metrics: add SSIM (#2671)

* metrics: add SSIM

* Update CHANGELOG.md

fix codefactor issue

fix doctest

fix doctest

fix test

* added test for raise Error

CHANGELOG.md

@@ -9,6 +9,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Added SSIM metrics ([#2671](https://github.com/PyTorchLightning/pytorch-lightning/pull/2671))
 - Added BLEU metrics ([#2535](https://github.com/PyTorchLightning/pytorch-lightning/pull/2535))
 
 - Added support for `Trainer(num_sanity_val_steps=-1)` to check all validation data before training ([#2246](https://github.com/PyTorchLightning/pytorch-lightning/pull/2246))

docs/source/metrics.rst

@@ -234,6 +234,12 @@ RMSLE
 .. autoclass:: pytorch_lightning.metrics.regression.RMSLE
     :noindex:
 
+SSIM
+^^^^
+
+.. autoclass:: pytorch_lightning.metrics.regression.SSIM
+    :noindex:
+
 ----------------
 
 Functional Metrics
@@ -403,6 +409,12 @@ psnr (F)
 .. autofunction:: pytorch_lightning.metrics.functional.psnr
     :noindex:
 
+ssim (F)
+^^^^^^^^
+
+.. autofunction:: pytorch_lightning.metrics.functional.ssim
+    :noindex:
+
 stat_scores_multiple_classes (F)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

pytorch_lightning/metrics/__init__.py

@@ -6,6 +6,7 @@
     PSNR,
     RMSE,
     RMSLE,
+    SSIM
 )
 from pytorch_lightning.metrics.classification import (
     Accuracy,
@@ -54,6 +55,7 @@
     "PSNR",
     "RMSE",
     "RMSLE",
+    "SSIM"
 ]
 __sequence_metrics = ["BLEUScore"]
 __all__ = __regression_metrics + __classification_metrics + ["SklearnMetric"] + __sequence_metrics

pytorch_lightning/metrics/functional/__init__.py

@@ -26,5 +26,6 @@
     psnr,
     rmse,
     rmsle,
+    ssim
 )
 from pytorch_lightning.metrics.functional.nlp import bleu_score

pytorch_lightning/metrics/functional/regression.py

@@ -1,3 +1,5 @@
+from typing import Sequence
+
 import torch
 from torch.nn import functional as F
 
