Input
Uploads or a local webcam provide the frame for inference.
▧00THE QUESTION
EDGE AI / IMAGE CLASSIFICATION / UNCERTAINTY
What if a model could say “check me”?
Most inference demos stop at a label. This diagnostics layer keeps the second thought: how certain was the model, and should a person review the result?
↓ Explore the inferenceLIVE SIGNALreview is part of the output
01THE PIPELINE
One image.
Four signals.
An uploaded frame becomes a prediction, then gets measured for uncertainty before it reaches an edge workflow.
Prediction
A pretrained model returns the top candidate classes.
→Uncertainty
Entropy and confidence gap reveal when the choice is unclear.
≈Decision
A flag saves the frame and metadata for review or retraining.
!02THE SIGNAL
A RESULT FROM A PUBLIC MODEL
The answer is only half the story.
A public ImageNet model can give an unfamiliar image a visually unrelated label. These diagnostics make uncertainty visible instead of pretending the label is ground truth.
3.70ENTROPYHow spread out the model's belief is.
0.14CONFIDENCE GAPDistance between the leading predictions.
TRUEREVIEW FLAGThe result deserves a second look.
03THE MODEL
Give it an image.
See its hesitation.
A compact inference experience with model selection and the reasoning left visible.
A familiar ImageNet baseline for comparing uncertainty behavior.
INFERENCE RESULTDEMO OUTPUT
PREDICTED CLASSAwaiting image
CONFIDENCE--
ENTROPY--
GAP--
REVIEW--
Upload an image to see how the interface communicates a prediction and its uncertainty.
04USE IT ON YOUR MODEL
Download the tool.
Keep your model.
The diagnostics layer is separate from the model. Bring a TorchScript model or a custom loader.
Public baseline
Run a pretrained baseline on ImageNet images.
python diagnose.py image.jpgTorchScript model
Bring a serialized model and your own class list.
python diagnose.py image.jpg --model model.tsCustom PyTorch model
Expose a loader for project-specific setup.
python diagnose.py image.jpg --loader my_model:load_model05METHOD & LIMITS
A useful signal
is not a verdict.
Entropy and confidence gap are decision aids, not proof that a prediction is wrong. They tell us when the model's evidence is weak enough to invite human review.
Swap in a model trained for your domain, then keep the same diagnostics and logging layer.