Expose training metrics for custom classification models

[marnunez]

I've been training a state classification model and noticed there's no way to tell how well it actually performs. The model compiles with accuracy metrics and uses a 20% validation split, but the history from model.fit() is discarded.

Would be useful to save the final accuracy and validation accuracy to .training_metadata.json so users can see if their model is overfitting or if they need more training samples.

Currently the metadata only has:

• last_training_date
• last_training_image_count

Adding something like final_accuracy, final_val_accuracy, and final_loss would help users understand if their model is actually learning or just memorizing the training data.

I have a small patch that does this (~15 lines changed), happy to create a PR if there's interest.

[sudheerlucky]

also it better to have wait call system to use train button to avoid multiple training at same time

kind of wail till training to complate then next one to start...

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.