Auto training pipeline for object detection models
This pipeline trains object detection model(YOLOv8) using real time inference data. It is for automating the supervised learning, specifically cutting out the manual labelling task and training the model for it to remember the object as per the label we want.
- Python >= 3.8
- GPU (optional but prefferable)
pip install autotrain-vision
Pip package : autotrain-vision
For more details, refer to the GitHub repository.
Error: Rebuilt the library with Windows, GTK+ 2.x or Carbon support.
Try
sudo apt install libgtk2.0-dev pkg-config
pip install opencv-contrib-python
from autotrain_vision import AutoTrain
at = AutoTrain(combined_folder="/path/to/local/folder")
at.run()
data_folder
(str, required): Path to local folder to store the new dataprev_data_folder
(str): Path to local previous folder containing images and labels foldernew_weights
(boolean): True if to not use any previous dataabs_yaml_file
(str): Absolute path to the YAML file for given prev_data_folderdraw_bb
(boolean): True to draw bounding boxes on previous image datasetimage_threshold
(int): Number of images to capture for creating new datasetnumber_aug
(int): Number of times to apply augmentationsepochs
(int): Number of epochs for trainingmap_threshold
(float): value<=1 ; Threshold to compare mAP50 scoreinference
(boolean): True to perform the inference on live feedinference_threshold
(float): value<=1 ; Threshold for inference confidence scorecamera_range
(int): Range of camera indexes to look for
weights.pt
: Weights file for trained model.
- Creates new annotated data
- Combines previous annotated data with newly captured
- Annotates the data with visible bounding boxes, given images and corresponding v8 txt label files
- Generates new weights (pt) file for given dataset
- Generates analysis graphs and metrices for validation
- Add multiple objects annotation in single frame
- RnD on Florence capabilities for giving in text prompt to ZSD model
- RnD on discarding faulty annotations from procured dataset