updates to Gaussian map for sliding window inference (#5302)

This updates the calculation of Gaussian map (weights) during sliding
window inference with "gaussian"

Current version had multiple small issues
- it computed Gaussian weight map (image) via nn.Conv1d sequence with an
empty image (with a single 1 in the middle). We don't need to run any
convolutions, it's much simpler to directly calculate the Gaussian map
(it's also faster and takes less memory)
- For patch_sizes of even size (e.g. 128x128) it centered Gaussian on
patch_size//2 which is 0.5 pixel off-center (I'm not sure why we did it.
- Finally the Gaussian 1d convolutions were done approximately (with
'erf' internal approximation and truncated to sigma=4). I'm not sure why
we need any approximations here at all, it's trivial to compute the
Gaussian weight map directly



### 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).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [ ] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

Signed-off-by: myron <[email protected]>

Author: myron

monai/data/utils.py

@@ -30,7 +30,6 @@
 from monai import config
 from monai.config.type_definitions import NdarrayOrTensor, NdarrayTensor, PathLike
 from monai.data.meta_obj import MetaObj
-from monai.networks.layers.simplelayers import GaussianFilter
 from monai.utils import (
     MAX_SEED,
     BlendMode,
@@ -1067,17 +1066,16 @@ def compute_importance_map(
     if mode == BlendMode.CONSTANT:
         importance_map = torch.ones(patch_size, device=device, dtype=torch.float)
     elif mode == BlendMode.GAUSSIAN:
-        center_coords = [i // 2 for i in patch_size]
+
         sigma_scale = ensure_tuple_rep(sigma_scale, len(patch_size))
         sigmas = [i * sigma_s for i, sigma_s in zip(patch_size, sigma_scale)]
 
-        importance_map = torch.zeros(patch_size, device=device)
-        importance_map[tuple(center_coords)] = 1
-        pt_gaussian = GaussianFilter(len(patch_size), sigmas).to(device=device, dtype=torch.float)
-        importance_map = pt_gaussian(importance_map.unsqueeze(0).unsqueeze(0))
-        importance_map = importance_map.squeeze(0).squeeze(0)
-        importance_map = importance_map / torch.max(importance_map)
-        importance_map = importance_map.float()
+        for i in range(len(patch_size)):
+            x = torch.arange(
+                start=-(patch_size[i] - 1) / 2.0, end=(patch_size[i] - 1) / 2.0 + 1, dtype=torch.float, device=device
+            )
+            x = torch.exp(x**2 / (-2 * sigmas[i] ** 2))  # 1D gaussian
+            importance_map = importance_map.unsqueeze(-1) * x[(None,) * i] if i > 0 else x
     else:
         raise ValueError(
             f"Unsupported mode: {mode}, available options are [{BlendMode.CONSTANT}, {BlendMode.CONSTANT}]."