Add RelativeAverageSpectralError and RootMeanSquaredErrorUsingSlidingWindow

CHANGELOG.md

@@ -11,6 +11,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- Added`RelativeAverageSpectralError` and `RootMeanSquaredErrorUsingSlidingWindow` to image package ([#816](https://github.com/PyTorchLightning/metrics/pull/816))
+
+
 - Added support for `SpecificityAtSensitivity` Metric ([#1432](https://github.com/Lightning-AI/metrics/pull/1432))
 
 

docs/source/image/relative_average_spectral_error.rst

@@ -0,0 +1,23 @@
+.. customcarditem::
+ :header: Relative Average Spectral Error (RASE)
+ :image: https://pl-flash-data.s3.amazonaws.com/assets/thumbnails/image_classification.svg
+ :tags: Image
+
+.. include:: ../links.rst
+
+######################################
+Relative Average Spectral Error (RASE)
+######################################
+
+Module Interface
+________________
+
+.. autoclass:: torchmetrics.RelativeAverageSpectralError
+ :noindex:
+ :exclude-members: update, compute
+
+Functional Interface
+____________________
+
+.. autofunction:: torchmetrics.functional.relative_average_spectral_error
+ :noindex:

docs/source/image/root_mean_squared_error_using_sliding_window.rst

@@ -0,0 +1,23 @@
+.. customcarditem::
+ :header: Root Mean Squared Error Using Sliding Window
+ :image: https://pl-flash-data.s3.amazonaws.com/assets/thumbnails/image_classification.svg
+ :tags: Image
+
+.. include:: ../links.rst
+
+############################################
+Root Mean Squared Error Using Sliding Window
+############################################
+
+Module Interface
+________________
+
+.. autoclass:: torchmetrics.RootMeanSquaredErrorUsingSlidingWindow
+ :noindex:
+ :exclude-members: update, compute
+
+Functional Interface
+____________________
+
+.. autofunction:: torchmetrics.functional.root_mean_squared_error_using_sliding_window
+ :noindex:

docs/source/links.rst

@@ -83,6 +83,7 @@
 .. _MultiScaleSSIM: https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf
 .. _UniversalImageQualityIndex: https://ieeexplore.ieee.org/document/995823
 .. _SpectralDistortionIndex: https://www.semanticscholar.org/paper/Multispectral-and-panchromatic-data-fusion-without-Alparone-Aiazzi/b6db12e3785326577cb95fd743fecbf5bc66c7c9
+.. _RelativeAverageSpectralError: https://www.semanticscholar.org/paper/Data-Fusion.-Definitions-and-Architectures-Fusion-Wald/51b2b81e5124b3bb7ec53517a5dd64d8e348cadf
 .. _WMAPE: https://en.wikipedia.org/wiki/WMAPE
 .. _CER: https://rechtsprechung-im-ostseeraum.archiv.uni-greifswald.de/word-error-rate-character-error-rate-how-to-evaluate-a-model
 .. _MER: https://www.isca-speech.org/archive_v0/archive_papers/interspeech_2004/i04_2765.pdf

requirements/image_test.txt

@@ -4,3 +4,4 @@
 scikit-image >0.17.1, <=0.19.3
 kornia >=0.6.7, <0.6.11
 pytorch-msssim ==0.2.1
+sewar >=0.4.4, <=0.4.5

src/torchmetrics/__init__.py

@@ -48,6 +48,8 @@
  ErrorRelativeGlobalDimensionlessSynthesis,
  MultiScaleStructuralSimilarityIndexMeasure,
  PeakSignalNoiseRatio,
+ RelativeAverageSpectralError,
+ RootMeanSquaredErrorUsingSlidingWindow,
  SpectralAngleMapper,
  SpectralDistortionIndex,
  StructuralSimilarityIndexMeasure,
@@ -161,6 +163,7 @@
  "PeakSignalNoiseRatio",
  "R2Score",
  "Recall",
+ "RelativeAverageSpectralError",
  "RetrievalFallOut",
  "RetrievalHitRate",
  "RetrievalMAP",
@@ -172,6 +175,7 @@
  "RetrievalPrecisionRecallCurve",
  "RetrievalRecallAtFixedPrecision",
  "ROC",
+ "RootMeanSquaredErrorUsingSlidingWindow",
  "SacreBLEUScore",
  "SignalDistortionRatio",
  "ScaleInvariantSignalDistortionRatio",

src/torchmetrics/functional/__init__.py

@@ -37,6 +37,8 @@
 from torchmetrics.functional.image.ergas import error_relative_global_dimensionless_synthesis
 from torchmetrics.functional.image.gradients import image_gradients
 from torchmetrics.functional.image.psnr import peak_signal_noise_ratio
+from torchmetrics.functional.image.rase import relative_average_spectral_error
+from torchmetrics.functional.image.rmse_sw import root_mean_squared_error_using_sliding_window
 from torchmetrics.functional.image.sam import spectral_angle_mapper
 from torchmetrics.functional.image.ssim import (
  multiscale_structural_similarity_index_measure,
@@ -155,6 +157,7 @@
  "peak_signal_noise_ratio",
  "r2_score",
  "recall",
+ "relative_average_spectral_error",
  "retrieval_average_precision",
  "retrieval_fall_out",
  "retrieval_hit_rate",
@@ -165,6 +168,7 @@
  "retrieval_reciprocal_rank",
  "retrieval_precision_recall_curve",
  "roc",
+ "root_mean_squared_error_using_sliding_window",
  "rouge_score",
  "sacre_bleu_score",
  "signal_distortion_ratio",

