def __quantize_model(converter: tensorflow.lite.TFLiteConverter,
model: Detector) -> Any:
print_debug('\nConverting model, this will take some time...')
# Get sample data to be used during converstion
image_files: List[str] = [
os.path.abspath(os.path.join(IMAGES_PATH, image))
for image in os.listdir(IMAGES_PATH)
]
annotation_files: List[str] = [
os.path.abspath(os.path.join(ANNOTATIONS_PATH, annotations))
for annotations in os.listdir(ANNOTATIONS_PATH)
]
try:
# Set optimizations to perform on model to be compatible with edge TPUs
converter.optimizations = [tensorflow.lite.Optimize.DEFAULT]
converter.representative_dataset = lambda: __input_data_generator(
model, image_files, annotation_files)
converter.target_spec.supported_ops = [
tensorflow.lite.OpsSet.TFLITE_BUILTINS_INT8
]
converter.inference_input_type = tensorflow.uint8
converter.inference_output_type = tensorflow.uint8
# Start conversion
return converter.convert()
except Exception as exception: # pylint: disable=broad-except
print_debug('Error: Could not convert model to TFLite')
print_debug(str(exception))
def _set_converter_options_for_calibration(
self, converter: tf.lite.TFLiteConverter) -> tf.lite.TFLiteConverter:
"""Verify converter options and set required experimental options."""
if not converter.optimizations:
converter.optimizations = [tf.lite.Optimize.DEFAULT]
if not converter.representative_dataset:
raise ValueError('converter object must set representative_dataset')
converter.experimental_mlir_quantizer = True
converter._experimental_calibrate_only = True # pylint: disable=protected-access
return converter
def _set_converter_options_for_float(
self,
converter: tf.lite.TFLiteConverter) -> tf.lite.TFLiteConverter:
"""Verify converter options and set required experimental options."""
if converter.optimizations:
converter.optimizations = []
return converter
def get_tflite_model(converter: tf.lite.TFLiteConverter, config: Config):
if config.optimization:
converter.optimizations.append(tf.lite.Optimize.DEFAULT)
converter.representative_dataset = config.representative_dataset
if config.supported_ops:
converter.target_spec.supported_ops.clear()
converter.target_spec.supported_ops.update(config.supported_ops)
if config.supported_types:
converter.target_spec.supported_types.extend(config.supported_types)
converter.inference_input_type = config.inference_input_type
converter.inference_output_type = config.inference_output_type
tflite_model = converter.convert()
return tflite_model