Fix lazy rand affine (#6774)

Fixes #6773.

### Description

Call `rand_affine_grid()` once before call
`rand_affine_grid.get_transformation_matrix()`, since its documented as
"Get the most recently applied transformation matrix", or the `.affine`
attribute will not be set.
Also, set `randomize=False` here since randomization if performed in the
beginning of the function.

### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.

---------

Signed-off-by: function2 <[email protected]>
Co-authored-by: Wenqi Li <[email protected]>

Author: function2-llx

monai/transforms/spatial/array.py

@@ -2523,6 +2523,8 @@ def __call__(
         img = convert_to_tensor(img, track_meta=get_track_meta())
         if lazy_:
             if self._do_transform:
+                if grid is None:
+                    self.rand_affine_grid(sp_size, randomize=randomize, lazy=True)
                 affine = self.rand_affine_grid.get_transformation_matrix()
             else:
                 affine = convert_to_dst_type(torch.eye(len(sp_size) + 1), img, dtype=self.rand_affine_grid.dtype)[0]

monai/transforms/spatial/dictionary.py

@@ -1141,6 +1141,7 @@ def __call__(
             grid = self.rand_affine.get_identity_grid(sp_size, lazy=lazy_)
             if self._do_transform:  # add some random factors
                 grid = self.rand_affine.rand_affine_grid(sp_size, grid=grid, lazy=lazy_)
+        grid = 0 if grid is None else grid  # always provide a grid to self.rand_affine
 
         for key, mode, padding_mode in self.key_iterator(d, self.mode, self.padding_mode):
             # do the transform

tests/test_rand_affine.py

@@ -145,6 +145,7 @@ def test_rand_affine(self, input_param, input_data, expected_val):
         g = RandAffine(**input_param)
         g.set_random_state(123)
         result = g(**input_data)
+        g.rand_affine_grid.affine = torch.eye(4, dtype=torch.float64)  # reset affine
         test_resampler_lazy(g, result, input_param, input_data, seed=123)
         if input_param.get("cache_grid", False):
             self.assertTrue(g._cached_grid is not None)