fixing bugs
Browse files
utils.py
CHANGED
@@ -4,48 +4,49 @@ import jax.numpy as jnp
|
|
4 |
import numpy as np
|
5 |
|
6 |
|
7 |
-
def repeat_vmap(fun, in_axes=
|
8 |
-
if in_axes is None:
|
9 |
-
in_axes = [0]
|
10 |
for axes in in_axes:
|
11 |
fun = jax.vmap(fun, in_axes=axes)
|
12 |
return fun
|
13 |
|
14 |
|
15 |
def make_grid(patch_size: int | tuple[int, int]):
|
|
|
|
|
16 |
if isinstance(patch_size, int):
|
17 |
-
|
|
|
|
|
18 |
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
|
23 |
-
grid
|
24 |
-
|
|
|
25 |
|
26 |
|
27 |
def interpolate_grid(coords, grid, order=0):
|
28 |
-
"""
|
29 |
-
coords: Tensor de shape (B, H, W, 2) ou (H, W, 2)
|
30 |
-
grid: Tensor de shape (B, H', W', C)
|
31 |
-
order: default 0
|
32 |
-
"""
|
33 |
try:
|
34 |
-
# Converter para array
|
35 |
coords = jnp.asarray(coords)
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
if coords.shape[-1] != 2 or coords.ndim != 4:
|
41 |
-
raise ValueError(f"Formato inválido: {coords.shape}. Esperado (B, H, W, 2)")
|
42 |
|
43 |
# Transformação de coordenadas
|
44 |
coords = coords.transpose((0, 3, 1, 2))
|
45 |
-
coords = coords.at[:, 0].set(
|
46 |
-
|
47 |
-
|
48 |
-
|
|
|
|
|
|
|
|
|
49 |
map_fn = jax.vmap(jax.vmap(
|
50 |
partial(jax.scipy.ndimage.map_coordinates, order=order, mode='nearest'),
|
51 |
in_axes=(2, None),
|
@@ -54,4 +55,4 @@ def interpolate_grid(coords, grid, order=0):
|
|
54 |
return map_fn(grid, coords)
|
55 |
|
56 |
except Exception as e:
|
57 |
-
raise RuntimeError(f"
|
|
|
4 |
import numpy as np
|
5 |
|
6 |
|
7 |
+
def repeat_vmap(fun, in_axes=[0]):
|
|
|
|
|
8 |
for axes in in_axes:
|
9 |
fun = jax.vmap(fun, in_axes=axes)
|
10 |
return fun
|
11 |
|
12 |
|
13 |
def make_grid(patch_size: int | tuple[int, int]):
|
14 |
+
"""Gera grid de coordenadas com segurança numérica"""
|
15 |
+
# Garantir tamanho mínimo de 8x8
|
16 |
if isinstance(patch_size, int):
|
17 |
+
h = w = max(8, patch_size)
|
18 |
+
else:
|
19 |
+
h, w = (max(8, ps) for ps in patch_size)
|
20 |
|
21 |
+
# Espaçamento preciso entre pontos
|
22 |
+
y_space = np.linspace(-0.5 + 1 / (2 * h), 0.5 - 1 / (2 * h), h)
|
23 |
+
x_space = np.linspace(-0.5 + 1 / (2 * w), 0.5 - 1 / (2 * w), w)
|
24 |
|
25 |
+
# Criar grid com dimensões (1, H, W, 2)
|
26 |
+
grid = np.stack(np.meshgrid(y_space, x_space, indexing='ij'), axis=-1)
|
27 |
+
return grid[np.newaxis, ...]
|
28 |
|
29 |
|
30 |
def interpolate_grid(coords, grid, order=0):
|
31 |
+
"""Interpolação segura com verificação de dimensões"""
|
|
|
|
|
|
|
|
|
32 |
try:
|
33 |
+
# Converter para JAX array e validar formato
|
34 |
coords = jnp.asarray(coords)
|
35 |
+
if coords.ndim != 4 or coords.shape[-1] != 2:
|
36 |
+
raise ValueError(
|
37 |
+
f"Dimensões inválidas: {coords.shape}. Esperado (B, H, W, 2)"
|
38 |
+
)
|
|
|
|
|
39 |
|
40 |
# Transformação de coordenadas
|
41 |
coords = coords.transpose((0, 3, 1, 2))
|
42 |
+
coords = coords.at[:, 0].set(
|
43 |
+
coords[:, 0] * (grid.shape[-3] - 1) + (grid.shape[-3] - 1) / 2
|
44 |
+
)
|
45 |
+
coords = coords.at[:, 1].set(
|
46 |
+
coords[:, 1] * (grid.shape[-2] - 1) + (grid.shape[-2] - 1) / 2
|
47 |
+
)
|
48 |
+
|
49 |
+
# Interpolação vetorizada
|
50 |
map_fn = jax.vmap(jax.vmap(
|
51 |
partial(jax.scipy.ndimage.map_coordinates, order=order, mode='nearest'),
|
52 |
in_axes=(2, None),
|
|
|
55 |
return map_fn(grid, coords)
|
56 |
|
57 |
except Exception as e:
|
58 |
+
raise RuntimeError(f"Erro de interpolação: {str(e)}") from e
|