Fix typos in comments

src/sharp/models/initializer.py

@@ -114,7 +114,7 @@ def __init__(
         self.feature_input_stop_grad = feature_input_stop_grad
 
     def prepare_feature_input(self, image: torch.Tensor, depth: torch.Tensor) -> torch.Tensor:
-        """Prepare the feature input to the Guassian predictor."""
+        """Prepare the feature input to the Gaussian predictor."""
         if self.feature_input_stop_grad:
             image = image.detach()
             depth = depth.detach()

src/sharp/models/monodepth.py

@@ -77,7 +77,7 @@ def __init__(
             nn.ReLU(),
         )
 
-        # Set the final convoultion layer's bias to be 0.
+        # Set the final convolution layer's bias to be 0.
         self.head[4].bias.data.fill_(0)
 
         self.grad_checkpointing = False

src/sharp/models/predictor.py

@@ -140,7 +140,7 @@ def forward(
         #         +------+-------+
         #                |
         #        +-------+--------+     # Optionally align monodepth to ground truth
-        #        |depth_alignement|     # with a local scale map.
+        #        |depth_alignment|     # with a local scale map.
         #        +-------+--------+
         #                |
         #                v
@@ -169,7 +169,7 @@ def forward(
         #
 
         # The logic to decide whether to align monodepth to the ground truth is wrapped
-        # in a submodule 'DepthAlignement' to facilitate the symbolic tracing of the
+        # in a submodule 'DepthAlignment' to facilitate the symbolic tracing of the
         # predictor. This way, the depth alignment submodule containing the conditional
         # logic can be excluded during the tracing and the graph of the predictors is
         # static.