tfx.components.trainer.rewriting.tflite_rewriter.TFLiteRewriter

Performs TFLite conversion.

Inherits From: BaseRewriter

tfx.components.trainer.rewriting.tflite_rewriter.TFLiteRewriter(
    name: Text,
    filename: Text = 'tflite',
    copy_assets: bool = True,
    copy_assets_extra: bool = True,
    quantization_optimizations: Optional[Sequence[tf.lite.Optimize]] = None,
    quantization_supported_types: Optional[Sequence[tf.DType]] = None,
    quantization_enable_full_integer: bool = False,
    signature_key: Text = None
)

Args
name	The name to use when identifying the rewriter.
filename	The name of the file to use for the tflite model.
copy_assets	Boolean whether to copy the assets directory to the rewritten model directory.
copy_assets_extra	Boolean whether to copy the assets.extra directory to the rewritten model directory.
quantization_optimizations	Options for optimizations in quantization. If None, no quantization will be applied(float32). Check https://www.tensorflow.org/lite/performance/post_training_quantization for details.
quantization_supported_types	Options for optimizations in quantization. Check https://www.tensorflow.org/lite/performance/post_training_quantization for details.
quantization_enable_full_integer	True to quantizae with FULL_INTEGER option.
signature_key	Key identifying SignatureDef containing TFLite inputs and outputs.

Attributes
name	The user-specified name of the rewriter.

Methods

perform_rewrite

View source

perform_rewrite(
    original_model: tfx.components.trainer.rewriting.rewriter.ModelDescription,
    rewritten_model: tfx.components.trainer.rewriting.rewriter.ModelDescription
)

Invoke all validations and perform the rewrite.

Args
original_model	A base_rewriter.ModelDescription object describing the original model.
rewritten_model	A base_rewriter.ModelDescription object describing the location and type of the rewritten model.

Raises
ValueError	if the model was not successfully rewritten.