Deployment of own quantized model

Report bugs
#1

Question/Issue:
When trying to deploy our customized quantized int8 model (onnx format) using the Edge Impulse, the page would report a series of errors. However, when we deploy the pruned fp32 model (onnx format), it works fine.

Project ID:
1132429

Context/Use case:
[Provide context or use case where the issue is encountered]

Summary:
[Provide a concise summary of the bug]

Steps to Reproduce:

  1. Upload the int8.onnx file through the “Upload your model” tab
  2. After uploading, let Edge Impulse start processing job

Expected Results:
Successful compiling as fp32 model

Actual Results:
A series of error report

Reproducibility:

  • [ ] Always

Environment:

  • **Edge Impulse Online Platform

Logs/Attachments:

1
1780×496 38.5 KB

#2

Hello @minghao8926,

Thanks for sharing the logs, I’ll see with the core engineering team if LayerNormalization operation could be added to the supported operation.

In the meantime, you can still quantize (post-training quantization) your model by adding enough representative features. This easiest way to do this is through our Python SDK:

Let me know if you encounter any issues.

Best,

Louis