Looong01 / ML-project

Machine Learning Final Project

[ Models ] & [ Path ] to be downloaded from Google Drive

• Dataset & a pickle dump of the preprocessed data (Please download all the files in this path and place them in the project root path)
  • ./dataset/
• Pretrained weights on COCO dataset
  • ./yolo/data/yolov4_coco_pretrained_weights.pth
• Weights of our best YOLO model (mAP 63.47%)
  • ./yolo/data/yolov4_test.pth
• vgg 16 model (for OpenCV demo)
  • ./vgg.pth

Transfer learning notes

• Before you run the Transfer_learning.ipynb, please make sure to download the dataset folder from the google drive and put it in the project root path

Opencv_demo notes

• You can either use the vgg model from the google drive to run the demo, or run the Transfer_learning.ipynb and save a model (Resnet 50 or vgg 16).

YOLO method notes

• ./yolo/
  • Main structure derives from reference project [1]
• ./yolo/annotation_preprocess.ipynb
  • Make annotation files for training YOLO models.
• ./yolo/train.ipynb
  • Train a YOLO model.
• ./yolo/predict.ipynb
  • Detect images or videos (based on yolo.py in the folder).
• ./yolo/eval.ipynb
  • Evaluate a YOLO model (generate ground-truth files and detection-results files).
• ./yolo/yolo.py
  • Use the given YOLO model to make predictions.
  • You may want to change the model weights path in this file to try out different models.
• ./yolo/data
  • Store train / val / test dataset annotations.
  • Store classes and anchors data.
  • Store font file for detection results labeling.
• ./yolo/loss_data
  • Store training losses
• ./yolo/map
  • Code for computing mAP (please run main.py after running eval.ipynb) [2]
• ./yolo/nets/
  • YOLO V4 implementation [1]
• ./yolo/utils/
  • Utility code

References

1. yolo4-pytorch
2. mAP