Train yolov3 on my own dataset

Environment: Ubuntu 16.04

Paper: YOLOv3.pdf

How to train

Pre-requirements

• OpenCV

• CUDA

Otional but highly recommended, will be a lot faster.

Download model and try a test

0. Download source code

 git clone https://github.com/pjreddie/darknet
 cd darknet

Instead of this original edition of yolov3, there is another AlexeyAB edition of yolov3, chose whatever you want.

1. Change makefile ---> vi Makefile

• Want to use GPU ---> GPU=1

• Want to use openCV ---> OPENCV=1

• Want to use CUDNN

 CUDNN=1
 NVCC=/usr/local/cuda-10.0/bin/nvcc

2. Compile ---> make

3. Test

Download a pre-trained model to test --> wget https://pjreddie.com/media/files/yolov3.weights

Then test the dog picture ---> ./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg

If you are lucky, you will get a picture like this

Customize the dataset

0. Background

My project is to detect five different kinds of objects: lizard,bird,car,dog,turtle and I use labelImg to label my pictures. After that, prepare a folder to save all the pictures and another folder to save all the .xml documents.

Do not delete these xml files, they are needed for computing mAP.

1. Generate .txt file

.txt file that I prepared: train.txt,val.txtand a lot of picture_name.txt in folder labels.

train.txt: Store the paths of pictures used for training (without .jpg). One path per line.

val.txt: Store the paths of pictures used for testing (without .jpg). One path per line.

picture_name.txt: Store bounding box information of the picture, one object per line. It should be customized to

(object-class) (x_center) (y_center) (width) (height)

for example:

0 0.66015625 0.28515625 0.6171875 0.5390625

• Put folder labels and folder pictures under the same directory.

2. Change some relevant files and codes

• Copy voc.names , rename to animal.names, and change the class names, one class name per line.

bird
car
lizard
dog
turtle

The order of the names should match class name index in picture_name.txt.

• Copy voc.data, rename to animal.data.

 classes =  5                  # Class number
 train   =  data/train.txt     # Path to train.txt
 valid   =  data/val.txt       # Path to val.txt
 names   =  data/animal.names  # Path to animal.names
 backup  =  backup/

• Copy yolov3-voc.cfg in folder cfg, rename to animal.cfg, make the following changes:

[net]
# Testing
# batch=1            # Use it when you test the model
# subdivisions=1     # Use it when you test the model
# Training
batch=64
subdivisions=32      # Set it bigger if memory full
...              
...
...
learning_rate=0.001  
burn_in=1000
max_batches = 30000  # Change according to your needs
policy=steps
steps=10000,20000    # Change with max_batches
...
...
...
[convolutional]
size=1
stride=1
pad=1
filters=30           # filters = 3*(classes + 5)
activation=linear

[yolo]
mask = 0,1,2
anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326
classes=5            # change to your class number
num=9
jitter=.3
ignore_thresh = .5
truth_thresh = 1
random=1             # set to 0 if memory full
...
...
...

# There are three [yolo] layers in total, change all of them.
# In front of every [yolo] layer there is a [convolutional] layer, change filter number.

3. How it looks like after the change

I have 150 pictures for training and 30 pictures for testing, each group, so my folder data is like

train_images

000001.JPEG...000750.JPEG

train_labels

000001.txt...000750.txt

val_images

000001.JPEG...000150.JPEG

val_labels

000001.txt...000150.txt

animal.data

animal.names

train.txt

val.txt

4. Download a pre-trained weight ---> wget https://pjreddie.com/media/files/darknet53.conv.74

5. Start training ---> sudo ./darknet detector train data/animal.data cfg/animal.cfg darknet53.conv.74

When I finished training and want to visulize the loss, I found I did't save it. So maybe try the following code for training instead of the previous one:

sudo ./darknet detector train data/animal.data cfg/animal.cfg darknet53.conv.74 2>1 | tee visualization/animal.log

Test the result

0. Detect a single picture

• Switch the .cfg file to test mode:

[net]
# Testing
batch=1  # Use it when you test the model
subdivisions=1    # Use it when you test the model
# Training
# batch=64
# subdivisions=32     # Set it smaller if memory full
...    
...
...

