About this module
In Module 1, you compared PyTorch eager, portable ExecuTorch, and ExecuTorch with XNNPACK for two transformer workloads.
Module 2 focuses on FP32 and INT8 models using XNNPACK, rather than repeating the portable-kernel comparison. In this module, you will compare three ways to run MobileNetV2, a lightweight convolutional neural network for image classification.
MobileNetV2’s convolution-heavy structure provides a contrast with the transformer workloads in Module 1 and can map more cleanly to optimized CPU kernels.
| Model and runtime path | Role in the comparison |
|---|---|
| FP32 PyTorch eager MobileNetV2 | Provides the full-stack latency and classification reference. |
| FP32 ExecuTorch with XNNPACK | Shows the effect of deploying the CNN via ExecuTorch and through optimized CPU kernels. |
| INT8 ExecuTorch with XNNPACK | Shows the effect of quantization on artifact size, latency, and classification output. |
The comparison considers runtime latency, .pte artifact size, delegation, and classification output.
Course complete
Nice work
You completed Optimize and deploy AI inference with ExecuTorch on Arm.