GhostNet: More Features from Cheap Operations - 75.7% top-1 (better than than MobileNetV3)

[AlexeyAB]

• paper: https://arxiv.org/abs/1911.11907v1

• source: https://github.com/iamhankai/ghostnet

• model: ghostnet.cfg.txt

---

GPU GeForce RTX 2070 - Darknet framework (GPU=1 CUDNN=1 CUDNN_HALF=1) CPU Intel Core i7 6700k - Darknet framework (OPENMP=1 AVX=1)

• darknet.cfg - GPU 360 FPS - CPU 63 FPS - 0.400 BFlops - 7.3M params - 61.1% Top1
• darknet19.cfg - GPU 179 FPS - CPU 14 FPS - 2.793 BFlops - 20.8M params - 72.9% Top1
• GhostNet-1.0 - GPU 61 FPS - CPU 12 FPS - 0.117 BFlops - 5.0M params - xx.x% Top1
• MixNet-M-GPU - GPU 82 FPS - CPU 4.6 FPS - 0.533 BFlops - 11.9M params - 71.5% Top1
• EfficientNetB0 - GPU 110 FPS - CPU 6.3 FPS - 0.450 BFlops - 4.9M params - 71.3% Top1
• darknet53.cfg - GPU 85 FPS - CPU 4.8 FPS- 9.285 BFlops - 41.6M params - 77.2% Top1

---

• GhostNet-1.0 - 5.0M params - 0.117 BFlops - xx.x% Top1 - xx.x% Top5 - MY URL
• GhostNet-1.0 - 5.2M params - 0.141 BFlops - 73.9% Top1 - 91.4% Top5 - Official
• MobileNetV3 - 5.4M params - 0.219 BFlops - 75.2% Top1 - --- Top5
• GhostNet-1.3 - 7.3M params - 0.226 BFlops - 75.7% Top1 - 92.7% Top5 - Official
• EfficientNetB0 - 4.9M params - 0.450 BFlops - 76.3% (71.3%) Top1 - 93.2% (90.4%) Top5 - MY URL
• MixNet-M - 5.0M params - 0.360 BFlops - 77.0% (71.5%) Top1 - 93.3% ( 90.5%) Top5 - https://github.com/AlexeyAB/darknet/issues/4203

Comparison table: https://github.com/AlexeyAB/darknet/issues/4203#issuecomment-548955416

---

maybe better than mobilenetv3, efficientnet, mixnet..., etc. https://github.com/iamhankai/ghostnet/issues/1

We measure the actual inference speed on an ARM-based mobile phone using the TFLite tool, we use single-threaded mode with batch size 1:

---

---

---

[AlexeyAB]

@WongKinYiu Hi, I added ghostnet.cfg.txt so you can try to train it for 600 000 iterations with batch_size=192 (mini_batch_size=96).

It can be trained for 2 weeks on GeForce RTX 2070.

May be it can be fast on CPU/Neurochips (OpenCV-dnn).

[WongKinYiu]

@AlexeyAB thank you!

[AlexeyAB]

@WongKinYiu I just added dropout after avg-pooling. So if you already started training, you can download new cfg-file and continue training.

[duducode]

我在cpu度测试darnet-19 速度1.3秒一张图片，测试图片大小500*374，请问哪里出问题了

[WongKinYiu]

Do you build darknet with (OPENMP=1 AVX=1)? And which CPU do you use?

[duducode]

Do you build darknet with (OPENMP=1 AVX=1)? And which CPU do you use?

thanks，i will try update my code and try again

[WongKinYiu]

@AlexeyAB

top-1 1.5%, top-5 5.6%.

[AlexeyAB]

@WongKinYiu Thanks! ghostnet.cfg.txt - 1.5% Top1? is it near ~0 ? Can you share cfg/weights file?

[WongKinYiu]

@AlexeyAB

ghostnet.cfg.txt ghostnet_final.zip

[iamhankai]

@WongKinYiu This repo may help: https://github.com/d-li14/ghostnet.pytorch

[WongKinYiu]

@iamhankai thank you very much.

[rsek147]

@WongKinYiu I have tested with your cfg/weights. The result is almost same as yours. I started training few days ago with almost same your cfg except for batch and subdivisions. My result is top-1 30%, top-5 64% (300000 iterations, continuing). This is strange despite of using almost same cfg.

[AlexeyAB]

@WongKinYiu Thanks!

@rsek147 Can you attach your cfg-file?

[DayBreak-u]

@AlexeyAB @WongKinYiu I think the ghostnet.cfg.txt is wrong，it can refer to https://github.com/d-li14/ghostnet.pytorch/blob/master/ghostnet.py , i use ghiost moudle in mobilenetv3 Small , it get 20% top1 after 20000 iters with 256 batch size

[maohu-96]

@AlexeyAB hi！ When I use GhostNet for training, the loss -nan，How to solve？ I am completing a fruit test（class=1） using the ghostnet.cfg file above. What parts should I modify?

Thanks！

[wsypy]

@WongKinYiu Hi, have you been retraining Ghostnet since?If so, can you share .cfg and .weights files?

[WongKinYiu]

I did not get good result.