@@ -182,3 +184,116 @@ def psnr(
     psnr_base_e = 2 * torch.log(data_range) - torch.log(mse_score)
     psnr = psnr_base_e * (10 / torch.log(torch.tensor(base)))
     return psnr
+
+
+def _gaussian_kernel(channel, kernel_size, sigma, device):
+    def gaussian(kernel_size, sigma, device):
+        gauss = torch.arange(
+            start=(1 - kernel_size) / 2, end=(1 + kernel_size) / 2, step=1, dtype=torch.float32, device=device
+        )
+        gauss = torch.exp(-gauss.pow(2) / (2 * pow(sigma, 2)))
+        return (gauss / gauss.sum()).unsqueeze(dim=0)  # (1, kernel_size)
+
+    gaussian_kernel_x = gaussian(kernel_size[0], sigma[0], device)
+    gaussian_kernel_y = gaussian(kernel_size[1], sigma[1], device)
+    kernel = torch.matmul(gaussian_kernel_x.t(), gaussian_kernel_y)  # (kernel_size, 1) * (1, kernel_size)
+
+    return kernel.expand(channel, 1, kernel_size[0], kernel_size[1])
+
+
+def ssim(
+    pred: torch.Tensor,
+    target: torch.Tensor,
+    kernel_size: Sequence[int] = (11, 11),
+    sigma: Sequence[float] = (1.5, 1.5),
+    reduction: str = "elementwise_mean",
+    data_range: float = None,
+    k1: float = 0.01,
+    k2: float = 0.03
+) -> torch.Tensor:
+    """
+    Computes Structual Similarity Index Measure
+
+    Args:
+        pred: Estimated image
+        target: Ground truth image
+        kernel_size: Size of the gaussian kernel. Default: (11, 11)
+        sigma: Standard deviation of the gaussian kernel. Default: (1.5, 1.5)
+        reduction: A method for reducing ssim over all elements in the ``pred`` tensor. Default: ``elementwise_mean``
+
+            Available reduction methods:
+            - elementwise_mean: takes the mean
+            - none: pass away
+            - sum: add elements
+
+        data_range: Range of the image. If ``None``, it is determined from the image (max - min)
+        k1: Parameter of SSIM. Default: 0.01
+        k2: Parameter of SSIM. Default: 0.03
+
+    Returns:
+        A Tensor with SSIM
+
+    Example:
+
+        >>> pred = torch.rand([16, 1, 16, 16])
+        >>> target = pred * 1.25
+        >>> ssim(pred, target)
+        tensor(0.9520)
+    """
+
+    if pred.dtype != target.dtype:
+        raise TypeError(
+            "Expected `pred` and `target` to have the same data type."
+            f" Got pred: {pred.dtype} and target: {target.dtype}."
+        )
+
+    if pred.shape != target.shape:
+        raise ValueError(
+            "Expected `pred` and `target` to have the same shape."
+            f" Got pred: {pred.shape} and target: {target.shape}."
+        )
+
+    if len(pred.shape) != 4 or len(target.shape) != 4:
+        raise ValueError(
+            "Expected `pred` and `target` to have BxCxHxW shape."
+            f" Got pred: {pred.shape} and target: {target.shape}."
+        )
+
+    if len(kernel_size) != 2 or len(sigma) != 2:
+        raise ValueError(
+            "Expected `kernel_size` and `sigma` to have the length of two."
+            f" Got kernel_size: {len(kernel_size)} and sigma: {len(sigma)}."
+        )
+
+    if any(x % 2 == 0 or x <= 0 for x in kernel_size):
+        raise ValueError(f"Expected `kernel_size` to have odd positive number. Got {kernel_size}.")
+
+    if any(y <= 0 for y in sigma):
+        raise ValueError(f"Expected `sigma` to have positive number. Got {sigma}.")
+
+    if data_range is None:
+        data_range = max(pred.max() - pred.min(), target.max() - target.min())
+
+    C1 = pow(k1 * data_range, 2)
+    C2 = pow(k2 * data_range, 2)
+    device = pred.device
+
+    channel = pred.size(1)
+    kernel = _gaussian_kernel(channel, kernel_size, sigma, device)
+    mu_pred = F.conv2d(pred, kernel, groups=channel)
+    mu_target = F.conv2d(target, kernel, groups=channel)
+
+    mu_pred_sq = mu_pred.pow(2)
+    mu_target_sq = mu_target.pow(2)
+    mu_pred_target = mu_pred * mu_target
+
+    sigma_pred_sq = F.conv2d(pred * pred, kernel, groups=channel) - mu_pred_sq
+    sigma_target_sq = F.conv2d(target * target, kernel, groups=channel) - mu_target_sq
+    sigma_pred_target = F.conv2d(pred * target, kernel, groups=channel) - mu_pred_target
+
+    UPPER = 2 * sigma_pred_target + C2
+    LOWER = sigma_pred_sq + sigma_target_sq + C2
+
+    ssim_idx = ((2 * mu_pred_target + C1) * UPPER) / ((mu_pred_sq + mu_target_sq + C1) * LOWER)
+
+    return reduce(ssim_idx, reduction)

pytorch_lightning/metrics/regression.py

@@ -1,11 +1,14 @@
+from typing import Sequence
+
 import torch
 
 from pytorch_lightning.metrics.functional.regression import (
     mae,
     mse,
     psnr,
     rmse,
-    rmsle
+    rmsle,
+    ssim
 )
 from pytorch_lightning.metrics.metric import Metric
 
@@ -229,3 +232,62 @@ def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
             A Tensor with psnr score.
         """
         return psnr(pred, target, self.data_range, self.base, self.reduction)
+
+
+class SSIM(Metric):
+    """
+    Computes Structual Similarity Index Measure
+
+    Example:
+
+        >>> pred = torch.rand([16, 1, 16, 16])
+        >>> target = pred * 1.25
+        >>> metric = SSIM()
+        >>> metric(pred, target)
+        tensor(0.9520)
+    """
+
+    def __init__(
+        self,
+        kernel_size: Sequence[int] = (11, 11),
+        sigma: Sequence[float] = (1.5, 1.5),
+        reduction: str = "elementwise_mean",
+        data_range: float = None,
+        k1: float = 0.01,
+        k2: float = 0.03
+    ):
+        """
+        Args:
+            kernel_size: Size of the gaussian kernel. Default: (11, 11)
+            sigma: Standard deviation of the gaussian kernel. Default: (1.5, 1.5)
+            reduction: A method for reducing ssim. Default: ``elementwise_mean``
+
+                Available reduction methods:
+                - elementwise_mean: takes the mean
+                - none: pass away
+                - sum: add elements
+
+            data_range: Range of the image. If ``None``, it is determined from the image (max - min)
+            k1: Parameter of SSIM. Default: 0.01
+            k2: Parameter of SSIM. Default: 0.03
+        """
+        super().__init__(name="ssim")
+        self.kernel_size = kernel_size
+        self.sigma = sigma
+        self.reduction = reduction
+        self.data_range = data_range
+        self.k1 = k1
+        self.k2 = k2
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """
+        Actual metric computation
+
+        Args:
+            pred: Estimated image
+            target: Ground truth image
+
+        Return:
+            torch.Tensor: SSIM Score
+        """
+        return ssim(pred, target, self.kernel_size, self.sigma, self.reduction, self.data_range, self.k1, self.k2)

