Tiny yolo on jetson tx2

[aprentis]

Hi everybody! I've trained yolo model for my own class of images. It works pretty good, but now I want to use this network on my nvidia jetson tx1 and TX2 .

So. Is it possible to use to use my weights file with tinyYolo model, which is much faster then traditional yolo model? Or should I train a new tiny model on my data to use it on jetson? Maybe I should use "small" model?

[AlexeyAB]

Hi, you should train a new tiny model.

[xhuvom]

the tiny-yolo.weights model (for COCO) provided on the website doesn't mention mAP. However, running ./darknet detector demo cfg/coco.data cfg/tiny-yolo.cfg tiny-yolo.weights detects nothing. But the tiny-voc model works fine with the mentioned mAP. I can't understand why is this!! Is it hard to detect 80 object classes with 15 layer network? @AlexeyAB

[Fred-Erik]

I think it's just because the weights file format has been changed, so you would have to learn a new model on COCO yourself.

[lzane]

@aprentis Did you try tiny yolo model on TX2? Could you please tell me how's the performance. I've train my own tiny yolo model with just two classes and only get 7 or 8 fps on my TX2 with Max-N mode. I wonder if I did something wrong?

[Fred-Erik]

@lzane I didn't try the demo mode you I can't tell you the FPS but when I use the test mode I get "Predicted in 0.042975 seconds" in Max-P mode, so that's about 20 FPS. But I have also forced my TX2 to not scale back when not in use using ~/jetson_clocks.sh, I noticed that makes quite a difference.

[lzane]

@Fred-Erik Thanks, I'll try to run ~/jetson_clocks.sh before testing.

[lzane]

I pull AlexeyAB/darknet and try demo mode with my TX2 in Max-N mode, and it reach 21FPS. Thank you very much.

On Dec 15, 2017, 12:26 AM +0800, Fred-Erik notifications@github.com, wrote:

@lzane I didn't try the demo mode you I can't tell you the FPS but when I use the test mode I get "Predicted in 0.042975 seconds" in Max-P mode, so that's about 20 FPS. But I have also forced my TX2 to not scale back when not in use using ~/jetson_clocks.sh, I noticed that makes quite a difference. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[Fred-Erik]

Okay! Have you also tried running jetson_clocks.sh and using Max-P mode? Do you get even higher FPS then?

[lzane]

@Fred-Erik No, But I don't think it will get higher FPS. The Max-N mode is more powerful than Max-P Model.

Look at the following table from http://www.jetsonhacks.com/2017/03/25/nvpmodel-nvidia-jetson-tx2-development-kit/

Mode	Mode Name	Denver 2	Frequency	ARM A57	Frequency	GPU Frequency
0	Max-N	2	2.0 GHz	4	2.0 GHz	1.30 Ghz
1	Max-Q	0		4	1.2 Ghz	0.85 Ghz
2	Max-P Core-All	2	1.4 GHz	4	1.4 GHz	1.12 Ghz
3	Max-P ARM	0		4	2.0 GHz	1.12 Ghz
4	Max-P Denver	2	2.0 GHz	0		1.12 Ghz

[Fred-Erik]

Ah I see! But when I compare between Max-N mode and after running jetson_clocks.sh, it does get about 10 times faster from 3 FPS to about 27 FPS. So I'm curious if AlexeyABs fork would perform even better.

[lzane]

Do you mean that you have faster FPS in Max-P mode than Max-N mode??? how can?

I have try AlexeyABs’ repo and reach about 24FPS with playing window on TX2 in Max-N mode.

---

Zane Lee 李泽帆

website: https://www.lzane.com github: https://www.github.com/lzane

On Dec 15, 2017, 9:14 PM +0800, Fred-Erik notifications@github.com, wrote:

Ah I see! But when I compare between Max-N mode and after running jetson_clocks.sh, it does get about 10 times faster from 3 FPS to about 27 FPS. So I'm curious if AlexeyABs fork would perform even better. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[hahakid]

only 5Hz for 1080P on Tx1