Type of transforms.Normalize()?

vision
1

after creating normalize transform like this

    normalize = transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
    print(type(normalize))

I get type

  • <class ‘torchvision.transforms.transforms.Normalize’>
    Is it ok or not? Asking because on another machine I was getting
  • <class ‘torchvision.transforms.Normalize’>
    and my code was running properly, now it isn’t…
2

The class is fine.
What kind of error do you get?
Your code works in this simple example:

normalize = transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
)

x = torch.randn(3, 24, 24)
y = normalize(x)
3

A timely response from me! :slight_smile:
Anyways, I am not sure if I should get

<class ‘torchvision.transforms.transforms.Normalize’>

or

<class ‘torchvision.transforms.Normalize’>

Asking because my code is not running properly on the new machine, but it was fine earlier.

The Error I get is

TypeError: pic should be Tensor or ndarray. Got <class ‘torch.DoubleTensor’>.

I can provide the full code if need be.

4

If I print the class, I get:

print(normalize.__class__)
> <class 'torchvision.transforms.transforms.Normalize'>

As far as I remember this error message might indicate you are passing an image tensor containing a batch dimension.
Could you check that and if it’s the case, pass the tensor as [channels, height, width]?

6

Hi. May I ask, how to define the mean value and std value for each image channel?
Moreover, can we set a parameter to make the CNN find the optimal parameter (mean value, std value or other weights/biases used in each channel) for the image processing? If so, can you tell me how to set the parameter?

7

You could calculate the mean and stddev from your training dataset. Alternatively, you could use the ImageNet stats if your dataset is from a similar domain (natural images).
Often just 0.5 works fine so you would have to try it out.

I’m not sure if it’s a good idea to learn these preprocessing steps as they often are a good way to speed up training or make it feasible in the first place.

8

ok. I got your idea. Thanks a lot.