Live inference on this server's CPU
Everything else in this demonstration is a measurement taken earlier and labelled with the time it was taken. This screen is the exception: drop an image in and both models run now, on Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz, through ONNX Runtime, and tell you how long they took.
Upload
The detector was trained on two DAGM texture families and the anomaly model was fitted on normal images of a third. A photograph of anything else will produce a confident and meaningless answer, which is itself worth demonstrating: a model only means something relative to the distribution it was shown.
Result
Nothing run yet.
What the models saw
Detections
Anomaly map
computed just now Nothing on this screen is precomputed. The
model files were built by scripts/export_onnx.py and
scripts/train_anomaly.py; the inference, the decode and the timing all happen
inside the request. Runtime: onnxruntime (MIT), numpy and Pillow
(BSD-3-Clause, MIT-CMU). PyTorch is not installed in the serving
environment.