darknet to tensrrt model

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.75k stars 7.96k forks source link

darknet to tensrrt model #4665

Open ou525 opened 4 years ago

ou525 commented 4 years ago

I trained a yolov3-spp model based on darkent, and at the same time I converted it to a tensorrt model, but I found that it has more false detections than tests based on darknet. precision:float32, no int8

AlexeyAB commented 4 years ago

@ou525

Can you show examples of detections on Darknet and TRT?
Try to detect by using Darknet with flag -letter_box at the end of detection command

isra60 commented 4 years ago

How do you convert a darknet model to TensorRT??

ou525 commented 4 years ago

@AlexeyAB @isra60 i use the nvidia demo, https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/tree/restructure/yolo/apps/trt-yolo

i add the flag -letter_box, but I don't find any difference

AlexeyAB commented 4 years ago

but I found that it has more false detections than tests based on darknet.

Can you show screenshot of detections?

Darknet with ./darknet detector test ... -letter_box
TRT

ou525 commented 4 years ago

darknet_box trt_box @AlexeyAB these two pictures are the most obvious place for this scene.

AlexeyAB commented 4 years ago

Do you get this issue only for yolov3-spp.cfg or also for yolov3.cfg and yolov3-tiny.cfg?
Can you show screenshot of detection on this image without -letter_box?

ou525 commented 4 years ago

@AlexeyAB sorry, I retested the add and without flag -letter_box, it does appear similar to the false detection under trt, but it does not always appear on the same picture. At the same time, I checked the content of letter_box. Is it caused by the different input of trt's blobFromImage () and darknet's letterbox_image ()?

i only use the yoov3-spp.cfg, trained without -letter_box.

I extracted the feature values of layer81 of the yolov3-spp network, which seems to be very different from those under trt. darknet.txt trt.txt

AlexeyAB commented 4 years ago

@AlexeyAB sorry, I retested the add and without flag -letter_box, it does appear similar to the false detection under trt, but it does not always appear on the same picture.

Do you mean that each run of the same command with the same image gives a different result? ./darknet detector test ... test.jpg -letter_box

ou525 commented 4 years ago

The test results are unique. The first results may be obtained by other model tests.

I replaced trt's letter_box with direct resize, and now the result seems to be normal.

AlexeyAB commented 4 years ago

You should

either use resize instead of letterbox in trt
or train your model on Darknet with [net] letter_box=1 in cfg-file (so you can use letterbox in trt)

ou525 commented 4 years ago

Thanks for your suggestions, I will try to do it.

ttdd11 commented 4 years ago

@ou525 How did you convert trained .weights and .cfg to tensorRT? Been struggling with this for a while, why guidance would be greatly appreciated.