tfmot.quantization.keras.quantize_annotate_model | TensorFlow Model Optimization (original) (raw)

tfmot.quantization.keras.quantize_annotate_model

Stay organized with collections Save and categorize content based on your preferences.

Annotate a tf.keras model to be quantized.

tfmot.quantization.keras.quantize_annotate_model(
    to_annotate
)

Used in the notebooks

Used in the guide
Cluster preserving quantization aware training (CQAT) Keras example Sparsity and cluster preserving quantization aware training (PCQAT) Keras example Pruning preserving quantization aware training (PQAT) Keras example

This function does not actually quantize the model. It merely specifies that the model needs to be quantized. quantize_apply can then be used to quantize the model.

This function is intended to be used in conjunction with thequantize_annotate_layer API. Otherwise, it is simpler to usequantize_model.

Annotate a model while overriding the default behavior for a layer:

quantize_config = MyDenseQuantizeConfig()

model = quantize_annotate_model(
  keras.Sequential([
    layers.Dense(10, activation='relu', input_shape=(100,)),
    quantize_annotate_layer(
        layers.Dense(2, activation='sigmoid'),
        quantize_config=quantize_config)
  ]))

# The first Dense layer gets quantized with the default behavior,
# but the second layer uses `MyDenseQuantizeConfig` for quantization.
quantized_model = quantize_apply(model)

Note that this function removes the optimizer from the original model.

Args
to_annotate	tf.keras model which needs to be quantized.

Returns
New tf.keras model with each layer in the model wrapped withQuantizeAnnotate. The new model preserves weights from the original model.

Raises
ValueError	if the model cannot be annotated.

Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License, and code samples are licensed under the Apache 2.0 License. For details, see the Google Developers Site Policies. Java is a registered trademark of Oracle and/or its affiliates.

Last updated 2023-05-26 UTC.