πŸ–ŒοΈ View as RGB images

from nbdev.showdoc import *
from fastcore.test import test_eq
from lovely_grad import monkey_patch

source

rgb

 rgb (x:tinygrad.tensor.Tensor, denorm:Any=None, cl:Any=False,
      gutter_px:int=3, frame_px:int=1, scale:int=1, view_width:int=966,
      ax:Optional[matplotlib.axes._axes.Axes]=None)
Type Default Details
x Tensor Tensor to display. [[…], C,H,W] or [[…], H,W,C]
denorm Any None Reverse per-channel normalizatoin
cl Any False Channel-last
gutter_px int 3 If more than one tensor -> tile with this gutter width
frame_px int 1 If more than one tensor -> tile with this frame width
scale int 1 Scale up. Can’t scale down.
view_width int 966 target width of the image
ax Optional None Use this Axes
Returns RGBProxy 
rgb(image)

rgb(image, scale=2)

two_images = Tensor.stack(image, image)
two_images
Tensor[2, 3, 196, 196] n=230496 x∈[-2.118, 2.640] ΞΌ=-0.388 Οƒ=1.073 CPU Realized ADD
in_stats = (    [0.485, 0.456, 0.406],  # Mean
                [0.229, 0.224, 0.225] ) # std
rgb(two_images, denorm=in_stats)

# Make 8 images with progressively higher brightness and stack them 2x2x2.
eight_images = (Tensor.stack(*([image]*8)) + Tensor(np.linspace(-2, 2, 8).astype(np.float32)).reshape(8,1,1,1))
eight_images = (eight_images
                    .mul(Tensor(in_stats[1]).reshape(1,3,1,1))
                    .add(Tensor(in_stats[0]).reshape(1,3,1,1))
                    .clip(0,1)
                    .reshape(2,2,2,3,196,196)
)
eight_images
Tensor[2, 2, 2, 3, 196, 196] n=921984 x∈[0., 1.000] ΞΌ=0.382 Οƒ=0.319 CPU Realized RESHAPE
rgb(eight_images)

# You can do channel-last too:
rgb(image.permute(1, 2, 0), cl=True)