Add Total Variation (TV) to metrax

src/metrax/__init__.py

@@ -45,6 +45,7 @@
 RougeN = nlp_metrics.RougeN
 SNR = audio_metrics.SNR
 SSIM = image_metrics.SSIM
+TotalVariation = image_metrics.TotalVariation
 WER = nlp_metrics.WER
 
 

src/metrax/image_metrics.py

@@ -666,3 +666,69 @@ def compute(self) -> jax.Array:
     """Returns the final Dice coefficient."""
     epsilon = 1e-7
     return (2.0 * self.intersection) / (self.sum_pred + self.sum_true + epsilon)
+
+@flax.struct.dataclass
+class TotalVariation(base.Average):
+  r"""Calculates and returns the Total Variation (TV) for one or more images.
+
+  The total variation is the sum of the absolute differences for neighboring
+  pixel-values in the input images. This measures how much noise is in the
+  images.
+
+  This implements the anisotropic 2-D version of the formula described here:
+
+  https://en.wikipedia.org/wiki/Total_variation_denoising
+  """
+
+  @staticmethod
+  def _calculate_total_variation(
+      images: jax.Array,
+  ) -> jax.Array:
+    """Computes Total Variation values.
+
+    Args:
+      images: 4-D Array of shape ``(batch, height, width, channels)`` or
+      3-D Array of shape ``(height, width, channels)``.
+
+    Returns:
+      Total variation of 'images'.
+
+      If `images` was 4-D, return a 1-D float Array of shape `[batch]` with the
+        total variation for each image in the batch.
+      If `images` was 3-D, return a scalar float with the total variation for
+        that image.
+    """
+    ndims = images.ndim 
+    if ndims == 3: # (height, width, channels)
+      # Shift images by one pixel along the height and width.
+      pixel_dif1 = jnp.abs(images[1:, :, :] - images[:-1, :, :])
+      pixel_dif2 = jnp.abs(images[:, 1:, :] - images[:, :-1, :])
+      sum_axis = None
+    elif ndims == 4: # (batch, height, width, channels)
+      # Shift images by one pixel along the height and width.
+      pixel_dif1 = jnp.abs(images[:, 1:, :, :] - images[:, :-1, :, :])
+      pixel_dif2 = jnp.abs(images[:, :, 1:, :] - images[:, :, :-1, :])
+      sum_axis = [1, 2, 3]
+    else:
+      raise ValueError(
+        f'Input images must be either 3 or 4-dimensional, got {ndims} dimensions instead.'
+      )
+
+    return jnp.sum(pixel_dif1, axis=sum_axis) + jnp.sum(pixel_dif2, axis=sum_axis)
+
+
+  @classmethod
+  def from_model_output(
+    cls,
+    predictions: jax.Array
+  ) -> 'TotalVariation':
+    """Computes the Total Variation for a batch of images and creates a TotalVariation metric instance.
+
+    Args:
+      predictions: A JAX array of predicted images, with shape ``(batch, H, W, C)``.
+
+    Returns:
+      A ``TotalVariation`` instance containing per‑image total variation values.
+    """
+    total_variation = cls._calculate_total_variation(predictions)
+    return super().from_model_output(values=total_variation)

src/metrax/image_metrics_test.py

@@ -212,6 +212,27 @@
 DICE_ALL_ZEROS = (np.array([0, 0, 0, 0]), np.array([0, 0, 0, 0]))
 DICE_NO_OVERLAP = (np.array([1, 1, 0, 0]), np.array([0, 0, 1, 1]))
 
+# Test data for TotalVariation
+# Case 1: Basic, float normalized [0,1], single channel (3D)
+TV_IMG_SHAPE_1 = (16, 16, 1)  # height, width, channels
+TV_IMG_1 = np.random.rand(*TV_IMG_SHAPE_1).astype(np.float32)
+
+# Case 2: Multi-channel (3), float normalized [0,1] (3D)
+TV_IMG_SHAPE_2 = (32, 32, 3)
+TV_IMG_2 = np.random.rand(*TV_IMG_SHAPE_2).astype(np.float32)
+
+# Case 3: Batch of single channel images (4D)
+TV_IMG_SHAPE_3 = (4, 16, 16, 1)  # batch, height, width, channels
+TV_IMG_3 = np.random.rand(*TV_IMG_SHAPE_3).astype(np.float32)
+
+# Case 4: Batch of multi-channel images (4D)
+TV_IMG_SHAPE_4 = (4, 32, 32, 3)  # batch, height, width, channels
+TV_IMG_4 = np.random.rand(*TV_IMG_SHAPE_4).astype(np.float32)
+
+# Case 5: Constant image (should have zero variation)
+TV_IMG_SHAPE_5 = (16, 16, 1)
+TV_IMG_5 = np.ones(TV_IMG_SHAPE_5, dtype=np.float32)
+
 
 class ImageMetricsTest(parameterized.TestCase):
 
@@ -553,6 +574,60 @@ def test_dice(self, y_true, y_pred):
 
     np.testing.assert_allclose(result, expected, rtol=1e-5, atol=1e-5)
 
+  @parameterized.named_parameters(
+      (
+          'tv_single_channel_3d',
+          TV_IMG_1,
+          False,  # is_batch
+      ),
+      (
+          'tv_multichannel_3d',
+          TV_IMG_2,
+          False,  # is_batch
+      ),
+      (
+          'tv_batch_single_channel_4d',
+          TV_IMG_3,
+          True,  # is_batch
+      ),
+      (
+          'tv_batch_multichannel_4d',
+          TV_IMG_4,
+          True,  # is_batch
+      ),
+      (
+          'tv_constant_image',
+          TV_IMG_5,
+          False,  # is_batch
+      ),
+  )
+  def test_total_variation_against_tensorflow(
+      self,
+      images_np: np.ndarray,
+      is_batch: bool,
+  ):
+    """Test that TotalVariation metric computes values close to tf.image.total_variation."""
+    # Calculate TV using Metrax
+    metrax_tv = metrax.TotalVariation.from_model_output(
+        predictions=jnp.array(images_np)
+    ).compute()
+
+    # Calculate TV using TensorFlow
+    tf_tv = tf.image.total_variation(tf.convert_to_tensor(images_np))
+    tf_mean = tf.reduce_mean(tf_tv).numpy()
+
+    # For constant image, TV should be 0
+    if np.array_equal(images_np, TV_IMG_5):
+      np.testing.assert_allclose(metrax_tv, 0.0, rtol=1e-5, atol=1e-5)
+      np.testing.assert_allclose(tf_mean, 0.0, rtol=1e-5, atol=1e-5)
+    else:
+      np.testing.assert_allclose(
+          metrax_tv,
+          tf_mean,
+          rtol=1e-5,
+          atol=1e-5,
+          err_msg='Total Variation mismatch',
+      )
 
 if __name__ == '__main__':
   absltest.main()

src/metrax/nnx/__init__.py

@@ -39,6 +39,7 @@
 RougeN = nnx_metrics.RougeN
 SNR = nnx_metrics.SNR
 SSIM = nnx_metrics.SSIM
+TotalVariation = nnx_metrics.TotalVariation
 WER = nnx_metrics.WER
 
 

src/metrax/nnx/nnx_metrics.py

@@ -193,6 +193,13 @@ def __init__(self):
     super().__init__(metrax.SSIM)
 
 
+class TotalVariation(NnxWrapper):
+  """An NNX class for the Metrax metric TotalVariation."""
+
+  def __init__(self):
+    super().__init__(metrax.TotalVariation)
+
+
 class WER(NnxWrapper):
   """An NNX class for the Metrax metric WER."""