src/torchmetrics/functional/image/__init__.py

@@ -15,6 +15,8 @@
 from torchmetrics.functional.image.ergas import error_relative_global_dimensionless_synthesis # noqa: F401
 from torchmetrics.functional.image.gradients import image_gradients # noqa: F401
 from torchmetrics.functional.image.psnr import peak_signal_noise_ratio # noqa: F401
+from torchmetrics.functional.image.rase import relative_average_spectral_error # noqa: F401
+from torchmetrics.functional.image.rmse_sw import root_mean_squared_error_using_sliding_window # noqa: F401
 from torchmetrics.functional.image.sam import spectral_angle_mapper # noqa: F401
 from torchmetrics.functional.image.ssim import ( # noqa: F401
  multiscale_structural_similarity_index_measure,

src/torchmetrics/functional/image/helper.py

@@ -1,4 +1,4 @@
-from typing import Sequence, Union
+from typing import Sequence, Tuple, Union
 
 import torch
 from torch import Tensor
@@ -57,6 +57,79 @@ def _gaussian_kernel_2d(
  return kernel.expand(channel, 1, kernel_size[0], kernel_size[1])
 
 
+def _uniform_weight_bias_conv2d(inputs: Tensor, window_size: int) -> Tuple[Tensor, Tensor]:
+ """Construct uniform weight and bias for a 2d convolution.
+
+ Args:
+ inputs: Input image
+ window_size: size of convolutional kernel
+
+ Return:
+ The weight and bias for 2d convolution
+ """
+ kernel_weight = torch.ones(1, 1, window_size, window_size, dtype=inputs.dtype, device=inputs.device)
+ kernel_weight /= window_size**2
+ kernel_bias = torch.zeros(1, dtype=inputs.dtype, device=inputs.device)
+ return kernel_weight, kernel_bias
+
+
+def _single_dimension_pad(inputs: Tensor, dim: int, pad: int, outer_pad: int = 0) -> Tensor:
+ """Apply single-dimension reflection padding to match scipy implementation.
+
+ Args:
+ inputs: Input image
+ dim: A dimension the image should be padded over
+ pad: Number of pads
+ outer_pad: Number of outer pads
+
+ Return:
+ Image padded over a single dimension
+ """
+ _max = inputs.shape[dim]
+ x = torch.index_select(inputs, dim, torch.arange(pad - 1, -1, -1).to(inputs.device))
+ y = torch.index_select(inputs, dim, torch.arange(_max - 1, _max - pad - outer_pad, -1).to(inputs.device))
+ return torch.cat((x, inputs, y), dim)
+
+
+def _reflection_pad_2d(inputs: Tensor, pad: int, outer_pad: int = 0) -> Tensor:
+ """Apply reflection padding to the input image.
+
+ Args:
+ inputs: Input image
+ pad: Number of pads
+ outer_pad: Number of outer pads
+
+ Return:
+ Padded image
+ """
+ for dim in [2, 3]:
+ inputs = _single_dimension_pad(inputs, dim, pad, outer_pad)
+ return inputs
+
+
+def _uniform_filter(inputs: Tensor, window_size: int) -> Tensor:
+ """Applies uniform filtew with a window of a given size over the input image.
+
+ Args:
+ inputs: Input image
+ window_size: Sliding window used for rmse calculation
+
+ Return:
+ Image transformed with the uniform input
+ """
+ inputs = _reflection_pad_2d(inputs, window_size // 2, window_size % 2)
+ kernel_weight, kernel_bias = _uniform_weight_bias_conv2d(inputs, window_size)
+ # Iterate over channels
+ inputs = torch.cat(
+ [
+ F.conv2d(inputs[:, channel].unsqueeze(1), kernel_weight, kernel_bias, padding=0)
+ for channel in range(inputs.shape[1])
+ ],
+ dim=1,
+ )
+ return inputs
+
+
 def _gaussian_kernel_3d(
  channel: int, kernel_size: Sequence[int], sigma: Sequence[float], dtype: torch.dtype, device: torch.device
 ) -> Tensor:
@@ -80,23 +153,6 @@ def _gaussian_kernel_3d(
  return kernel.expand(channel, 1, kernel_size[0], kernel_size[1], kernel_size[2])
 
 
-def _single_dimension_pad(inputs: Tensor, dim: int, pad: int) -> Tensor:
- """Reflective padding of input along a specific dimension.
-
- Args:
- inputs: tensor to pad
- dim: dimension to pad along
- pad: amount of padding to add
-
- Returns:
- padded input
- """
- _max = inputs.shape[dim] - 2
- x = torch.index_select(inputs, dim, torch.arange(pad, 0, -1, device=inputs.device))
- y = torch.index_select(inputs, dim, torch.arange(_max, _max - pad, -1, device=inputs.device))
- return torch.cat((x, inputs, y), dim)
-
-
 def _reflection_pad_3d(inputs: Tensor, pad_h: int, pad_w: int, pad_d: int) -> Tensor:
  """Reflective padding of 3d input.
 
