CHaiDNN icon indicating copy to clipboard operation
CHaiDNN copied to clipboard

CHaiDNN normalization input

Open eerobert opened this issue 6 years ago • 6 comments

I have a network trained with pixel/255.0 as normalized input

however, i cannot get the correct results. i wonder do chaidnn accept input of pixels/255.0 as input layer? any examples?

eerobert avatar Jan 03 '19 09:01 eerobert

i notice that the xiInputRead will call only loadMeanSubtractedDatafromBuffptr and there are no other loadNormalizedInputDatafromBuffptr(inp_mode == 1) or loadPixelMeanSubtractedDatafromBuffptr (inp_mode == 2)

So does it mean it do not support normalized input with pixel/255.0?

eerobert avatar Jan 03 '19 11:01 eerobert

I suggest you to create your own input layer if the ones provided don't fit your needs, it's simple.

NeutrinoXY avatar Jan 07 '19 10:01 NeutrinoXY

so are you suggesting to modify the xiInputRead function? can you elaborate how simple is to add a new xiInputRead function? thanks

eerobert avatar Jan 08 '19 06:01 eerobert

still stuck in modifying xiInputRead function. i wish to know what is the data type expected for such function?

eerobert avatar Jan 29 '19 03:01 eerobert

anyone can help?

eerobert avatar Feb 04 '19 03:02 eerobert

Hi, since the pixel/255 will result in input value (-1,1], what is the expected representation of such value? in mean-subtracted case, it is rather understandable. the example use truncated input (as in Trim2FP_Offline). However in the case of pixel/255, the Trim2FixedPoint will makes every input to 0

eerobert avatar Feb 27 '19 10:02 eerobert