🏁 Image grid

lo(number_images)

array[100, 64, 64, 3] n=1228800 x∈[0., 1.000] μ=0.784 σ=0.171

# number_images[10].rgb(cl=True)

t = number_images

n_images = t.shape[0]
n_channels = t.shape[-1]
xy_shape = t.shape[1:3]

n_rows, n_cols = fit_columns(t, view_width=966)

# We need to form the images inro a rectangular area. For this, we might
# need to add some dummy images to the last row, whoch might be not be full.
n_extra_images = n_rows*n_cols - t.shape[0]
if n_extra_images:
    extra_images = np.ones((n_extra_images, *t.shape[1:]))
    # extra_images = torch.ones((n_extra_images, *t.shape[1:]))
    t = np.concatenate([ t, extra_images ])

lo(t)

array[104, 64, 64, 3] n=1277952 x∈[0., 1.000] μ=0.792 σ=0.173

# This is where the fun begins! Imagine 't' is tensor[20, 128, 128, 3].
# and we want 5 rows, 4 columns each.

t = t.reshape(n_rows, n_cols, *t.shape[-3:])
# Now t is tensor[5, 4, 128, 128, 3]

t = t.transpose(0, 2, 1, 3, 4)
# now t is tensor[5, 128, 4, 128, 3]
# If we just squick dimensions 0,1 and 2,3 togerther, we get the image we want.
t = t.reshape(n_rows*xy_shape[0], n_cols*xy_shape[1], n_channels)

lo(t).rgb(cl=1)

---

source

tile2d

 tile2d (t:numpy.ndarray, view_width=966)

Tile images in a grid.

	Type	Default	Details
t	ndarray		Array containing images, shape=(n,h,w,c)
view_width	int	966	Try to protuce an images at most this wide

lo(tile2d(number_images[:17]))

array[192, 512, 3] n=294912 x∈[0., 1.000] μ=0.852 σ=0.160

lo(tile2d(number_images[:16])).rgb

Combine with utils.pad.pad_frame_gutters to make it look better

lo(tile2d(pad_frame_gutters(number_images[:16]), view_width=1200)).rgb

---

source

hypertile

 hypertile (t:numpy.ndarray, frame_px=1, gutter_px=3, view_width=966)

Recursively tile images on a 2d grid

	Type	Default	Details
t	ndarray		torch.Tensor, # input tensor, shape=([…], B, H, W, C)
frame_px	int	1	Frame width for the innermost group
gutter_px	int	3	Gutter width for the innermost group
view_width	int	966	Try to protuce an images at most this wide

hyperimages = number_images.reshape(2, 5, 10, 64, 64, 3)
lo(hyperimages)

array[2, 5, 10, 64, 64, 3] n=1228800 x∈[0., 1.000] μ=0.784 σ=0.171

lo(hypertile(hyperimages)).rgb