Spaces:
Running
Running
# Copyright (c) ONNX Project Contributors | |
# | |
# SPDX-License-Identifier: Apache-2.0 | |
import numpy as np | |
import onnx | |
from onnx.backend.test.case.base import Base | |
from onnx.backend.test.case.node import expect | |
from onnx.reference.ops.op_affine_grid import ( | |
apply_affine_transform, | |
construct_original_grid, | |
) | |
def create_affine_matrix_3d( | |
angle1, | |
angle2, | |
offset_x, | |
offset_y, | |
offset_z, | |
shear_x, | |
shear_y, | |
shear_z, | |
scale_x, | |
scale_y, | |
scale_z, | |
): | |
rot_x = np.stack( | |
[ | |
np.ones_like(angle1), | |
np.zeros_like(angle1), | |
np.zeros_like(angle1), | |
np.zeros_like(angle1), | |
np.cos(angle1), | |
-np.sin(angle1), | |
np.zeros_like(angle1), | |
np.sin(angle1), | |
np.cos(angle1), | |
], | |
axis=-1, | |
).reshape(-1, 3, 3) | |
rot_y = np.stack( | |
[ | |
np.cos(angle2), | |
np.zeros_like(angle2), | |
np.sin(angle2), | |
np.zeros_like(angle2), | |
np.ones_like(angle2), | |
np.zeros_like(angle2), | |
-np.sin(angle2), | |
np.zeros_like(angle2), | |
np.cos(angle2), | |
], | |
axis=-1, | |
).reshape(-1, 3, 3) | |
shear = np.stack( | |
[ | |
np.ones_like(shear_x), | |
shear_x, | |
shear_y, | |
shear_z, | |
np.ones_like(shear_x), | |
shear_x, | |
shear_y, | |
shear_x, | |
np.ones_like(shear_x), | |
], | |
axis=-1, | |
).reshape(-1, 3, 3) | |
scale = np.stack( | |
[ | |
scale_x, | |
np.zeros_like(scale_x), | |
np.zeros_like(scale_x), | |
np.zeros_like(scale_x), | |
scale_y, | |
np.zeros_like(scale_x), | |
np.zeros_like(scale_x), | |
np.zeros_like(scale_x), | |
scale_z, | |
], | |
axis=-1, | |
).reshape(-1, 3, 3) | |
translation = np.transpose(np.array([offset_x, offset_y, offset_z])).reshape( | |
-1, 1, 3 | |
) | |
rotation_matrix = rot_y @ rot_x @ shear @ scale # (N, 3, 3) | |
rotation_matrix = np.transpose(rotation_matrix, (0, 2, 1)) | |
affine_matrix = np.hstack((rotation_matrix, translation)) | |
affine_matrix = np.transpose(affine_matrix, (0, 2, 1)) | |
return affine_matrix.astype(np.float32) | |
def create_affine_matrix_2d( | |
angle1, offset_x, offset_y, shear_x, shear_y, scale_x, scale_y | |
): | |
rot = np.stack( | |
[np.cos(angle1), -np.sin(angle1), np.sin(angle1), np.cos(angle1)], axis=-1 | |
).reshape(-1, 2, 2) | |
shear = np.stack( | |
[np.ones_like(shear_x), shear_x, shear_y, np.ones_like(shear_x)], axis=-1 | |
).reshape(-1, 2, 2) | |
scale = np.stack( | |
[scale_x, np.zeros_like(scale_x), np.zeros_like(scale_x), scale_y], axis=-1 | |
).reshape(-1, 2, 2) | |
translation = np.transpose(np.array([offset_x, offset_y])).reshape(-1, 1, 2) | |
rotation_matrix = rot @ shear @ scale # (N, 3, 3) | |
rotation_matrix = np.transpose(rotation_matrix, (0, 2, 1)) | |
affine_matrix = np.hstack((rotation_matrix, translation)) | |
affine_matrix = np.transpose(affine_matrix, (0, 2, 1)) | |
return affine_matrix.astype(np.float32) | |
def create_theta_2d(): | |
angle = np.array([np.pi / 4, np.pi / 3]) | |
offset_x = np.array([5.0, 2.5]) | |
offset_y = np.array([-3.3, 1.1]) | |
shear_x = np.array([-0.5, 0.5]) | |
shear_y = np.array([0.3, -0.3]) | |
scale_x = np.array([2.2, 1.1]) | |
scale_y = np.array([3.1, 0.9]) | |
theta_2d = create_affine_matrix_2d( | |
angle, offset_x, offset_y, shear_x, shear_y, scale_x, scale_y | |
) | |
return theta_2d | |
def create_theta_3d(): | |
angle1 = np.array([np.pi / 4, np.pi / 3]) | |
angle2 = np.array([np.pi / 6, np.pi / 2]) | |
offset_x = np.array([5.0, 2.5]) | |
offset_y = np.array([-3.3, 1.1]) | |
offset_z = np.array([-1.1, 2.2]) | |
shear_x = np.array([-0.5, 0.5]) | |
shear_y = np.array([0.3, -0.3]) | |
shear_z = np.array([0.7, -0.2]) | |
scale_x = np.array([2.2, 1.1]) | |
scale_y = np.array([3.1, 0.9]) | |
scale_z = np.array([0.5, 1.5]) | |
theta_3d = create_affine_matrix_3d( | |
angle1, | |
angle2, | |
offset_x, | |
offset_y, | |
offset_z, | |
shear_x, | |
shear_y, | |
shear_z, | |
scale_x, | |
scale_y, | |
scale_z, | |
) | |
return theta_3d | |
class AffineGrid(Base): | |
def export_2d_no_reference_evaluator() -> None: | |
theta_2d = create_theta_2d() | |
N, C, H, W = len(theta_2d), 3, 5, 6 | |
data_size = (H, W) | |
for align_corners in (0, 1): | |
node = onnx.helper.make_node( | |
"AffineGrid", | |
inputs=["theta", "size"], | |
outputs=["grid"], | |
align_corners=align_corners, | |
) | |
original_grid = construct_original_grid(data_size, align_corners) | |
grid = apply_affine_transform(theta_2d, original_grid) | |
test_name = "test_affine_grid_2d" | |
if align_corners == 1: | |
test_name += "_align_corners" | |
expect( | |
node, | |
inputs=[theta_2d, np.array([N, C, H, W], dtype=np.int64)], | |
outputs=[grid], | |
name=test_name, | |
) | |
def export_3d_no_reference_evaluator() -> None: | |
theta_3d = create_theta_3d() | |
N, C, D, H, W = len(theta_3d), 3, 4, 5, 6 | |
data_size = (D, H, W) | |
for align_corners in (0, 1): | |
node = onnx.helper.make_node( | |
"AffineGrid", | |
inputs=["theta", "size"], | |
outputs=["grid"], | |
align_corners=align_corners, | |
) | |
original_grid = construct_original_grid(data_size, align_corners) | |
grid = apply_affine_transform(theta_3d, original_grid) | |
test_name = "test_affine_grid_3d" | |
if align_corners == 1: | |
test_name += "_align_corners" | |
expect( | |
node, | |
inputs=[theta_3d, np.array([N, C, D, H, W], dtype=np.int64)], | |
outputs=[grid], | |
name=test_name, | |
) | |