Merge pull request #13 from iesl/regularizers

Added delta regularizer for now. More can be added later.

box_embeddings/modules/regularization/__init__.py

@@ -0,0 +1,2 @@
+from .regularizer import BoxRegularizer
+from .l2_side_regularizer import l2_side_regularizer, L2SideBoxRegularizer

box_embeddings/modules/regularization/l2_side_regularizer.py

@@ -0,0 +1,58 @@
+from typing import List, Tuple, Union, Dict, Any, Optional
+from box_embeddings.common.registrable import Registrable
+import torch
+from box_embeddings.parameterizations.box_tensor import BoxTensor
+from box_embeddings.modules.regularization.regularizer import BoxRegularizer
+
+eps = 1e-23
+
+
+def l2_side_regularizer(
+    box_tensor: BoxTensor, log_scale: bool = False
+) -> Union[float, torch.Tensor]:
+    """Applies l2 regularization on all sides of all boxes and returns the sum.
+
+    Args:
+        box_tensor: TODO
+        log_scale: mean in log scale
+
+    Returns:
+        (None)
+    """
+    z = box_tensor.z  # (..., box_dim)
+    Z = box_tensor.Z  # (..., box_dim)
+
+    if not log_scale:
+        return torch.mean((Z - z) ** 2)
+    else:
+        return torch.mean(torch.log(torch.abs(Z - z) + eps))
+
+
+@BoxRegularizer.register("l2_side")
+class L2SideBoxRegularizer(BoxRegularizer):
+
+    """Applies l2 regularization on side lengths."""
+
+    def __init__(self, weight: float, log_scale: bool = False) -> None:
+        """TODO: to be defined.
+
+        Args:
+            weight: Weight (hyperparameter) given to this regularization in the overall loss.
+            log_scale: Whether the output should be in log scale or not.
+                Should be true in almost any practical case where box_dim>5.
+
+
+        """
+        super().__init__(weight, log_scale=log_scale)
+
+    def _forward(self, box_tensor: BoxTensor) -> Union[float, torch.Tensor]:
+        """Applies l2 regularization on all sides of all boxes and returns the sum.
+
+        Args:
+            box_tensor: TODO
+
+        Returns:
+            (None)
+        """
+
+        return l2_side_regularizer(box_tensor, log_scale=self.log_scale)

box_embeddings/modules/regularization/regularizer.py

@@ -0,0 +1,47 @@
+from typing import List, Tuple, Union, Dict, Any, Optional
+from box_embeddings.common.registrable import Registrable
+import torch
+from box_embeddings.parameterizations.box_tensor import BoxTensor
+
+
+class BoxRegularizer(torch.nn.Module, Registrable):
+
+    """Base box-regularizer class"""
+
+    def __init__(
+        self, weight: float, log_scale: bool = True, **kwargs: Any
+    ) -> None:
+        """
+        Args:
+            weight: Weight (hyperparameter) given to this regularization in the overall loss.
+            log_scale: Whether the output should be in log scale or not.
+                Should be true in almost any practical case where box_dim>5.
+            kwargs: Unused
+        """
+        super().__init__()  # type:ignore
+        self.weight = weight
+        self.log_scale = log_scale
+
+    def forward(self, box_tensor: BoxTensor) -> Union[float, torch.Tensor]:
+        """Calls the _forward and multiplies the weight
+
+        Args:
+            box_tensor: Input box tensor
+
+        Returns:
+            scalar regularization loss
+        """
+
+        return self.weight * self._forward(box_tensor)
+
+    def _forward(self, box_tensor: BoxTensor) -> Union[float, torch.Tensor]:
+        """The method that does all the work and needs to be overriden
+
+        Args:
+            box_tensor: Input box tensor
+
+        Returns:
+            0
+        """
+
+        return 0.0

core_requirements.txt

@@ -1,2 +1,2 @@
-torch >= 1.7.0
+torch >= 1.6.0
 numpy