heechul / fastspconv-cvpr2020

Code for "Fast Sparse ConvNets" CVPR2020 submissions

cvpr2020

This directory contains all of the ARM computational kernels used the paper "Fast Sparse ConvNets" submitted to CVPR 2020.

spmm-NxM-[scalar,neonfma].c - Sparse Matrix Multiplication with an unroll of N in the HW dimension and a block size of M in the channel output dimension.

dwconv-KxKsSpP-[scalar,neonfma].c - Depthwise Convolution with a filter of KxK, a stride of S and symmetric padding of P.

gavgpool-[scalar,neon]-xN.c - Global average pooling unrolled over N rows.

conv-3x3s2p1c3x4-[scalar,neonfma]-KxK - Full 3x3 stride 2 convolution with HWC input and CHW output. Operates on 3 input channels (only supported) and 4 output channels in the inner loop. Produces a KxK output.

TF-Lite Models

The MobileNetV1 models have a block size of 4 in the last block, otherwise they are unstructured.

The MobileNetV2 models have a block size of 2 from block 11 onwards, otherwise they are unstructured. The exception is the width 1.8, 80% sparse model which unstructured throughout.

The first full convolution and the final fully connected layer are both dense in all models.

EfficientNet models are fully unstructured and the final fully connected layer is sparse.

Model	Top-1 Accuracy	Sparsity	Latency (ms) SD 835	Download
MobileNetV1 .75	64.4%	90%	21	link
MobileNetV1 1.0	68.4%	90%	31	link
MobileNetV1 1.4	72.0%	90%	58	link
MobileNetV2 .8	65.2%	85%	26	link
Cache Aware MobileNetV2 1.0	69.7%	85%	33	link
MobileNetV2 1.15	70.2%	85%	40	link
MobileNetV2 1.4	72.0%	85%	54	link
MobileNetV2 1.8	74.9%	80%	102	link
MobileNetV2 2.0	74.5%	85%	93	link
EfficientNet B0	75.1%	80%	80	link
EfficientNet B1	76.7%	85%	110	link