EfficientNetb0-Yolo speed slow

[Code-Fight]

Hi @AlexeyAB , Thanks for you work.

I try to run EfficientNetb0-Yolo on 1080ti,But it only 10 fps.

The EfficientNetb0-Yolo only 3.7 BFLOPs,why it runing so slow?

I would like to increase the inference speed,what should I do?

Thank you very much.

[LukeAI]

depthwise convolution models like efficientnet are very slow on GPU. If you need GPU realtime, consider yolov3-spp or csresnext50

[Code-Fight]

Hi, @LukeAI
I got it ,thank you very much Can I running on CPU realtime? I try to run on CPU by openCV ,but it faild. What should I do? Thanks

[AlexeyAB]

Ask there, when will be added EfficientNet-Yolo to the OpenCV-dnn https://github.com/opencv/opencv/issues/15987

At least there were added Swish and Mish activations: https://github.com/opencv/opencv/pull/15808

[Code-Fight]

Hi @AlexeyAB Thanks for you reply. I will follow it

[WongKinYiu]

@Code-Fight @AlexeyAB

Our YOLOv3-tiny-PRN can run 48 fps on i7-8750H CPU and 71 fps on i9-9900K, respectively, using opencv dnn module (80 classes). https://github.com/opencv/opencv/pull/15739 Maybe you can also considerate about it before efficientnet-YOLO-PRN is supported by opencv.

[Code-Fight]

Hi @WongKinYiu ， Thank for your work for "YOLOv3-tiny-PRN". I hava used this model and it work well. Thanks again.

[nyj-ocean]

@LukeAI @WongKinYiu @AlexeyAB I found that both of efficientnet and cspresnext50 use groups in their .cfg like following:

[convolutional]
batch_normalize=1
filters=128
size=3
groups=32
stride=1
pad=1
activation=leaky

Since comment like https://github.com/AlexeyAB/darknet/issues/4447#issuecomment-561611786

depthwise convolution models like efficientnet are very slow on GPU

Why can cspresnext50 achieve real-time on GPU?

[WongKinYiu]

Because it uses group convolution, not depth-wise convolution.

groups=1 is equals to vanilla convolution, its base is equal to k. groups=k is equals to depth-wise convolution, its base is equal to 1. In ResNeXt50, groups=32, the base is 4~32, so its speed is between above two groups setting.

CSPResNeXt50 can achieve real-time on GPU is because it has 40% less computation than YOLOv3.

Also you can see EfficientNet-EdgeTPU, even though it has 7 times computation then EfficientNet, it run faster than EfficientNet. So if you want get real-time performance on mobile GPU using depth-wise convolution, maybe you have to make its computation lower to 0.4 BFLOPs, for example, ThunderNet.

[ashuezy]

@WongKinYiu Your so aweasome that i don't have words. Fabulous work with opencv-integration.

[nyj-ocean]

@WongKinYiu Thanks a lot. That lets me learn more about group convolution and depth-wise convolution.

[mmaaz60]

@Code-Fight @AlexeyAB

Our YOLOv3-tiny-PRN can run 48 fps on i7-8750H CPU and 71 fps on i9-9900K, respectively, using opencv dnn module (80 classes). opencv/opencv#15739 Maybe you can also considerate about it before efficientnet-YOLO-PRN is supported by opencv.

Hi @WongKinYiu

I ran the speed test of this network on the Intel CPU. It looks like it is almost 5 times slower than the Tiny Yolov3 PRN network on CPU as well. Below are the results,

OpenCV: 3.4.10-pre (https://github.com/opencv/opencv/tree/377dd04224630e835cce8c7d67e651cae73fd3b3) CPU: Intel(R) Core(TM) i5-6200U CPU @ 2.30 GHz Hard Drive Type: HDD Display: Off Yolov3-Tiny-PRN: 21.62 FPS EfficientNetB0-Yolov3: 4.72 FPS

It looks like depth wise convolutions are slow on CPU as well. Any thoughts?

Thanks

[AlexeyAB]

Take a look at the comparison: https://github.com/AlexeyAB/darknet/issues/5079

[WongKinYiu]

also you can use efficientnet-lite instead. https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/lite

[mmaaz60]

It would be interesting to implement efficientnet-lite-yolov3 in this repo.

[AlexeyAB]

I added https://github.com/AlexeyAB/darknet/blob/master/cfg/efficientnet-lite3.cfg

[becauseofAI]

I added https://github.com/AlexeyAB/darknet/blob/master/cfg/efficientnet-lite3.cfg

@AlexeyAB Where can we see the performance?

[WongKinYiu]

@AlexeyAB

efficientnet-lite3.cfg: top-1 69.1%, top-5 89.6%.

[WongKinYiu]

@AlexeyAB

efficientnet-lite3-leaky.cfg: top-1 73.0%, top-5 92.4%.

[linyib]

@AlexeyAB @WongKinYiu

Hi, Who has efficientnet-lite3.weights file, can you share it with me?