1. Build a small Keras CNN for MNIST-style images [N, 28, 28, 1] (channels-last). 2. Convert with tf2onnx.convert.from_keras (SavedModel path optional). 3. Validate with onnx.checker. 4. Compare ...