apextoaster
/
onnx-web
1
0
Fork 0
Code Actions

use adjusted tile for gradients, repeat mask based on output shape

This commit is contained in:
Sean Sube 2023-06-03 13:17:20 -05:00
parent ec4ee49806
commit 51c5b63090
Signed by: ssube
GPG Key ID: 3EED7B957D362AF1
1 changed files with 13 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
api/onnx_web/chain/utils.py
View File

@ -42,20 +42,20 @@ def get_tile_grads(
width: int, width: int,
height: int, height: int,
) -> Tuple[Tuple[float, float, float, float], Tuple[float, float, float, float]]: ) -> Tuple[Tuple[float, float, float, float], Tuple[float, float, float, float]]:
grad_x = [1, 1, 1, 1] grad_x = [0, 1, 1, 0]
grad_y = [1, 1, 1, 1] grad_y = [0, 1, 1, 0]
if left > 0: if left <= 0:
grad_x[0] = 0 grad_x[0] = 1
if top > 0: if top <= 0:
grad_y[0] = 0 grad_y[0] = 1
if (left + tile) >= width: if (left + tile) >= width:
grad_x[3] = 0 grad_x[3] = 1
if (top + tile) >= height: if (top + tile) >= height:
grad_y[3] = 0 grad_y[3] = 1
return (grad_x, grad_y) return (grad_x, grad_y)
@ -104,22 +104,22 @@ def process_tile_grid(
# gradient blending # gradient blending
points = [0, adj_tile * scale, (tile - adj_tile) * scale, (tile * scale) - 1] points = [0, adj_tile * scale, (tile - adj_tile) * scale, (tile * scale) - 1]
grad_x, grad_y = get_tile_grads(left, top, tile, width, height) grad_x, grad_y = get_tile_grads(left, top, adj_tile, width, height)
mult_x = [np.interp(i, points, grad_x) for i in range(tile * scale)] mult_x = [np.interp(i, points, grad_x) for i in range(tile * scale)]
mult_y = [np.interp(i, points, grad_y) for i in range(tile * scale)] mult_y = [np.interp(i, points, grad_y) for i in range(tile * scale)]
mask = np.ones_like(equalized[:, :, 0]) * mult_x mask = np.ones_like(equalized[:, :, 0]) * mult_x
mask = (mask.T * mult_y).T mask = (mask.T * mult_y).T
for c in range(3): for c in range(3):
equalized[:, :, c] *= mask.astype(np.uint8) equalized[:, :, c] = (equalized[:, :, c] * mask).astype(np.uint8)
# accumulation # accumulation
value[ value[
top * scale : (top + tile) * scale, left * scale : (left + tile) * scale, : top * scale : (top * scale) + equalized.shape[0], left * scale : (left * scale) + equalized.shape[1], :
] += equalized ] += equalized
count[ count[
top * scale : (top + tile) * scale, left * scale : (left + tile) * scale, : top * scale : (top * scale) + equalized.shape[0], left * scale : (left * scale) + equalized.shape[1], :
] += mask ] += np.repeat(mask, 3, axis=2)
pixels = np.where(count > 0, value / count, value) pixels = np.where(count > 0, value / count, value)
return Image.fromarray(pixels) return Image.fromarray(pixels)