python-yolo-training-with-jupyter-notebooks

Table of Contents

• Summary
• References
• How to use
  • Labeling images
  • Updating the yolov4-tiny-custom.cfg file
  • Uploading the custom-data folder to Google Drive
  • Running code blocks of the custom Jupyter notebook in Google Colab
  • Init and update git submodules
  • Utilize your trained weights in the sample object detector app

Summary

Repository demonstrating how to train a custom CNN model based on yolo-v4-tiny architecture. This can be utilized for image classification, image localization or object detection applications.

References

• Github repository of Techzizou - yolov4-tiny-custom_Training
• Medium post by Techzizou - Train a custom yolov4 tiny object detector using google colab
• Jupyter notebook of Techzizou
• darknet

How to use

Labeling images

Add your images (e.g. .JPG files) to the img folder and label your images. For labeling images utilize for example one of the following:

• labelimg
• Yolo_mark

Modify train.txt to encompass all JPG files for training purposes, and adjust test.txt to incorporate all JPG files for validation. Also update the obj.names file to list your classes.

Updating the yolov4-tiny-custom.cfg file

Adjust in the yolov4-tiny-custom.cfg the width, height, batch, subdivision, max_batches, steps, classes and filters hyperparameter values. Refer to section 3(a) Create and upload the labeled custom dataset “obj.zip” file to the “yolov4-tiny” folder on your drive here or copy the original file which can be found here yolov4-tiny-custom.cfg and replace this yolov4-tiny-custom.cfg.

Uploading the custom-data folder to Google Drive

Upload the data-custom folder to Google Drive.

Running code blocks of the custom Jupyter notebook in Google Colab

Open custom-yolov4-tiny-training.ipynb in Colab and run each code block

Init and update git submodules

Execute the following steps in order to initialize the git submodule containing the object detector app:

# On Unix terminals
git submodule init --update
# or on Windows OS
git submodule init
git submodule update

Utilize your trained weights in the sample object detector app

Copy custom .weights, .cfg and .names files to appropriate destination pathes:

# On Unix terminals
cp <your download folder path>/yolov4-tiny-custom_best.weights python-object-detection-with-yolo-and-opencv/weights
cp data-custom/yolov4-tiny-custom.cfg python-object-detection-with-yolo-and-opencv/cfg
cp data-custom/obj.names python-object-detection-with-yolo-and-opencv/object-names
# On Powershell (Windows OS)
Copy-Item "<your download folder path>\yolov4-tiny-custom_best.weights" -Destination "python-object-detection-with-yolo-and-opencv\weights"
Copy-Item "data-custom\yolov4-tiny-custom.cfg" -Destination "python-object-detection-with-yolo-and-opencv\cfg"
Copy-Item "data-custom\obj.names" -Destination "python-object-detection-with-yolo-and-opencv\object-names"

In python-object-detection-with-yolo-and-opencv install the pip package requirements if not yet done and launch the object detector app via

cd python-object-detection-with-yolo-and-opencv
python object_detector_app.py --model custom-yolov4-tiny