@@ -113,8 +169,9 @@ def _reflection_pad_3d(inputs: Tensor, pad_h: int, pad_w: int, pad_d: int) -> Te
  inputs = F.pad(inputs, (pad_h, pad_h, pad_w, pad_w, pad_d, pad_d), mode="reflect")
  else:
  rank_zero_warn(
- "An older version of pyTorch is used. For optimal speed, please upgrade to at least pyTorch 1.10."
+ "An older version of pyTorch is used."
+ " For optimal speed, please upgrade to at least PyTorch v1.10 or higher."
  )
  for dim, pad in enumerate([pad_h, pad_w, pad_d]):
- inputs = _single_dimension_pad(inputs, dim + 2, pad)
+ inputs = _single_dimension_pad(inputs, dim + 2, pad, outer_pad=1)
  return inputs

src/torchmetrics/functional/image/rase.py

@@ -0,0 +1,101 @@
+# Copyright The PyTorch Lightning team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Tuple
+
+import torch
+from torch import Tensor
+
+from torchmetrics.functional.image.helper import _uniform_filter
+from torchmetrics.functional.image.rmse_sw import _rmse_sw_compute, _rmse_sw_update
+
+
+def _rase_update(
+ preds: Tensor, target: Tensor, window_size: int, rmse_map: Tensor, target_sum: Tensor, total_images: Tensor
+) -> Tuple[Tensor, Tensor, Tensor]:
+ """Calculates the sum of RMSE map values for the batch of examples and update intermediate states.
+
+ Args:
+ preds: Deformed image
+ target: Ground truth image
+ window_size: Sliding window used for RMSE calculation
+ rmse_map: Sum of RMSE map values over all examples
+ target_sum: target...
+ total_images: Total number of images
+
+ Return:
+ Intermediate state of RMSE map
+ Updated total number of already processed images
+ """
+ _, rmse_map, total_images = _rmse_sw_update(
+ preds, target, window_size, rmse_val_sum=None, rmse_map=rmse_map, total_images=total_images
+ )
+ target_sum += torch.sum(_uniform_filter(target, window_size) / (window_size**2), dim=0)
+ return rmse_map, target_sum, total_images
+
+
+def _rase_compute(rmse_map: Tensor, target_sum: Tensor, total_images: Tensor, window_size: int) -> Tensor:
+ """Compute RASE.
+
+ Args:
+ rmse_map: Sum of RMSE map values over all examples
+ target_sum: target...
+ total_images: Total number of images.
+ window_size: Sliding window used for rmse calculation
+
+ Return:
+ Relative Average Spectral Error (RASE)
+ """
+ _, rmse_map = _rmse_sw_compute(rmse_val_sum=None, rmse_map=rmse_map, total_images=total_images)
+ target_mean = target_sum / total_images
+ target_mean = target_mean.mean(0) # mean over image channels
+ rase_map = 100 / target_mean * torch.sqrt(torch.mean(rmse_map**2, 0))
+ crop_slide = round(window_size / 2)
+
+ return torch.mean(rase_map[crop_slide:-crop_slide, crop_slide:-crop_slide])
+
+
+def relative_average_spectral_error(preds: Tensor, target: Tensor, window_size: int = 8) -> Tensor:
+ """Computes Relative Average Spectral Error (RASE) (RelativeAverageSpectralError_).
+
+ Args:
+ preds: Deformed image
+ target: Ground truth image
+ window_size: Sliding window used for rmse calculation
+
+ Return:
+ Relative Average Spectral Error (RASE)
+
+ Example:
+ >>> from torchmetrics.functional import relative_average_spectral_error
+ >>> g = torch.manual_seed(22)
+ >>> preds = torch.rand(4, 3, 16, 16)
+ >>> target = torch.rand(4, 3, 16, 16)
+ >>> relative_average_spectral_error(preds, target)
+ tensor(5114.6641)
+
+ Raises:
+ ValueError: If ``window_size`` is not a positive integer.
+ """
+ if not isinstance(window_size, int) or isinstance(window_size, int) and window_size < 1:
+ raise ValueError("Argument `window_size` is expected to be a positive integer.")
+
+ img_shape = target.shape[1:] # [num_channels, width, height]
+ rmse_map = torch.zeros(img_shape, dtype=target.dtype, device=target.device)
+ target_sum = torch.zeros(img_shape, dtype=target.dtype, device=target.device)
+ total_images = torch.tensor(0.0, device=target.device)
+
+ rmse_map, target_sum, total_images = _rase_update(preds, target, window_size, rmse_map, target_sum, total_images)
+ rase = _rase_compute(rmse_map, target_sum, total_images, window_size)
+ return rase