When to Stop Training
Training an object detection model requires careful consideration of when to halt the training process. Stopping training at the appropriate moment can significantly impact the model's performance, generalization capabilities, and efficiency.
This guide aims to provide researchers, developers, and practitioners with valuable insights to determine the optimal stopping point during model training.
If this is your first training, you may follow the Starter Checklist.
Monitor Training Progress
During training, continuously monitor the progress of the model. Keep track of critical performance metrics, such as:
Loss
mAP
IOU
Iterations
scores. Loss and mAP will be show on a graph just like below:
All metrics can wildy vary by:
Data variety
Data size
Annotation accuracy
Model size
Numbers below are only provided for setting an initial ground for newcomers.
Loss
Loss is shown with blue points on training graph and represents how far is the model accuracy from the training data provided.
Training Loss can be monitored to keep track on model accuracy and overtraining. Several ranges of Training Loss value can indicate:
**2.0 ≥** Loss
A generic model that can give an idea on how generic and accurate the database is. For non-specific databases, this should give you a somehow accurate model, ready to assess training procedure.
As shown in the graph above, loss values around 2.0 may not produce accurate models.
**1.0 ≥** Loss
A specialized database can be used to achieve loss values below 1.0 and are good indicators to a specialized database.
**0.5 ≥** Loss
Fine tuned model that is ready to test and deploy. After reaching this value, improving loss value can take a lot longer time than the first stages of the training.
mAP
The mAP (mean average precision) metric combines both precision and recall to provide a comprehensive evaluation of the model's accuracy in detecting objects in an image.
It is calculated by comparing prediction boxes with ground truth annotations overlapping.
mAP as a percentage doesn't reflect if a detection is correct or not. It reflects average overlapping on bounding box areas between training dataset and prediction.
During training, reaching values around %90 is generally considered a good model. Values above %90 usually considered as over-fitting.
IOU
IOU (Intersection over Union) measures the overlap between predicted and true bounding boxes for individual object detections. mAP evaluates the overall performance of the object detection model across all object categories, considering both precision and recall.
You can track each IOU in Training Window loggings:
Fine Tuning
Training Time
Define a maximum training time budget based on available computational resources and project constraints. If the model does not achieve satisfactory performance within the allocated time, consider stopping training and exploring other approaches such as:
Manually analyze annotation accuracy
Check class variety
Choose different model sizes and batch sizes
Increase database size
Over-Fitting
Avoid overfitting by monitoring the training and validation losses. This usually happens when training loses most of its momentum going downwards and starts creeping down.
However, in specialized databases or cases over-fitting is not always a bad thing. Proved you have enough data, an over-trained model can serve you well.
Balancing Time and Performance
Balance the training time with the desired model performance. In some cases, additional training iterations may improve performance, but the returns may diminish over time. Weigh the benefits against the computational cost and the urgency of the project.
Usually, depending on class numbers and database size, training process length can vary between a day or a week.
Starter Checklist
Database:
Model:
Training (stop if):
Last updated
Was this helpful?