tf.lite example with custom train model issue
terryll opened this issue · comments
I followed the CVInterop.Lite.Net example and everything work. However, when I custom train an tf.lite classification model (ssdMobileV2), the return results are byte[8] instead of float[8] for _outputTensor.Data.
If byte expected?
If so, how do I interpreted the results?
Thanks in advance,
Terry
The Tensor object can return byte[] if Tensord.DataType is UInt8:
https://www.emgu.com/wiki/files/emgutf/2.8.0/document-lite/html/8f3e0a49-e7c4-3cb0-7213-1f4fe1ae9227.htm
You can follow this instructions for model loading issues:
https://www.emgu.com/wiki/index.php/My_model_doesn%27t_work#Your_demo_works.2C_I_replaced_with_my_model_and_it_fails
You should always make sure your own model can be loaded and perform inference in python first, then replicate the same pyhon routine in C# with Emgu TF Lite.
Hope that helps.
Thanks for the suggestion. It ends up that I did not resigned the image causing unexpected results. However, the processing speed is much longer than expected. If I understand correctly, the tf.lite may need to rebuild based on the particular CPU in order to get the performance expected. I gave up that option for now.
I try to use emgu.tf. When I run the example (CVInterop.NetCore), there is an error message "Error CS0246 The type or namespace name 'ConfigProto' could not be found ". Did I missed something on the installation ?
I have went through the document and google the error, But not able to go pass this compile error on the sample program.
Thanks in advance,
Terry