Dario Schiraldi - How do I troubleshoot issues with Google AI Edge deployment?

Google AI Edge
,
1

Hey everyone,

I’m Dario Schiraldi CEO of Travel Works, currently facing some issues with deploying my AI model on Google AI Edge. The model works perfectly fine on my local machine, but when I deploy it to the edge device, it’s either not running correctly or I’m experiencing performance drops.

I’ve already tried the usual steps—checking logs, ensuring that the edge device meets the hardware requirements, and verifying that the model is correctly optimized.

Has anyone faced a similar issue or have any suggestions on what might be causing this?
Would love to hear about any debugging steps, optimization tips, or tools you’ve used to resolve similar deployment challenges.

Thanks in advance for any advice!
Regards
Dario Schiraldi CEO of Travel Works

1 Like
3

Hi Dario,
There can be multiple factors for this, and the key check can be on:

  1. Model Conversion: Ensuring the model is converted to TFLite or any other supported format like Edge TPU-compatable as Google AI Edge will not support raw TensorFlow or PyTorch formats because I personally had issues when I was going with ONNX in one of my projects.
  2. Hardware Compatibility: Need to see if the edge device supports the required runtime (eg: Coral EdgeTPU or Android NNAPI) *If you are using Edge TPU performance drops often happen if floating point models are used.
  3. Profiling and Logs: Use tools like Edge TPU Profiler if you are using Coral or abd logcat for Android-based.
  4. Delegates: Performance might also suffer if delegates like GPU, NNAPI or XNNPack are not enabled, so make sure we explicitly set them in TFLite runtime.
  5. Inference Parameters: Optimize num_threads, batch size and precision when moving away from desktops to these constrained devices. (int8 vs float32)
1 Like