tfmot.quantization.keras.graph_transformations.transforms.LayerPattern

Defines a tree sub-graph pattern of Keras layers to match in a model.

tfmot.quantization.keras.graph_transformations.transforms.LayerPattern(
    class_name, config=None, inputs=None
)

LayerPattern can be used to describe various common patterns in model graphs that we need to find.

Matches a Conv+BN+ReLU6 and DepthwiseConv+BN+ReLU6 pattern. pattern = LayerPattern('ReLU', {'max_value': 6.0}, [ LayerPattern('BatchNormalization', {}, [ LayerPattern('Conv2D|DepthwiseConv2D', {} []) ]) ])

Matches multiple Conv2Ds feeding into a Concat. pattern = LayerPattern('Concat', {}, [ LayerPattern('Conv2D', {}, []), LayerPattern('Conv2D', {}, []) ])

class_name Type of keras layer (such as Conv2D, Dense etc.)
config Map of arguments of the layer to match. For eg., for ReLU(6.0) it would be {'max_value': 6.0}.
inputs input layers to the layer.