Merge branch 'master' into kernel_distances

prototorch/functions/distances.py

@@ -3,8 +3,11 @@
 import numpy as np
 import torch
 
-from prototorch.functions.helper import (_check_shapes, _int_and_mixed_shape,
-                                         equal_int_shape)
+from prototorch.functions.helper import (
+    _check_shapes,
+    _int_and_mixed_shape,
+    equal_int_shape,
+)
 
 
 def squared_euclidean_distance(x, y):
@@ -261,5 +264,86 @@ def tangent_distance(signals, protos, subspaces, squared=False, epsilon=1e-10):
             return diss.permute([1, 0, 2]).squeeze(-1)
 
 
+class KernelDistance:
+    r"""Kernel Distance
+
+    Distance based on a kernel function.
+    """
+    def __init__(self, kernel_fn):
+        self.kernel_fn = kernel_fn
+
+    def __call__(self, x_batch: torch.Tensor, y_batch: torch.Tensor):
+        return self._single_call(x_batch, y_batch)
+
+    def _single_call(self, x, y):
+        remove_dims = []
+        if len(x.shape) == 1:
+            x = x.unsqueeze(0)
+            remove_dims.append(0)
+        if len(y.shape) == 1:
+            y = y.unsqueeze(0)
+            remove_dims.append(-1)
+
+        output = self.kernel_fn(x, x).diag().unsqueeze(1) - 2 * self.kernel_fn(
+            x, y) + self.kernel_fn(y, y).diag()
+
+        for dim in remove_dims:
+            output.squeeze_(dim)
+
+        return torch.sqrt(output)
+
+
+class BatchKernelDistance:
+    r"""Kernel Distance
+
+    Distance based on a kernel function.
+    """
+    def __init__(self, kernel_fn):
+        self.kernel_fn = kernel_fn
+
+    def __call__(self, x_batch: torch.Tensor, y_batch: torch.Tensor):
+        remove_dims = 0
+        # Extend Single inputs
+        if len(x_batch.shape) == 1:
+            x_batch = x_batch.unsqueeze(0)
+            remove_dims += 1
+        if len(y_batch.shape) == 1:
+            y_batch = y_batch.unsqueeze(0)
+            remove_dims += 1
+
+        # Loop over batches
+        output = torch.FloatTensor(len(x_batch), len(y_batch))
+        for i, x in enumerate(x_batch):
+            for j, y in enumerate(y_batch):
+                output[i][j] = self._single_call(x, y)
+
+        for _ in range(remove_dims):
+            output.squeeze_(0)
+
+        return output
+
+    def _single_call(self, x, y):
+        kappa_xx = self.kernel_fn(x, x)
+        kappa_xy = self.kernel_fn(x, y)
+        kappa_yy = self.kernel_fn(y, y)
+
+        squared_distance = kappa_xx - 2 * kappa_xy + kappa_yy
+
+        return torch.sqrt(squared_distance)
+
+
+class SquaredKernelDistance(KernelDistance):
+    r"""Squared Kernel Distance
+
+    Kernel distance without final squareroot.
+    """
+    def single_call(self, x, y):
+        kappa_xx = self.kernel_fn(x, x)
+        kappa_xy = self.kernel_fn(x, y)
+        kappa_yy = self.kernel_fn(y, y)
+
+        return kappa_xx - 2 * kappa_xy + kappa_yy
+
+
 # Aliases
 sed = squared_euclidean_distance

prototorch/functions/kernels.py

@@ -0,0 +1,28 @@
+"""
+Experimental Kernels
+"""
+
+import torch
+
+
+class ExplicitKernel:
+    def __init__(self, projection=torch.nn.Identity()):
+        self.projection = projection
+
+    def __call__(self, x, y):
+        return self.projection(x) @ self.projection(y).T
+
+
+class RadialBasisFunctionKernel:
+    def __init__(self, sigma) -> None:
+        self.s2 = sigma * sigma
+
+    def __call__(self, x, y):
+        remove_dim = False
+        if len(x.shape) > 1:
+            x = x.unsqueeze(1)
+            remove_dim = True
+        output = torch.exp(-torch.sum((x - y)**2, dim=-1) / (2 * self.s2))
+        if remove_dim:
+            output = output.squeeze(1)
+        return output

