Color Space Transformation

chirathpansilu · March 7, 2020, 4:08pm

Guys, Is There a way to convert color spaces in fastai ?

Especially I want to convert RGB space to YUV space for a image classification project ?

Is there a direct way or do i have to make a custom transformer ?

nestorDemeure · March 7, 2020, 5:12pm

I am not aware of a fastai implementation but you will find differentiable pytorch version of those conversions in torchgeometry

chirathpansilu · March 7, 2020, 6:05pm

Oh it’s awesome , But Does Not contain the YUV space

Thanks for the suggestion

Did some digging and found this https://github.com/jorge-pessoa/pytorch-colors (Contains several spaces including YUV)

But I think it will take some work to make a DataBunch

muellerzr · March 7, 2020, 6:24pm

If you’re using v2 it shouldn’t be too bad at all actually. Look into PILImage (if you feel like jumping in)

JonathanSum · March 7, 2020, 6:50pm

I did something smilar before. I am not sure this is what you want to do.

![image|690x484](

chirathpansilu · March 7, 2020, 7:02pm

It Looks Interesting

But I’m Still On Lesson 3 , But I will Definitely give it a look when time comes

Thank You

chirathpansilu · March 7, 2020, 7:04pm

I’m Using v1,

But can You elaborate what you meant by that

Thank You

muellerzr · March 7, 2020, 7:10pm

v2 makes utilizing your own data very easy, and creating a custom image class you’re using is very very simplistic. (Also in July this will be merged into the fastai library so it’ll be the new go-to).

The PILImage is here:

github.com

fastai/fastai2/blob/master/fastai2/vision/core.py#L87


    return im.convert(mode) if mode else im


# Cell
def image2tensor(img):
    "Transform image to byte tensor in `c*h*w` dim order."
    res = tensor(img)
    if res.dim()==2: res = res.unsqueeze(-1)
    return res.permute(2,0,1)


# Cell
class PILBase(Image.Image, metaclass=BypassNewMeta):
    _bypass_type=Image.Image
    _show_args = {'cmap':'viridis'}
    _open_args = {'mode': 'RGB'}
    @classmethod
    def create(cls, fn:(Path,str,Tensor,ndarray,bytes), **kwargs)->None:
        "Open an `Image` from path `fn`"
        if isinstance(fn,TensorImage): fn = fn.permute(1,2,0).type(torch.uint8)
        if isinstance(fn,Tensor): fn = fn.numpy()
        if isinstance(fn,ndarray): return cls(Image.fromarray(fn))
        if isinstance(fn,bytes): fn = io.BytesIO(fn)

I’d recommend watching Jeremy’s code walkthroughs to understand the nitty gritty like that, otherwise I have a study group currently running where we explore it (and is a good way to learn v2 until the v4 of the course goes live in July)

chirathpansilu · March 8, 2020, 3:22am

That’s Interesting

Thank You