affine transformation is not working as expected (monai0.4.0)

[YufanHe]

The parameters on the affine transformation is not corresponding to the generated image

image = np.zeros((128,128))
image[30:80, 30:80] = 1 
'''The affine transformations are applied in rotate, shear, translate, scale order.'''
fixed = Affine(rotate_params=0, translate_params=[0,0], scale_params=[0.5,0.5], shear_params=[0,0], \
                mode='bilinear', padding_mode='zeros')(image[np.newaxis], spatial_size=image.shape)[0]
plt.imshow(image);plt.show();
plt.imshow(fixed);plt.show();

The scale is 0.5 but the generated image is larger. The problem comes from not inversing the affine matrix for the affine_grid
as in
https://kornia.readthedocs.io/en/latest/_modules/kornia/geometry/transform/imgwarp.html#warp_affine

Also the documentation says 'The affine transformations are applied in rotate, shear, translate, scale order but this is reversed.
The source code from the AffineGrid in monai:

from monai.transforms.utils import (
    create_control_grid,
    create_grid,
    create_rotate,
    create_scale,
    create_shear,
    create_translate,
)
def get_affine_monai(rotate_params, shear_params, translate_params, scale_params, spatial_dims=2):
    affine = np.eye(spatial_dims + 1)
    affine = affine @ create_rotate(spatial_dims, rotate_params)
    affine = affine @ create_shear(spatial_dims, shear_params)
    affine = affine @ create_translate(spatial_dims, translate_params)
    affine = affine @ create_scale(spatial_dims, scale_params)
    return affine 
print(get_affine_monai(0, [0,0],[40,40],2))

This affine matrix shows that the scaling is not affecting the translation, and from the code we can see that the scaling is before translation (Affine = rotate * shear * translate * scale) * [x, y ,1]

[wyli]

I think that is just a difference in the conventions, here we try to follow scipy's affine_transform convention

[wyli]

thi is the current expected behaviour, discussed/confirmed by Dong offline