In Pytorch no normalize but has toTensor

1

I converted the PyTorch to CoreML and then used it in my iOS project.
But I found in PyTorch there is no normalize code to do image preprocess , so it means I should do image preprocess option in CoreML.
Does it have same effect between toTensor and Normalize

2

You can do image normalization with torchvision.transforms.Normalize(mean, std, inplace=False), if that’s what you want to do.

3

@Marc_Dumon

Screen Shot 2020-01-08 at 12.22.50
Screen Shot 2020-01-08 at 12.22.50608×792 24.7 KB
my model file without scaler
Screen Shot 2020-01-08 at 12.23.12
Screen Shot 2020-01-08 at 12.23.12762×408 10.5 KB
standard model file
I met one weird issue - I input the image and then predict wrongly.I used CoreML but the model doesn’t have scaler

@ptrblck

4

Could you rephrase because I have no idea what your problem is.

5

@Marc_Dumon because onnx doesn’t support preprocess so in CoreML I must do the job manually.So I mean the model doesn’t have scaler.

6

In pytorch you can do normalization, resizing etc with torchvision.transforms.

7

Yeah ,if don’t do normalization ,what will happen ?