• Recompile:

make clean and make

• Now choose a picture from test set and test:

sudo ./darknet detector test data/animal.data ./cfg/animal.cfg ./backup/animal_20000.weights ./data/val_images/000133.JPEG

I trained a total number of 128w pictures, my results:

1. Compute Recall and IoU

• Definition:

True posotive ---> tp

True negative ---> tn

False positive ---> fp

False negative ---> fn

Precesion = tp/(tp+fp)

Recall = tp/(tp+fn)

• Compute

Find function validiate_detector_recall in examples/detector.c, make two changes:

1' Change the path: list *plist = get _path("data/val.txt") // to my path to val.txt

2' Change the for loop as directed by fix error in validate_detector_recall #952

++total;
box t = {truth[j].x, truth[j].y, truth[j].w, truth[j].h};
float best_iou = 0;
// for(k = 0; k < l.w*l.h*l.n; ++k){
for(k = 0; k < nboxes; ++k){  // Use this line
    float iou = box_iou(dets[k].bbox, t);
    if(dets[k].objectness > thresh && iou > best_iou){
        best_iou = iou;

make clean and make

./darknet detector recall data/animal.data cfg/animal.cfg backup/animal_20000.weights

• My Results:

On test set ---> RPs/Img: 1.37 IOU: 71.64% Recall:90.00%

On train set ---> RPs/Img: 1.03 IOU: 91.60% Recall:100.00%

2. Compute AP and mAP

• Definition

P (Precision) ---> for one image in one class

AP (AveragePrecision) ---> for all images in one class

mAP (MeanAveragePrecision) ---> for all images in all classes

• Compute

1' Detect test set and save the results classname.txt to darknet/results folder.

./darknet detector valid data/animal.data cfg/animal.cfg backup/animal.backup -out ""

2' Download voc_eval.py

3' Build compute_map.py:

from voc_eval import voc_eval

rec,prec,ap = voc_eval('/home/karry/Documents/Yolo/darknet/results/{}.txt', '/home/karry/Documents/Yolo/darknet/val_xml/{}.xml', '/home/karry/Documents/Yolo/darknet/val.txt', 'classname','.')

print('ap',ap)

The first path is for the detection results in the first step.

The second path is for .xml documents of test set.

The third path is for val.txt which contains picture name of test set, no picture path, no .jpg, only picture name. One name per line.

The forth one is class name.

The fifth one is the path to save annots.pkl file.

4' Run compute_map.py.

Get AP for this class.

• My results

Class	AP
Dog	0.8522680776014109
Car	0.8850865226727296
Bird	0.7822037102682264
Lizard	0.9223655913978495
Turtle	0.7221268965829435
mAP	0.8328

A debug experience

0. Problem description

When I run the test code for a single image, I got pictures with correct lable but no bbox:

1. How I solved it

Open program src/image.c, find function draw_box_width(), change the loop function:

from

for(i = 0; i < w; i++)

to

for(i = 0; i <= w; i++)

Then recompile ---> make clean and make

Run the detection command:

sudo ./darknet detector test data/animal.data ./cfg/animal.cfg ./backup/animal_20000.weights ./data/val_images/000133.JPEG

Problem solved.

2. Why it happens

Generally it's because my picture is too small, only 128 x 128 pixels.

When I use the test code, I'm actually calling function test_detector() in example/detector.c, and he will call draw_detections() defined in program src/image.c

void draw_detections(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes)
{
    int i,j;

    for(i = 0; i < num; ++i){
        char labelstr[4096] = {0};
        int class = -1;
        for(j = 0; j < classes; ++j){
            if (dets[i].prob[j] > thresh){
                if (class < 0) {
                    strcat(labelstr, names[j]);
                    class = j;
                } else {
                    strcat(labelstr, ", ");
                    strcat(labelstr, names[j]);
                }
                printf("%s: %.0f%%\n", names[j], dets[i].prob[j]*100);
            }
        }
        if(class >= 0){
            int width = im.h * .006;
            ...
            ...
            ...

In the function above, im.h = 128 so width will be 0. And then he will pass these parameters to draw_box_width()

void draw_box_width(image a, int x1, int y1, int x2, int y2, int w, float r, float g, float b)
{
    int i=0;
    for(i = 0; i < w; i++){
        draw_box(a, x1+i, y1+i, x2-i, y2-i, r, g, b);
    }
}

At that time, i = 0, w = 0, the for loop won't run, he won't call draw_box() so I can't get my boundingbox. By adding a = can make the program run and that's how I solved the problem.

• Pictures with their im.h smaller than 167 can have the same problem.