How the input images should be quantized before feeding to the model and how the model outputs should be dequantized when executing in GAP8

[gwgknudayanga]

Hi,

I trained a model for object detection simple feeding the raw inputs to the network. So the input pixels are in 0 to 255 range.
When the model is running in GAP8, can i feed these raw pixel values directly? Or otherwise do i need to explicitly quantized these to a particular range? (i hope it's just reducing 128)

• Also how can i dequantize the model outputs from GAP8?
• With ReadImageFromFile we can read images. But how can I read the .npy file into the c run the file because my data is not images?

[gwgknudayanga]

@marco-fariselli could you please help me in this regard?