tests/metrics/functional/test_regression.py

@@ -2,13 +2,15 @@
 import torch
 import numpy as np
 from skimage.metrics import peak_signal_noise_ratio as ski_psnr
+from skimage.metrics import structural_similarity as ski_ssim
 
 from pytorch_lightning.metrics.functional import (
     mae,
     mse,
     psnr,
     rmse,
-    rmsle
+    rmsle,
+    ssim
 )
 
 
@@ -93,3 +95,50 @@ def test_psnr_against_sklearn(sklearn_metric, torch_metric):
         sk_score = torch.tensor(sk_score, dtype=torch.float, device=device)
         pl_score = torch_metric(pred, target, data_range=n_cls_target)
         assert torch.allclose(sk_score, pl_score)
+
+
+@pytest.mark.parametrize(['size', 'channel', 'plus', 'multichannel'], [
+    pytest.param(16, 1, 0.125, False),
+    pytest.param(32, 1, 0.25, False),
+    pytest.param(48, 3, 0.5, True),
+    pytest.param(64, 4, 0.75, True),
+    pytest.param(128, 5, 1, True)
+])
+def test_ssim(size, channel, plus, multichannel):
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    pred = torch.rand(1, channel, size, size, device=device)
+    target = pred + plus
+    ssim_idx = ssim(pred, target)
+    np_pred = np.random.rand(size, size, channel)
+    if multichannel is False:
+        np_pred = np_pred[:, :, 0]
+    np_target = np.add(np_pred, plus)
+    sk_ssim_idx = ski_ssim(np_pred, np_target, win_size=11, multichannel=multichannel, gaussian_weights=True)
+    assert torch.allclose(ssim_idx, torch.tensor(sk_ssim_idx, dtype=torch.float, device=device), atol=1e-2, rtol=1e-2)
+
+    ssim_idx = ssim(pred, pred)
+    assert torch.allclose(ssim_idx, torch.tensor(1.0, device=device))
+
+
+@pytest.mark.parametrize(['pred', 'target', 'kernel', 'sigma'], [
+    pytest.param([1, 1, 16, 16], [1, 16, 16], [11, 11], [1.5, 1.5]),  # shape
+    pytest.param([1, 16, 16], [1, 16, 16], [11, 11], [1.5, 1.5]),  # len(shape)
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, 11], [1.5]),  # len(kernel), len(sigma)
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11], [1.5, 1.5]),  # len(kernel), len(sigma)
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11], [1.5]),  # len(kernel), len(sigma)
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, 0], [1.5, 1.5]),  # invalid kernel input
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, 10], [1.5, 1.5]),  # invalid kernel input
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, -11], [1.5, 1.5]),  # invalid kernel input
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, 11], [1.5, 0]),  # invalid sigma input
+    pytest.param([1, 1, 16, 16], [1, 1, 16, 16], [11, 0], [1.5, -1.5]),  # invalid sigma input
+])
+def test_ssim_invalid_inputs(pred, target, kernel, sigma):
+    pred_t = torch.rand(pred)
+    target_t = torch.rand(target, dtype=torch.float64)
+    with pytest.raises(TypeError):
+        ssim(pred_t, target_t)
+
+    pred = torch.rand(pred)
+    target = torch.rand(target)
+    with pytest.raises(ValueError):
+        ssim(pred, target, kernel, sigma)

tests/metrics/test_regression.py

@@ -6,7 +6,7 @@
 from skimage.metrics import peak_signal_noise_ratio as ski_psnr
 
 from pytorch_lightning.metrics.regression import (
-    MAE, MSE, RMSE, RMSLE, PSNR
+    MAE, MSE, RMSE, RMSLE, PSNR, SSIM
 )
 
 
@@ -58,3 +58,13 @@ def test_psnr():
     target = torch.tensor([0., 1, 2, 2])
     score = psnr(pred, target)
     assert isinstance(score, torch.Tensor)
+
+
+def test_ssim():
+    ssim = SSIM()
+    assert ssim.name == 'ssim'
+
+    pred = torch.rand([16, 1, 16, 16])
+    target = pred * 1.25
+    score = ssim(pred, target)
+    assert isinstance(score, torch.Tensor)