artyze / yolo_segmentation Goto Github PK

Thank you for the code. I've tried to train our own dataset in GPU. When I set GPU and CUDNN in preprocessor, I got many unresolved external symbol error such as activate_array_gpu, which is defined in activation_kernels.cu. What can I do to fix them?

Hi,
ThanX great work as such this project.
I have started reproducing your great result, but a no pretrained weight to check our result.

Can you send me pretrained weights file to below,

[email protected]

As of now, we can not get good result of our project, but we try hardly to reproduct and will send you our result sometimes.

Best regards.

My environments
OS: ubuntu 18.04

I prepared data like below.

data file

classes= 4
train = /home/yes89929/segmentation/yolo_segmentation/data/list/cityscapes/train.lst
valid = /home/yes89929/segmentation/yolo_segmentation/data/list/cityscapes/val.lst
names = /home/yes89929/segmentation/yolo_segmentation/data/label.names
backup = backup
results = /home/yes89929/segmentation/yolo_segmentation/result/
labels = /home/yes89929/segmentation/yolo_segmentation/aachen_resize/

train.lst file

/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknr8wpn0000n3a6bqi0tcty2_leftImg8bit.png
/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknr5qx27000u3a6bef9amy90_leftImg8bit.png
/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknjitehd11na3a6b8jfct2n0_leftImg8bit.png
/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknpkd4t50xx83a6bc6e4q18x_leftImg8bit.png
/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknpw2l3p03b53a6bd02uuz5o_leftImg8bit.png

Terminal command
./darknet segmenter train cfg/maskyolo.data cfg/segment.cfg cfg/instance_segment_161000.weights

Output message
I added some code for printing function name and image path.

load_data_seg 
load_image_color: /home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknjkfr0y11pk3a6bxdpgf8pl_leftImg8bit.png
load_image: /home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknjkfr0y11pk3a6bxdpgf8pl_leftImg8bit.png
stbi_load: /home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknjkfr0y11pk3a6bxdpgf8pl_leftImg8bit.png
fopen result: false

Message printed from original code.

"annot load image "/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknjkfr0y11pk3a6bxdpgf8pl_leftImg8bit.png
STB Reason: can't fopen

I made code to test fopen(), and confirmed fopen() work.
code

#include <iostream>

int main(int, char**) {
    std::cout << "Hello, world!\n";
    FILE* f = fopen("/home/yes89929/segmentation/yolo_segmentation/data/cityscapes/leftImg8bit/train/cknskurbm000e3a6blrlysviw_leftImg8bit.png", "rb");
    if (!f)
    {
        printf("fail\n");
    }
    else
    {
        printf("success\n");
    }
}

Output message

Hello, world!
success

I have no idea about this problem.

Hello ,
i have successfully launched the training after changing the code and using backup/segment.backup made by segment.backup when i do training for the first time .
My problem is that avg value doesn't decrease and the rate of learning doesn't increase during the training.
How can i fix this issue ?

• https://i.imgur.com/sQTukMF.png

Thanks

c++ 怎么调用这个修改后的darknet

Hi, I got "CUDA Error: an illegal memory access was encountered" when I was testing an image with [logistic], I also tried [softmax] which run succefully but pred.png was totally dark.
Would you mind helping to solve this problem?

Thanks
Julien

Couldn't open file: /backup//segment_40.weights

cityscapes have lane annotation?

@ArtyZe are you looking into the implementation of yolact into your current repo do u have any future thoughts on thsi

Nice Job!
I download your repo, with GPU=1/OPENCV=1/CUDNN=1 .
make -j8
Error occurs in ./src/image.c L1464
It seems you don't declare i,j,k and c,h,w before.When I annotate the wrong code, compile goes well, but I can't use the original detetctor demo func.
I can compile with the original darknet code on my aarch64 TX2.
I'm looking forward to your reply, Thx

按照ReadMe以及提供的模型，输出结果图如下：

命令行：./darknet segmenter test ./demo/demo.data demo/segment1.cfg demo/segment1.backup demo/orig.png

麻烦帮忙看下，thx

Hello ,
What should i change to training custom dataset for exemple ?
dataset with one class

Hi.
Thanks you for sharing good code.

Now I`m training your model. But It seems like using 1 GPU.

Can I use multiple GPUs by just using arguments on terminal?

OR

need I modify some code?

Can anybody share with me a sample data set and train.txt to be include in data set configuration file. including the format of train.txt. Thanks.

