onnx--onnx/onnx/reference/ops/op_affine_grid.py

# Copyright (c) ONNX Project Contributors

# SPDX-License-Identifier: Apache-2.0
from __future__ import annotations

import numpy as np

from onnx.reference.op_run import OpRun


def construct_original_grid(data_size, align_corners):
    is_2d = len(data_size) == 2
    size_zeros = np.zeros(data_size)
    original_grid = [np.ones(data_size)]
    for dim, dim_size in enumerate(data_size):
        if align_corners == 1:
            step = 2.0 / (dim_size - 1)
            start = -1
            stop = 1 + 0.0001
            a = np.arange(start, stop, step)
        else:
            step = 2.0 / dim_size
            start = -1 + step / 2
            stop = 1
            a = np.arange(start, stop, step)
        if dim == 0:
            if is_2d:
                y = np.reshape(a, (dim_size, 1)) + size_zeros
                original_grid = [y, *original_grid]
            else:
                z = np.reshape(a, (dim_size, 1, 1)) + size_zeros
                original_grid = [z, *original_grid]
        elif dim == 1:
            if is_2d:
                x = np.reshape(a, (1, dim_size)) + size_zeros
                original_grid = [x, *original_grid]
            else:
                y = np.reshape(a, (1, dim_size, 1)) + size_zeros
                original_grid = [y, *original_grid]
        else:
            x = np.reshape(a, (1, dim_size)) + size_zeros
            original_grid = [x, *original_grid]
    return np.stack(original_grid, axis=2 if is_2d else 3)


def apply_affine_transform(theta_n, original_grid_homo):
    # theta_n: (N, 2, 3) for 2D, (N, 3, 4) for 3D
    # original_grid_homo: (H, W, 3) for 2D, (D, H, W, 4) for 3D
    assert theta_n.ndim == 3, (
        "theta_n shall have shape of (N, 2, 3) for 2D, (N, 3, 4) for 3D"
    )
    if original_grid_homo.ndim == 3:
        N, dim_2d, dim_homo = theta_n.shape
        assert dim_2d == 2 and dim_homo == 3
        H, W, dim_homo = original_grid_homo.shape
        assert dim_homo == 3
        # reshape to [H * W, dim_homo] and then transpose to [dim_homo, H * W]
        original_grid_transposed = np.transpose(
            np.reshape(original_grid_homo, (H * W, dim_homo))
        )
        grid_n = np.matmul(
            theta_n, original_grid_transposed
        )  # shape (N, dim_2d, H * W)
        # transpose to (N, H * W, dim_2d) and then reshape to (N, H, W, dim_2d)
        grid = np.reshape(np.transpose(grid_n, (0, 2, 1)), (N, H, W, dim_2d))
        return grid.astype(np.float32)
    assert original_grid_homo.ndim == 4
    N, dim_3d, dim_homo = theta_n.shape
    assert dim_3d == 3 and dim_homo == 4
    D, H, W, dim_homo = original_grid_homo.shape
    assert dim_homo == 4
    # reshape to [D * H * W, dim_homo] and then transpose to [dim_homo, D * H * W]
    original_grid_transposed = np.transpose(
        np.reshape(original_grid_homo, (D * H * W, dim_homo))
    )
    grid_n = np.matmul(
        theta_n, original_grid_transposed
    )  # shape (N, dim_3d, D * H * W)
    # transpose to (N, D * H * W, dim_3d) and then reshape to (N, D, H, W, dim_3d)
    grid = np.reshape(np.transpose(grid_n, (0, 2, 1)), (N, D, H, W, dim_3d))
    return grid.astype(np.float32)


class AffineGrid(OpRun):
    def _run(self, theta, size, align_corners=None):
        align_corners = align_corners or self.align_corners
        _, _, *data_size = size
        original_grid = construct_original_grid(data_size, align_corners)
        grid = apply_affine_transform(theta, original_grid)
        return (grid,)