zeki096 / YoloV3_CustomData

Train YOLOV3 on custom dataset using pretrained weights.

YoloV3

---

YoloV3 Simplified for training on Colab with custom dataset.

A very 'smal' Coco sample imageset is added in the folder called smalcoco. This is to make sure you can run it without issues on Colab.

Full credit goes to TSAI, and if you are looking for much more detailed explainiation and features, please refer to the original source.

You'll need to download the weights from the original source.

1. Create a folder called weights in the root (YoloV3) folder
2. Download from: https://drive.google.com/open?id=1LezFG5g3BCW6iYaV89B2i64cqEUZD7e0
3. Place 'yolov3-spp-ultralytics.pt' file in the weights folder:

• to save time, move the file from the above link to your GDrive
• then drag and drop from your GDrive opened in Colab to weights folder

Run this command to check if the setup is ok.

python train.py --data data/smalcoco/smalcoco.data --batch 10 --cache --epochs 25 --nosave

For custom dataset:

1. Download and Annotate images in yolo format. Tool
2. images and annotated labels should go here.

data
  --customdata
    --images/
      --img001.jpg
      --img002.jpg
      --...
    --labels/
      --img001.txt
      --img002.txt
      --...
    custom.data #data file
    custom.names #your class names
    custom.txt #list of name of the images you want your network to be trained on. Currently we are using same file for test/train

5. As you can see above you need to create custom.data file. For 1 class example, your file will look like this:

  classes=1
  train=data/customdata/custom.txt
  test=data/customdata/custom.txt
  names=data/customdata/custom.names

6. As you it a poor idea to keep test and train data same, but the point of this repo is to get you up and running with YoloV3 asap. You'll probably do a mistake in writing to custom.txt file. This is how our file looks like (please note the .s and /s):

./data/customdata/images/img001.jpg
./data/customdata/images/img002.jpg
./data/customdata/images/img003.jpg
...

7. You need to add custom.names file as you can see above. For our example, we downloaded images of drones. Our custom.names file look like this:

drone

8. drone above will have a class index of 0.
9. For COCO's 80 classes, VOLOv3's output vector has 255 dimensions ( (4+1+80)*3). Now we have 1 class, so we would need to change it's architecture.
10. Copy the contents of 'yolov3-spp.cfg' file to a new file called 'yolov3-custom.cfg' file in the data/cfg folder.
11. Search for 'filters=255' (you should get entries entries). Change 255 to 18 = (4+1+1)*3
12. Search for 'classes=80' and change all three entries to 'classes=1'
13. Since you are lazy (probably), you'll be working with very few samples. In such a case it is a good idea to change:

• burn_in to 100
• max_batches to 5000
• steps to 4000,4500

14. Don't forget to perform the weight file steps mentioned in the sectio above.
15. Run this command python train.py --data data/customdata/custom.data --batch 10 --cache --cfg cfg/yolov3-custom.cfg --epochs <epochs> --nosave