Partial-Order-Pruning

Partial Order Pruning: for Best Speed/Accuracy Trade-off in Neural Architecture Search, CVPR 2019

---

Abstract

Achieving good speed and accuracy trade-off on target platform is very important in deploying deep neural networks. Most existing automatic architecture search approaches only pursue high performance but ignores such an important factor. In this work, we propose an algorithm "Partial Order Pruning" to prune architecture search space with partial order assumption, quickly lift the boundary of speed/accuracy trade-off on target platform, and automatically search the architecture with the best speed and accuracy trade-off. Our algorithm explicitly take profile information about the inference speed on target platform into consideration. With the proposed algorithm, we present several "Dongfeng(东风)" networks that provide high accuracy and fast inference speed on various application GPU platforms. By further searching decoder architecture, our DF-Seg real-time segmentation models yields state-of-the-art speed/accuracy trade-off on both embedded device and high-end GPU.

---

Experiments

1.We conduct backbone architecture searching experiments on TX2:

模型(Model)	ImageNet Val. Top-1 Acc.	Lat.(ms)/224x224/batch_size=1	GPU Mem./224x224/batch_size=50	GPU Mem./1024x1024/batch_size=1	FLOPS
东风一(DF1)	69.78%	2.5	1863	1042	746M
东风二(DF2)	73.92%	5.0	2867	1567	1.77G
东风二甲(DF2A)	76.00%	6.5	4844	2356	1.97G
ResNet-18	69.0%	4.4	2784	1397	1.8G
ResNet-50	75.3%	10.6	9914	4488	3.8G
ResNet-101	76.4%	-	14900	6813	7.6G

Comparing to ResNet18/50/101, our Dongfeng networks achieve similar accuracy with much lower latency on target platform. We have evaluated the GPU memory consumptions of inferencing with different backbone networks, in BVLC_CAFFE. Our Dongfeng networks consume much less GPU memory. Specifically, our DF2A has a similar accuracy with ResNet-50/101, but consumes 2x~3x less GPU memory. Faster inference and Less GPU memory consumption are of great importance in deploying, and also enables researcher to implement more fancy, sophisticated algorithm.

2.With our Dongfeng backbone network (searched on TX2), we conduct decoder architecture search experiments on both TX2 and 1080Ti:

Model	Cityscapes mIoU (Val/Test)	FPS(TX2/TensorRT-3.0.4)	FPS(1080Ti/TensorRT-3.0.4)	FPS(1080Ti/TensorRT-3.0.4)	FPS(Titan X/Caffe)
Resolution	1024x2048	640x384	1024x2048	1024x1024	1024x2048
DFlite-Seg-d8	71.7/-	-	157.4	263.4	45.7
DF1-Seg-d8	72.4/71.4	92.7	136.9	232.6	40.2
DFlite-Seg	73.4/-	-	118.4	202.5	33.8
DF1-Seg	74.1/73.0	71.4	106.1	182.1	30.7
DF2-Seg1	75.9/74.8	44.9	67.2	-	20.5
DF2-Seg2	76.9/75.3	39.9	56.3	-	17.7

3.Dongfeng models are designed for GPU platforms. We further conduct backbone and decoder architecture searching experiments on Snapdragon 845 CPU platform:

Model	Cityscapes mIoU (Val/Test)	FPS(Snapdragon 845)
Resolution	1024x2048	640x384
PL1A-Seg	68.7/69.1	52.0

---

Snapshots

欢迎使用“东风”系列模型，万事俱备，只欠东风！ Everything is ready, all you need is Dongfeng!

1.Please set weight_decay to 0.0001 during finetuning, otherwise performance will be negatively affected.

2.I would recommond using DF2 rather than DF2A in dense prediction. Although its accracy on ImageNet Val. is lower, DF2 provides better segmentation accuracy in our preliminary experiments. I attribute this to the larger receptive field of DF2.

df1.caffemodel https://drive.google.com/open?id=1yA9DLSy3PEMQD3R92vKr6CEaOJ0NrOVm

df2.caffemodel https://drive.google.com/open?id=1K0QPFD6XtKnMsrrOLSarnk3C1zmhZqpC

df2a.caffemodel https://drive.google.com/open?id=1H5T-nz1D2DCLtma-alkR_CxGAHKewbql

df1seg.caffemodel https://drive.google.com/open?id=1v-UCb1VIHGtIR9eXiPgdUu9wkDpemNTW

df1seg_mergebn.caffemodel https://drive.google.com/open?id=17ZROC9dJAN8dxkpvTzHQRTS9soCwBOXJ

df2seg1.caffemodel https://drive.google.com/open?id=1mCdozRO4BxDV-NS6secKuvaTNWEFcv_u

df2seg1_mergebn.caffemodel https://drive.google.com/open?id=1RfdYtc7YzM5zYoANsRiB-XJMkvYfy_mv

df2seg2.caffemodel https://drive.google.com/open?id=1D7bgq7h9OUQVY4LYA-x0B2FY8pnVgz3o

df2seg2_mergebn.caffemodel https://drive.google.com/open?id=1FtqRSEN90ynTgMeGH3ee5DC8ubvSHZ3z

df-lite_seg_mergebn.caffemodel https://drive.google.com/open?id=1se9wAkZFyNGYInjrhtXTMIZy39ucMaDu

Citation

If Dongfeng networks are useful for your research, please consider citing:

@inproceedings{li2019partial,
author = {Xin Li,
	Yiming Zhou,
	Zheng Pan,
	Jiashi Feng},
title = {Partial Order Pruning: for Best Speed/Accuracy Trade-off in Neural Architecture Search},
booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2019}
}