Model Conversion for Edge (TensorFlow Lite Micro, TFLite, Edge TPU)
Three different target platforms — three different conversion pipelines. TFLite Micro for MCU, TFLite for mobile/SBC, Edge TPU (Google Coral) for hardware-accelerated inference.
TFLite (mobile / Raspberry Pi / x86 edge)
Standard conversion:
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_model = converter.convert()
Supports: INT8, FP16, dynamic range quantization. GPU delegate, NNAPI, Hexagon DSP.
TFLite Micro (MCU, <1 MB)
Subset of TFLite operations, portable C++:
xxd -i model.tflite > model_data.cc # convert to C array
Supported on: STM32, Arduino, ESP32, nRF52840. Operation set limited — compatibility checker mandatory.
Edge TPU (Google Coral)
Edge TPU requires INT8 quantization. Only TPU-supported operations execute in hardware (rest — CPU fallback):
edgetpu_compiler model_quant.tflite # Google Coral compiler
Performance: 4 TOPS (Coral USB), 4 TOPS (Coral PCIe M.2). Excellent for image classification and object detection.
Limitation: models >8 MB don't fit entirely on Edge TPU — partial fallback reduces acceleration. Design for <8 MB for maximum acceleration.