tfmot.sparsity.keras.PruneForLatencyOnXNNPack

Specifies to prune only 1x1 Conv2D layers in the model.

Inherits From: PruningPolicy

Used in the notebooks

Used in the guide
Pruning for on-device inference w/ XNNPACK

PruneForLatencyOnXNNPack checks that the model contains a subgraph that can leverage XNNPACK's sparse inference and applies pruning wrapper only to Conv2D with kernel_size = (1, 1).

Reference
Fast Sparse ConvNets XNNPACK Sparse Inference

Methods

allow_pruning

View source

allow_pruning(
    layer
)

Allows to prune only 1x1 Conv2D layers.

ensure_model_supports_pruning

View source

ensure_model_supports_pruning(
    model
)

Ensures that the model contains only supported layers.