Results
Evaluate your model’s strengths and weaknesses and find areas for improvement.
What are my training results?
What are my training results?
Your results show you which images your model is predicting correctly and incorrectly. Correct predictions have green labels and incorrect predictions have red labels. The width of the label bar represents how confident the model was in that prediction.
Hovering over a predicted label will show the true label you gave that image. The more correct the predictions, the better the model is performing.
You can view and sort your images in different ways to check:
- Whether your model is successfully learning all the labels with View > All Images selected.
- Approximately how well it will work on new images with View > Test Images selected. Learn more about test images.
- Which images and labels confuse your model by selecting your individual labels in the sidebar.
How do I see labels that are confused with one another?
How do I see labels that are confused with one another?
Select your individual labels in the sidebar to see results specific to that label. Lobe will show you which other labels are commonly confused with this one.
Try to spot what similarities exist between the images that confuse the model and the images in the true label. For example, you may notice the same background color across the confused images and the images in your label. If you notice these patterns, try to collect varying images that look like the confused image to better teach Lobe what patterns to ignore. In this example, collect more images in your selected label with varying background colors.
How do I see images that confuse the model?
How do I see images that confuse the model?
To see where your model is most confused, look at the most confident incorrect predictions and the least confident correct predictions.
- Select View > Correct First to view your predictions sorted by most confident to least confident. Try to find any patterns where your model was very confident in its wrong prediction.
- Select View > Incorrect First to view your predictions sorted by least confident to most confident. See where your model is least confident with its correct predictions.
Collect more variations of images that have similar patterns to these two cases to help your model improve.
What are test images?
What are test images?
Test images are a random subset of your examples that Lobe hides from your model during training. Lobe automatically splits your examples into two parts:
- Random 80% is used to train your model.
- Random 20% is held out and used to test your model.
See the results on only your held out test images by selecting View > Test Images. If your overall performance on All Images is very high and your performance with Test Images is lower, your model may be memorizing your examples instead of learning to generalize to unseen images. This is often called overfitting.
To help prevent overfitting, you can:
- Collect more images overall so the training set includes more variations.
- Make sure your images don’t look too similar to each other.
Results look good, what should I do next?
Results look good, what should I do next?
When you are happy with the results, you should try out the model in the real world! Experiment on new images with Play, or Export the model to use in your app. Try to trick your model with new images that test its limits, and add those images as examples to improve your model.
How will my model perform in the real world?
How will my model perform in the real world?
Look at the results on your test images to get a better sense how your model will perform on unseen data. Select View > Test Images to see the random subset of your examples that your model did not use during training.
Can I fix labeling errors?
Can I fix labeling errors?
Yes, you can fix labeling errors directly from results. You may find some predictions marked as incorrect because they were incorrectly labeled examples. Fix these labeling mistakes by clicking on the predicted label text box and typing or selecting the true label.