The measured distribution for layer det_onnx/conv1 is different than expected. This could happened when the calibration set isn’t normalized.
Try preprocessing the dataset like in this tutorial
Thank you for your answer,
I wrote in my model_stcript file:
normalization1 = normalization([0.0, 0.0, 0.0], [255.0, 255.0, 255.0]
But this warning still appears, is it related to the dataset itself? My dataset is the icdar2017rctw_test. zip image set. Does it not meet the expected input distribution of conv1? But I don’t know the input expectation of conv, my model is the det_onnx model in pp OCR.
Hi @user61,
As suggested by Petar, the issue is most likely in the pre-processing. The pre-processing can be applied directly on the Hailo device (using the normalization command), or on the host CPU (in C++/Python). In both cases, you have to apply the same pre-processing (mean and standard deviation) that was used when training the model.
Please check which values are used in the original repository. Are you able to infer the ONNX with onnxruntime and get good results?