quantize Namespace Reference

Functions
None	_tflite_ptq_quantizer (tf.keras.Model model=None, tf.data.Dataset quantization_ds=None, bool fake=False, str output_dir=None, Optional[str] export_dir=None, tuple input_shape=None, str quantization_granularity=None, str quantization_input_type=None, str quantization_output_type=None, str quantization_split=None, str quantization_path=None)
str	quantize (DictConfig cfg=None, Optional[tf.data.Dataset] quantization_ds=None, Optional[bool] fake=False, Optional[str] float_model_path=None)

Function Documentation

_tflite_ptq_quantizer()

None quantize._tflite_ptq_quantizer ( tf.keras.Model   model = None, tf.data.Dataset   quantization_ds = None, bool   fake = False, str   output_dir = None, Optional[str]   export_dir = None, tuple   input_shape = None, str   quantization_granularity = None, str   quantization_input_type = None, str   quantization_output_type = None, str   quantization_split = None, str   quantization_path = None  )

private

Perform post-training quantization on a TensorFlow Lite model.

Args:
    model (tf.keras.Model): The TensorFlow model to be quantized.
    quantization_ds (tf.data.Dataset): The quantization dataset if it's provided by the user else the training
    dataset. Defaults to None
    fake (bool): Whether to use fake data for representative dataset generation.
    output_dir (str): Path to the output directory. Defaults to None.
    export_dir (str): Name of the export directory. Defaults to None.
    input_shape (tuple: The input shape of the model. Defaults to None.
    quantization_granularity (str): 'per_tensor' or 'per_channel'. Defaults to None.
    quantization_input_type (str): The quantization type for the input. Defaults to None.
    quantization_output_type (str): The quantization type for the output. Defaults to None.
    quantization_path (str): the quantization dataset path if it's provided by the user.  Defaults to None.
    quantization_split (str): The Fraction of the data to use for the quantization

Returns:
    None

Definition at line 44 of file quantize.py.

Referenced by quantize().

Here is the caller graph for this function:

quantize()

str quantize.quantize	(	DictConfig	cfg = None,
		Optional[tf.data.Dataset]	quantization_ds = None,
		Optional[bool]	fake = False,
		Optional[str]	float_model_path = None
	)

Quantize the TensorFlow model with training data.

Args:
    cfg (DictConfig): The configuration dictionary. Defaults to None.
    quantization_ds (tf.data.Dataset): The quantization dataset if it's provided by the user else the training
    dataset. Defaults to None.
    fake (bool, optional): Whether to use fake data for representative dataset generation. Defaults to False.
    float_model_path (str, optional): Model path to quantize

Returns:
    quantized model path (str)

Definition at line 135 of file quantize.py.

References _tflite_ptq_quantizer(), and preprocess.apply_rescaling().

Here is the call graph for this function: