notesum.ai

Published at November 10

RL-Pruner: Structured Pruning Using Reinforcement Learning for CNN Compression and Acceleration

cs.CV

cs.AI

Released Date: November 10, 2024

Authors: Boyao Wang¹, Volodymyr Kindratenko¹

Aff.: ¹University of Illinois Urbana-Champaign

Arxiv: http://arxiv.org/abs/2411.06463v1

Architecture	Reward Strategy	Pruned Accuracy / FLOPs Ratio / Para. Num. Ratio at Different Sparsity Levels
(Accuracy, FLOPs, Para. Num.)	Reward Strategy	20%	40%	60%	80%
VGG-19 (72.89, 418.63M, 39.33M)	Accuracy-Based	73.09 / 22.62 / 35.30	73.11 / 35.11 / 62.18	73.00 / 43.55 / 79.46	71.25 / 55.71 / 89.34
	FLOPs-Based	73.05 / 26.09 / 36.58	72.65 / 39.40 / 63.12	72.50 / 47.33 / 81.04	71.51 / 53.96 / 90.38
	Parameter-Based	73.00 / 24.54 / 36.02	72.08 / 38.94 / 63.70	71.98 / 47.43 / 81.98	71.67 / 56.84 / 91.44
ResNet-56 (72.25, 127.93M, 0.86M)	Accuracy-Based	70.79 / 51.01 / 37.67	68.33 / 73.36 / 66.19	59.65 / 87.02 / 89.32	37.74 / 95.51 / 97.77
	FLOPs-Based	71.21 / 49.92 / 34.53	68.36 / 75.69 / 68.81	61.49 / 84.50 / 88.13	40.22 / 94.06 / 97.28
	Parameter-Based	71.41 / 51.05 / 37.28	68.46 / 75.86 / 70.80	60.68 / 87.19 / 88.32	40.02 / 95.65 / 97.69
DenseNet-121 (79.83, 907.93M, 7.05M)	Accuracy-Based	78.60 / 38.59 / 36.22	77.08 / 67.26 / 64.60	73.33 / 85.72 / 82.60	69.10 / 95.26 / 94.50
	FLOPs-Based	79.08 / 44.15 / 37.91	76.43 / 72.77 / 67.41	71.67 / 87.51 / 83.84	67.71 / 95.44 / 94.67
	Parameter-Based	78.44 / 39.76 / 39.10	76.21 / 68.32 / 67.25	73.85 / 85.98 / 84.14	69.73 / 94.95 / 95.01
GoogLeNet (77.62, 535.66M, 6.40M)	Accuracy-Based	77.61 / 31.78 / 34.09	76.78 / 53.88 / 60.55	74.17 / 72.65 / 81.05	69.96 / 89.85 / 95.23
	FLOPs-Based	77.62 / 35.67 / 34.01	77.00 / 58.45 / 61.06	74.16 / 75.66 / 82.17	69.37 / 86.96 / 95.39
	Parameter-Based	77.64 / 32.21 / 34.49	76.47 / 55.57 / 60.75	73.85 / 71.10 / 81.18	69.47 / 88.73 / 94.88
MobileNetV3-Large (56.18, 7.47M, 4.03M)	Accuracy-Based	56.26 / 33.01 / 36.22	55.32 / 61.63 / 63.95	52.66 / 79.98 / 83.81	42.41 / 92.78 / 95.85
	FLOPs-Based	56.26 / 33.01 / 36.22	55.32 / 61.63 / 63.95	53.28 / 80.17 / 83.74	41.65 / 93.14 / 95.91
	Parameter-Based	56.76 / 33.12 / 36.51	55.74 / 61.46 / 65.23	53.08 / 79.54 / 84.53	44.10 / 92.57 / 96.08