Documentation Link: https://github.com/rshah1990/Smartathon-theme1/blob/main/Hackathon.pdf
- We have used YOLO V5 pretrained model trained on COCO dataset as starting point. We have retrained model on our data set to reduce training time & more accurate model
- We have tried data augmentation to reduce class imbalance
- Used customized stratified train test split to avoid data bias
- Split 80-20% of class with minimum number of samples , if number of samples is 1 than move it into training
- For next class identify how many classes is already available in training (since its multilabel ) & how many ideally should be based on 80-20 split
- We randomly sample those many numbers of rows from the training if number of classes less than expected else don’t sample anything move everything to test set
- Experiment tracking using weights and biases
- Different technique tried to improve data Quality