|
|
| --- |
|
|
| ## Requirements |
|
|
| ``` |
| tensorflow |
| numpy |
| ``` |
|
|
| --- |
|
|
| ## Usage |
|
|
| ### 1. Load the Base Model |
| ```python |
| import tensorflow as tf |
| |
| tf.config.set_visible_devices([], 'GPU') # Hide GPU to avoid CuDNN issues |
| model = tf.keras.models.load_model('path/to/modelv2.keras') |
| ``` |
|
|
| ### 2. Run Quantization |
| Open and run the notebook cells in order: |
| 1. **Cell 1–2** — Load the model |
| 2. **Cell 3–4** — Dynamic range quantization → `model_dynamic_quant.tflite` |
| 3. **Cell 5–6** — Float16 quantization → `model_fp16_quant.tflite` |
| 4. **Cell 7–8** — Full INT8 quantization → `model_full_int8.tflite` |
|
|
| --- |
|
|
| ## Important Notes |
|
|
| - **Representative dataset:** The INT8 quantization cell uses random dummy data for calibration. For production use, replace `dummy_data` in `representative_data_gen()` with real video frames from your training set to get accurate quantization ranges. |
|
|
| - **LSTM compatibility flags:** The model contains dynamic LSTM loops. The following flags are set in all conversion paths to prevent conversion failures: |
| ```python |
| converter.target_spec.supported_ops = [ |
| tf.lite.OpsSet.TFLITE_BUILTINS, |
| tf.lite.OpsSet.SELECT_TF_OPS |
| ] |
| converter._experimental_lower_tensor_list_ops = False |
| ``` |
|
|
| - **Static input shape:** The INT8 path uses `tf.function` with a `tf.TensorSpec` to lock the input shape to `(1, 16, 224, 224, 3)` before conversion — this is required for correct INT8 LSTM quantization. |
|
|
| --- |
|
|
| ## Output Files |
|
|
| | File | Method | Precision | |
| |---|---|---| |
| | `model_dynamic_quant.tflite` | Dynamic Range | Weights: INT8, Activations: float32 | |
| | `model_fp16_quant.tflite` | Float16 | Weights & Activations: float16 | |
| | `model_full_int8.tflite` | Full Integer | Weights & Activations: INT8 | |
