1
4

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 3 years have passed since last update.

DNNのモデル毎の処理量(GOP)をまとめる

Last updated at Posted at 2019-05-19

#目的
DNNのモデル毎の処理量(GOP = Giga Operations)をまとめる。
1画面(フレーム)あたりの処理量の意味。
正確な情報が必要な場合は、**「情報5」**とかを見られるのがいいと思う。
(「情報1」~「情報4」に比して)

:pencil:
GOP(Giga Operations)
GOPS(Giga Operations per Second)

#DNNのモデル毎の処理量

情報1

文献『AN ANALYSIS OF DEEP NEURAL NETWORK MODELS FOR PRACTICAL APPLICATIONS』を読む。

model Gop 備考
MobileNet-v1 1.4 ※目分量の精度悪し
MobileNet-v2 0.7 ※目分量の精度悪し
AlexNet 2.2 ※目分量の精度悪し
GoogleNet 2.5 ※目分量の精度悪し
ResNet-50 8.0 ※目分量の精度悪し
VGG19 39 ※目分量の精度悪し
VGG16 31 ※目分量の精度悪し

情報2

資料『JETSON AGX XAVIER AND THE NEW ERA OF AUTONOMOUS MACHINE』を読む

model Gop 備考
Image Recognition
MobileNet 0.6 224x224,
17GOPS@30Hz
AlexNet 0.7 227x227,
22GOPS@30Hz
GoogleNet 2 224x224,
60GOPS@30Hz
ResNet-50 4 224x224,
120GOPS@30Hz
VGG19 20 224x224,
600GOPS@30Hz
Object Detection
YOLO-v3 65 416x416,
1,950GOPS@30Hz
SSD-VGG 91 512x512,
2,730GOPS@30Hz
Faster-RCNN 172 600x850,
5,160GOPS@30Hz

情報3

文献『From Model to FPGA: Software-Hardware Co-Design for Efficient Neural Network Acceleration』を読む。

model Gop 備考
VGG16
Image classification
30.68 13 Conv layers
YOLO Tiny
General object detection
5.54 9 Conv layers
Customized Network
Face alignment
0.1046 ※1 9 Conv layers
※1 資料では、104.6 Mop と記載。

情報4

文献『Machine learning for embedded deep dive』を読む

model Gop 備考
Inception v1 3.2
Tiny Yolov3 5.6
Tiny Yolov2 7
ResNet50 7.7
VGG16 30
Yolov2 36
Yolov3 65
SSD 117

情報5

以下などに沢山情報があります。

#まとめ
特に、なし。

#今後
まだ、単に、情報を集めているレベルです。
コメントなどあれば、お願いします。:candy:

#関連(本人)
文献『AN ANALYSIS OF DEEP NEURAL NETWORK MODELS FOR PRACTICAL APPLICATIONS』を読む。

資料『JETSON AGX XAVIER AND THE NEW ERA OF AUTONOMOUS MACHINE』を読む

文献『From Model to FPGA: Software-Hardware Co-Design for Efficient Neural Network Acceleration』を読む。

文献『Machine learning for embedded deep dive』を読む

1
4
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
1
4

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?