industrial_abnormal_yolo

Using `darknet` to train yolov3 model

How to run

1. Clone Darknet

There are 2 ways :

You can directly use codes in darknet folder. (You may need follow tutorial on the websit shown in the second way on Windows platform to recompile)
You can simply delete existed darknet folder and run git clone https://github.com/AlexeyAB/darknet.git in the same folder. Then you can follow the tutorial on this website https://github.com/AlexeyAB/darknet .

2. Prepare data

Put all of images in darknet/train/ or any other path you like.
For each one of images, generate a .txt file with same name in which is like this:
```
0 0.6895424836601307 0.44140625 0.05147058823529412 0.0693359375
0 0.31147875816993464 0.47998046875 0.05024509803921569 0.0732421875
```
Meaning of each number:
1. Class number
2. The x-coordinate of normalized center point
3. The y-coordinate of normalized center point
4. The x-coordinate of normalized target's width
5. The x-coordinate of normalized target's height
You can easily get tutorial about this by Google.
Generate a txt file in which each line is a path to an image for training. Name it as train.txt
Prepare validation set in the same way.
Create obj.name file in which each line corresponds to the name of each class number.

Contact me if you want to try industrial abnormal data.

3. Setting

Open darknet/Makefile and you can set your compile configurations.
```
GPU=0
CUDNN=0
OPENCV=0
OPENMP=0
DEBUG=0
```
0 means disabling this option, and 1 means enabling.
Open darknet/cfg/. Then you can see a lot of yolo models. Choose one of them, for example yolov3.cfg . At the beginning, you set model parameters. Here are some parameters you may change:
```
# Testing uncomment following two lines when testing
# batch=1
# subdivisions=1
# Training
batch=64
subdivisions=64
width=416
height=416
```
ATTENTION Most important part to set is the class number. There are three [yolo] part in the model. Change all of their classes to yours and set all of the filters in [convoluntional] parts just before [yolo] to be 3*(classes+1+4) .

Create obj.data and its content is like following:

classes= 1						# number of classes
train = ./train.txt		# path to train set folder
valid = ./train.txt		# path to valid set folder
names = ./obj.names		# path to obj.names created in previous step
backup = ./output/		# path to folder which save the model weights

4. Train

Download pre-trained weights by https://pjreddie.com/media/files/darknet53.conv.74.

run command in darknet folder (Mac or Linux)

It will generate train log in train.log

./darknet detector train obj.data cfg/yolov3.cfg darknet53.conv.74 >> train.log

5. Test

The model weights are in the path you set in obj.data. You can just use following command to test one image

./darknet detector test obj.data cfg/yolov3.cfg output/yolov3_1000.weights train/20190628141831095.jpg

6. Evaluation

Run python log_transformer.py to transform train log.
Run python iou_draw.py to visualize IOU.
Run python loss_draw.py to visualize loss.
Run command ./darknet detector valid obj.data cfg/yolov3.cfg output/yolov3_200.weights

You can find valid results in darknet/results/comp4_det_test_abnormal.txt

The content looks like following:
```
20190628144629832 0.182705 177.825806 588.880005 2448.000000 749.905396
20190628144629832 0.096559 1019.806885 1163.286255 1929.701904 1384.190796
20190628144629832 0.095504 1.000000 153.420364 2448.000000 643.597717
20190628144629832 0.088120 1005.774109 885.752563 1432.554199 1645.059204
20190628144629832 0.081529 978.469360 294.323853 1779.933472 2048.000000
```
Meaning of numbers:
1. Time
2. Confidence Ratio
3. The x-coordinate of left-top point
4. The y-coordinate of left-top point
5. The x-coordinate of right-bottom point
6. The y-coordinate of right-bottom point
Copy .txt file of each validation sample into darknet/lables/
Run python convert_gt.py to transform coordinates.
Run python convert_results.py to transform validation log.
Visualization part is based on mAP.

Runcd mAP/scripts/extra

python intersect-gt-and-dr.py to get the intersection of validation results and ground truth.
Then goto mAP folder and run python main.py. You can see several analysis results in folder results.

Abingcbc/industrial_abnormal_yolo