Hi, Thanks for your project. Could you help to share the weights and cfg file? It is hard for me to train from scratch by my estimation. I would like to get your weights as pre-trained weights for my project.

您好，能提供一下cfg和weights吗？上传到百度或google云盘，或者发我邮箱：[email protected]

I've noticed that in predict_segmenter method, it will call
if (l.out_w && l.out_h && l.out_c){ return float_to_image(l.out_w, l.out_h, l.out_c, l.output); }
image float_to_image(int w, int h, int c, float *data) { printf("c = %d, w = %d, h = %d\n", c, w, h); image out = make_image(w,h,1); for(int ii=0; ii<w*h; ii++){ int max_id = 0; float max_value = 0; for(int jj = 0; jj < c; ++jj){ if(max_value < data[jj*w*h + ii]){ max_value = data[jj*w*h + ii]; max_id = jj; } } if(max_value > 0.05){ out.data[ii] = (float)max_id; }else{ out.data[ii] = 0; } } return out; }

I thought that value of l.output will be between 0 and 1 as softmax is used at end, but they are not (I saw most of them are <0 or >1 values). What is l.output and why > 0.05 is used? Thanks much!

Hello ,
I tried to train my own dataset using this project but it failed ,
For the first time , i didn't get any errors when runing this command :

./darknet segmenter train data/obj.data cfg/segment.cfg segment.backup
this is the output for the first time ( but training doesn't work ):

33 logistic x entropy 173056
Loading weights from segment.backup...Done!
the net n is 34
Learning Rate: 1e-07, Momentum: 0.9, Decay: 0.0005
2357
Saving weights to backup/segment.weights

When i tried again , i get this error :

   33 logistic x entropy                             173056
   Loading weights from segment.backup...Done!
   the net n is 34
   Learning Rate: 1e-07, Momentum: 0.9, Decay: 0.0005
   2357
   Saving weights to backup/segment.weights
   Cannot load image "/home/starinfo/yolo_segmentation/mask/Img_810_img.txt"
   STB Reason: can't fopen

How can i solve those issues ?
Thanks

I was wondering if the pre-trained model for tiny version is available online.

at first,thank you for the project.when I try to use the opencv lib,there are some small bug finded:
./src/image.c: In function ‘test_resize’:
./src/image.c:1464:9: error: ‘k’ undeclared (first use in this function)

Hi, to make use of ur repo on window. I try to copy your files into the AlexeyAB files and build darknet with msvs but however it fails with error such as

C1083 Cannot open include file: 'darknet.h': No such file or directory darknet c:\darknet-master\darknet-master\src\activations.h 3

I'm not sure if i did it in a correct way. Do you have any guide to install it on window ?

I think I can imagine what cfg setting is.
and whats the sample label rule do you set? is it a "binary image+src image" for every target? or some rule like "0 2532 12 2624 12 2881 10 3012 10 ..."

@ArtyZe Can you please elaborate the process , datasets preparation and training for multiple classes?

@ArtyZe
Hi, thank you for the project it's really helpful, but I want to know how I can organize my dataset (data file)for the training and what tools do you use for the segmentation-annotation. ????

I can not find Find_replace function in "segmenter"

Can you pls upload the weights and cfg files somewhere

@ArtyZe thanks for the source code and sharing the model . i have few queries

1. what is the difference btw your instance segmeter.c and segmenter.c code , since the mode u share gives output for both the function
2. the time take to predict the on an image is >27 secs which very less , assumption is since its built on c/c++ the fps should be good any ideas on this
3. Any procedure for training on instance segmentation for custom data , please share

Thanks in advance

while(1){
image aug = random_augment_image(im, 0, .75, 320, 448, 320, 320);
show_image(aug, "aug", 1);
free_image(aug);

for(k = 0; k < c; ++k){
  for(j = 0; j < h; ++j){
    for(i = 0; i < w; ++i){
      int dst_index = i + w*j + w*h*k;
      int src_index = k + c*i + c*w*j;
      im.data[dst_index] = (float)data[src_index]/255.;
    }
  }
some variable is not define, so some problem happens;

./src/image.c: In function ‘test_resize’:
./src/image.c:1467:22: error: ‘h’ undeclared (first use in this function)
for(j = 0; j < h; ++j){
^
compilation terminated due to -Wfatal-errors.
Makefile:85: recipe for target 'obj/image.o' failed

have u consider of combine detection and segmentation?