pytorch模型compression-prune,sparsity,quantization

sparsity

Pruning individual weight elements is called element-wise pruning, and it is also sometimes referred to as fine-grained pruning

Filter Pruner

相连卷积层变换

conv + bn + conv 变换

non-serial data-dependencies

element-wise-summed and fed into a third Convolution

channel pruner

mobile_net quantization 实验数据

图片大小 27 x 33 pixel

model	size	top1(accuracy)	inference time
未量化模型	9.2M	99.1	7.46 ms(cpu) / 4.85 ms(gpu)
量化后模型	3.1M	98.9	1.75 ms(gpu)
半精度模型	4.7M	99.1	5.28 ms(gpu)

图片大小 136 x 91 pixel

model	size	top1(accuracy)	inference time
未量化模型	9.2M	99.1	16 ms(cpu) / 4.83 ms(gpu)
量化后模型	3.1M	98.9	5.99 ms(gpu)
半精度模型	4.7M	99.1	5.13 ms(gpu)

图片大小 620 x 827 pixel

model	size	top1(accuracy)	inference time
未量化模型	9.2M	99.1	480 ms(cpu) / 21.2 ms(gpu)
量化后模型	3.1M	98.9	123 ms(gpu)
半精度模型	4.7M	99.1	17.6 ms(gpu)