prototorch/modules/models.py

@@ -1,8 +1,7 @@
 import torch
 from torch import nn
 
-from prototorch.functions.distances import (euclidean_distance_matrix,
-                                            tangent_distance)
+from prototorch.functions.distances import euclidean_distance_matrix, tangent_distance
 from prototorch.functions.helper import _check_shapes, _int_and_mixed_shape
 from prototorch.functions.normalization import orthogonalization
 from prototorch.modules.prototypes import Prototypes1D

tests/test_functions.py

@@ -5,8 +5,13 @@ import unittest
 import numpy as np
 import torch
 
-from prototorch.functions import (activations, competitions, distances,
-                                  initializers, losses)
+from prototorch.functions import (
+    activations,
+    competitions,
+    distances,
+    initializers,
+    losses,
+)
 
 
 class TestActivations(unittest.TestCase):

tests/test_kernels.py

@@ -0,0 +1,98 @@
+"""ProtoTorch kernels test suite."""
+
+import unittest
+
+import numpy as np
+import torch
+
+from prototorch.functions.distances import KernelDistance
+from prototorch.functions.kernels import ExplicitKernel, RadialBasisFunctionKernel
+
+
+class TestExplicitKernel(unittest.TestCase):
+    def setUp(self):
+        self.single_x = torch.randn(1024)
+        self.single_y = torch.randn(1024)
+
+        self.batch_x = torch.randn(32, 1024)
+        self.batch_y = torch.randn(32, 1024)
+
+    def test_single_values(self):
+        kernel = ExplicitKernel()
+        self.assertEqual(
+            kernel(self.single_x, self.single_y).shape, torch.Size([]))
+
+    def test_single_batch(self):
+        kernel = ExplicitKernel()
+        self.assertEqual(
+            kernel(self.single_x, self.batch_y).shape, torch.Size([32]))
+
+    def test_batch_single(self):
+        kernel = ExplicitKernel()
+        self.assertEqual(
+            kernel(self.batch_x, self.single_y).shape, torch.Size([32]))
+
+    def test_batch_values(self):
+        kernel = ExplicitKernel()
+        self.assertEqual(
+            kernel(self.batch_x, self.batch_y).shape, torch.Size([32, 32]))
+
+
+class TestRadialBasisFunctionKernel(unittest.TestCase):
+    def setUp(self):
+        self.single_x = torch.randn(1024)
+        self.single_y = torch.randn(1024)
+
+        self.batch_x = torch.randn(32, 1024)
+        self.batch_y = torch.randn(32, 1024)
+
+    def test_single_values(self):
+        kernel = RadialBasisFunctionKernel(1)
+        self.assertEqual(
+            kernel(self.single_x, self.single_y).shape, torch.Size([]))
+
+    def test_single_batch(self):
+        kernel = RadialBasisFunctionKernel(1)
+        self.assertEqual(
+            kernel(self.single_x, self.batch_y).shape, torch.Size([32]))
+
+    def test_batch_single(self):
+        kernel = RadialBasisFunctionKernel(1)
+        self.assertEqual(
+            kernel(self.batch_x, self.single_y).shape, torch.Size([32]))
+
+    def test_batch_values(self):
+        kernel = RadialBasisFunctionKernel(1)
+        self.assertEqual(
+            kernel(self.batch_x, self.batch_y).shape, torch.Size([32, 32]))
+
+
+class TestKernelDistance(unittest.TestCase):
+    def setUp(self):
+        self.single_x = torch.randn(1024)
+        self.single_y = torch.randn(1024)
+
+        self.batch_x = torch.randn(32, 1024)
+        self.batch_y = torch.randn(32, 1024)
+
+        self.kernel = ExplicitKernel()
+
+    def test_single_values(self):
+        distance = KernelDistance(self.kernel)
+        self.assertEqual(
+            distance(self.single_x, self.single_y).shape, torch.Size([]))
+
+    def test_single_batch(self):
+        distance = KernelDistance(self.kernel)
+        self.assertEqual(
+            distance(self.single_x, self.batch_y).shape, torch.Size([32]))
+
+    def test_batch_single(self):
+        distance = KernelDistance(self.kernel)
+        self.assertEqual(
+            distance(self.batch_x, self.single_y).shape, torch.Size([32]))
+
+    def test_batch_values(self):
+        distance = KernelDistance(self.kernel)
+        self.assertEqual(
+            distance(self.batch_x, self.batch_y).shape, torch.Size([32, 32]))