Open USMANHEART opened 5 years ago
@USMANHEART
I fixed it. Update your code from GitHub.
@AlexeyAB Thanks for your kind reply. Let me check
Now one problem is resolved. But when I try to use GPU and OPENCV =1 I get this error
In file included from ./src/main.c:7:0: ./src/additionally.h:32:38: fatal error: opencv2/core/fast_math.hpp: No such file or directory
^
compilation terminated. make: *** [obj/main.o] Error 1
I just compiled and run Darknet successfull on Linux Debian 8.2 x64 + OpenCV 3.4.0
and Windws 7 x64 + OpenCV 3.2.0
I have followed that step. I comment line 32.
I am using Linux, OpenCV version 3.1
I have followed that step. I comment line 32.
It requires cuDNN >= 7.1.1 https://github.com/AlexeyAB/yolo2_light#yolo2_light
I have followed that step. I comment line 32.
It requires cuDNN >= 7.1.1 https://github.com/AlexeyAB/yolo2_light#yolo2_light
So you have CUDA 8.0
Now you should install cuDNN v7.1.4 (May 16, 2018), for CUDA 8.0
https://developer.nvidia.com/rdp/cudnn-archive
Could you please help me a little more regarding label names. I want to display the name of the labels also in the picture. Which file I should modify?
And what about the training procedure? I have modified your yolo config file which also include "input_calibration=" parameter is there any mistake in my training command?
@USMANHEART
This repository is made:
This repository only for Detection Yolo, not for Training. You should use this repository: https://github.com/AlexeyAB/darknet
Could you please help me a little more regarding label names. I want to display the name of the labels also in the picture. Which file I should modify?
Currently it draws labels only on Video, but doesn't draw labels for detection on Images: https://github.com/AlexeyAB/yolo2_light/blob/85fc1b388aa00f0243220ed48dac95ac18401b22/src/main.c#L80-L148
I have to design an INT model of yolov3. I need to train in INT. Can you please guide me which repo i must use? Do you have any repo that I can use for training or any reference where I can train an INT model?
@USMANHEART
Currently you can train your model by using Float-32bit using this repo: https://github.com/AlexeyAB/darknet as usual and must be used one of these cfg-files:
And then Detect objects by using your (Float-32bit) cfg & weights files by using the command with flag -quantized
using this repo https://github.com/AlexeyAB/yolo2_light so will be used INT8 inference:
./darknet detector test coco.names yolov3-tiny.cfg yolov3-tiny.weights -thresh 0.24 dog.jpg -quantized
I am actually using Yolov3. I have trained the model in Yolov3 . And I am having weights from yolo v3. But I noticed this thing in your code. I can't use quantized in yolo v3? It's written there that quantization only for yolo v2. Please make me clear for that one.
But I noticed this thing in your code. I can't use quantized in yolo v3? It's written there that quantization only for yolo v2.
This is old comment, that I will remove. Yolo v3 supports quantization.
If you trained your ownd yolov3.cfg
then add to the [net] section parameter
input_calibration = 15.497, 12.537, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40
and run detection:
./darknet detector test coco.names yolov3.cfg yolov3.weights -thresh 0.15 dog.jpg -quantized
I am unable to make this project. I have set GPU/CUDNN/OPENCV and all other arguments as 0. But I am unable to use make command
compilation terminated due to -Wfatal-errors. make: *** [obj/yolov2_forward_network.o] Error 1