tfmot.quantization.keras.default_8bit.default_8bit_transforms.SeparableConvQuantize

View source on GitHub

Break SeparableConv into a DepthwiseConv and Conv layer.

Inherits From: Transform

SeparableConv is a composition of a DepthwiseConv and a Conv layer. For the purpose of quantization, a FQ operation needs to be placed between the output of DepthwiseConv and the following Conv.

This is needed since there is a dynamic tensor in between the two layers, and it's range information needs to be captured by the FakeQuant op to ensure full int8 quantization of the layers is possible.

Splitting the layer into 2 ensures that each individual layer is handled correctly with respect to quantization.

Methods

custom_objects

View source

custom_objects()

Dictionary of custom objects introduced by the replacement function.

A Transform may introduce custom Classes and types unknown to Keras. This function should return a dictionary containing these objects in case such types are introduced. It allows model construction to serialize/deserialize these objects.

pattern

View source

pattern()

Return the LayerPattern to find in the model graph.

replacement

View source

replacement(
    match_layer
)

Generate a replacement sub-graph for the matched sub-graph.

The fundamental constraint of the replacement is that the replacement sub-graph should consume the same input tensors as the original sub-graph and also produce a final list of tensors which are same in number and shape as the original sub-graph. Not following this could crash model creation, or introduce bugs in the new model graph.

sub-graph, and output layers feeding from the tip of the tree as parameters. These would be needed for complex replace cases.