An image
as given
what the model sees (32×32)
CIFAR-10 test images
The true label is in each filename, so an answer can be checked rather than admired.
All ten classes
loading the model…
Read the answer sceptically. This is not a CIFAR-10 result. ResNet-18
reaches ~95% trained the 100–200 epochs the literature uses; this is one epoch on a
CPU, undertrained on purpose, because what was being demonstrated is that the
platform carried a real workload end to end — not that the model is good. Expect
roughly two answers in three to be right.
CIFAR-10 is 32×32. Whatever you give it is centre-cropped and area-averaged down to that first, and the panel says what it did. A confident prediction about a 32×32 crop of a photograph is still a prediction about a 32×32 crop of a photograph.
CIFAR-10 is 32×32. Whatever you give it is centre-cropped and area-averaged down to that first, and the panel says what it did. A confident prediction about a 32×32 crop of a photograph is still a prediction about a 32×32 crop of a photograph.
The run that produced these weights
Why you can trust the export
The cluster serves
this model with serve.py, a PyTorch process. A browser cannot run that,
so this page runs the same weights through a different runtime. That
substitution is only worth something if it is checked — so
scripts/export-onnx.py re-evaluates both over the full test set and
refuses to write the export unless they agree.