Fix simplify ssim implementation (#2563)

* Fixed SSIM issue with variable batch

* Fixed mypy issue

* simplify ssim implementation

* Fix bug

Co-authored-by: vfdev-5 <vfdev.5@gmail.com>

ignite/metrics/ssim.py

@@ -98,8 +98,7 @@ def __init__(
 
     @reinit__is_reduced
     def reset(self) -> None:
-        # Not a tensor because batch size is not known in advance.
-        self._sum_of_batchwise_ssim = 0.0  # type: Union[float, torch.Tensor]
+        self._sum_of_ssim = torch.tensor(0.0, device=self._device)
         self._num_examples = 0
         self._kernel = self._gaussian_or_uniform_kernel(kernel_size=self.kernel_size, sigma=self.sigma)
 
@@ -176,11 +175,12 @@ def update(self, output: Sequence[torch.Tensor]) -> None:
         b2 = sigma_pred_sq + sigma_target_sq + self.c2
 
         ssim_idx = (a1 * a2) / (b1 * b2)
-        self._sum_of_batchwise_ssim += torch.mean(ssim_idx, (1, 2, 3), dtype=torch.float64).to(self._device)
+        self._sum_of_ssim += torch.mean(ssim_idx, (1, 2, 3), dtype=torch.float64).sum().to(self._device)
+
         self._num_examples += y.shape[0]
 
-    @sync_all_reduce("_sum_of_batchwise_ssim", "_num_examples")
+    @sync_all_reduce("_sum_of_ssim", "_num_examples")
     def compute(self) -> torch.Tensor:
         if self._num_examples == 0:
             raise NotComputableError("SSIM must have at least one example before it can be computed.")
-        return torch.sum(self._sum_of_batchwise_ssim / self._num_examples)  # type: ignore[arg-type]
+        return self._sum_of_ssim / self._num_examples

tests/ignite/metrics/test_ssim.py

@@ -68,8 +68,17 @@ def test_invalid_ssim():
         ssim.compute()
 
 
-def _test_ssim(y_pred, y, data_range, kernel_size, sigma, gaussian, use_sample_covariance, device):
-    atol = 7e-5
+@pytest.mark.parametrize("device", ["cpu"] + ["cuda"] if torch.cuda.is_available() else [])
+@pytest.mark.parametrize(
+    "shape, kernel_size, gaussian, use_sample_covariance",
+    [[(8, 3, 224, 224), 7, False, True], [(12, 3, 28, 28), 11, True, False]],
+)
+def test_ssim(device, shape, kernel_size, gaussian, use_sample_covariance):
+    y_pred = torch.rand(shape, device=device)
+    y = y_pred * 0.8
+
+    sigma = 1.5
+    data_range = 1.0
     ssim = SSIM(data_range=data_range, sigma=sigma, device=device)
     ssim.update((y_pred, y))
     ignite_ssim = ssim.compute()
@@ -89,23 +98,34 @@ def _test_ssim(y_pred, y, data_range, kernel_size, sigma, gaussian, use_sample_c
 
     assert isinstance(ignite_ssim, torch.Tensor)
     assert ignite_ssim.dtype == torch.float64
-    assert ignite_ssim.device == torch.device(device)
-    assert np.allclose(ignite_ssim.cpu().numpy(), skimg_ssim, atol=atol)
-
-
-def test_ssim():
-    device = "cuda:0" if torch.cuda.is_available() else "cpu"
-    y_pred = torch.rand(8, 3, 224, 224, device=device)
-    y = y_pred * 0.8
-    _test_ssim(
-        y_pred, y, data_range=1.0, kernel_size=7, sigma=1.5, gaussian=False, use_sample_covariance=True, device=device
-    )
+    assert ignite_ssim.device.type == torch.device(device).type
+    assert np.allclose(ignite_ssim.cpu().numpy(), skimg_ssim, atol=7e-5)
+
+
+def test_ssim_variable_batchsize():
+    # Checks https://github.com/pytorch/ignite/issues/2532
+    sigma = 1.5
+    data_range = 1.0
+    ssim = SSIM(data_range=data_range, sigma=sigma)
+
+    y_preds = [
+        torch.rand(12, 3, 28, 28),
+        torch.rand(12, 3, 28, 28),
+        torch.rand(8, 3, 28, 28),
+        torch.rand(16, 3, 28, 28),
+        torch.rand(1, 3, 28, 28),
+        torch.rand(30, 3, 28, 28),
+    ]
+    y_true = [v * 0.8 for v in y_preds]
+
+    for y_pred, y in zip(y_preds, y_true):
+        ssim.update((y_pred, y))
 
-    y_pred = torch.rand(12, 3, 28, 28, device=device)
-    y = y_pred * 0.8
-    _test_ssim(
-        y_pred, y, data_range=1.0, kernel_size=11, sigma=1.5, gaussian=True, use_sample_covariance=False, device=device
-    )
+    out = ssim.compute()
+    ssim.reset()
+    ssim.update((torch.cat(y_preds), torch.cat(y_true)))
+    expected = ssim.compute()
+    assert torch.allclose(out, expected)
 
 
 def _test_distrib_integration(device, tol=1e-4):
@@ -186,7 +206,7 @@ def _test_distrib_accumulator_device(device):
         y = y_pred * 0.65
         ssim.update((y_pred, y))
 
-        dev = ssim._sum_of_batchwise_ssim.device
+        dev = ssim._sum_of_ssim.device
         assert dev == metric_device, f"{type(dev)}:{dev} vs {type(metric_device)}:{metric_device}"