Compare MobileNetV2 deployment results

After the individual benchmark runs, you will view the results together in the dashboard. This makes it easier to compare average latency, and run-to-run variation.

Run on the Raspberry Pi:

    

        
        
course_support/serve_results.py \
  --results ~/executorch_on_arm/module-2-cnn-inference-with-quantization/results \
  --host 127.0.0.1 \
  --port 8080

Open on the development machine:

    

        
        
http://localhost:8080

The summary shows the three MobileNetV2 runs side by side:

Model	Runtime	Warmups	Runs	Mean (ms)	Median (ms)	Min (ms)	Max (ms)
mobilenetv2	PyTorch eager	10	50	72.55	72.15	71.77	76.96
mobilenetv2	ExecuTorch with XNNPACK	10	50	10.89	10.88	10.73	11.63
mobilenetv2	INT8 ExecuTorch with XNNPACK	10	50	4.01	3.98	3.95	4.56

The PyTorch eager and FP32 ExecuTorch with XNNPACK plots show the latency distribution and run-to-run variation for the first two paths:

Image Alt Text:MobileNetV2 PyTorch and FP32 XNNPACK latency plots.

The INT8 plots show the quantized path:

Image Alt Text:MobileNetV2 INT8 XNNPACK latency plots.

Use the dashboard, artifact sizes, and top-5 output to answer:

Question	Your conclusion
Which of the three runtime configurations best meets the latency and artifact-size requirements?
Which measurements support that choice?
Did the top-5 output remain useful after quantization?
What accuracy or application testing is still required before production use?

Compare MobileNetV2 deployment results

Nice work