XNNPACK provide sparse inference using sparse kernel on multiple platform. but sparse inference of XNNPACK is limited.
your TFLite model follow sub-graph rule to use Sparse inference
specially, MobileNet v1 provided by keras (i.e. tensorflow.keras.application.MobileNet) doesn't follow sub-graph rule to use sparse inference.
if you use Sparse MobileNet v1, Use FC layer and no dimemsion keeping avgpool, last convolution is replaced by FC layer
Model sparsity (#nnz / sum(prunable layer weight) need to go over 66%
if you use sparse model at any model sparsity, you must change XNNPACK sub-graph code, and tensorflow backend compiler .
XNNPACK provide sparse inference using sparse kernel on multiple platform. but sparse inference of XNNPACK is limited.
your TFLite model follow sub-graph rule to use Sparse inference specially, MobileNet v1 provided by keras (i.e. tensorflow.keras.application.MobileNet) doesn't follow sub-graph rule to use sparse inference. if you use Sparse MobileNet v1, Use FC layer and no dimemsion keeping avgpool, last convolution is replaced by FC layer
Model sparsity (#nnz / sum(prunable layer weight) need to go over 66% if you use sparse model at any model sparsity, you must change XNNPACK sub-graph code, and tensorflow backend compiler .