[metrics] Update SSIM (#4566)

* [metrics] Update SSIM

* [metrics] Update SSIM

* [metrics] Update SSIM

* [metrics] Update SSIM

* [metrics] update ssim

* dist_sync_on_step True

* [metrics] update ssim

* Update tests/metrics/regression/test_ssim.py

Co-authored-by: chaton <thomas@grid.ai>

* Update pytorch_lightning/metrics/functional/ssim.py

Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>

* ddp=True

* Update test_ssim.py

Co-authored-by: chaton <thomas@grid.ai>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: Nicki Skafte <skaftenicki@gmail.com>

pytorch_lightning/metrics/functional/ssim.py

@@ -19,17 +19,17 @@
 from torch.nn import functional as F
 
 
-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)
+def _gaussian(kernel_size: int, sigma: int, dtype: torch.dtype, device: torch.device):
+    dist = torch.arange(start=(1 - kernel_size) / 2, end=(1 + kernel_size) / 2, step=1, dtype=dtype, device=device)
+    gauss = torch.exp(-torch.pow(dist / sigma, 2) / 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)
+def _gaussian_kernel(channel: int, kernel_size: Sequence[int], sigma: Sequence[float],
+                     dtype: torch.dtype, device: torch.device):
+    gaussian_kernel_x = _gaussian(kernel_size[0], sigma[0], dtype, device)
+    gaussian_kernel_y = _gaussian(kernel_size[1], sigma[1], dtype, 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])
 
@@ -82,9 +82,15 @@ def _ssim_compute(
     device = preds.device
 
     channel = preds.size(1)
-    kernel = _gaussian_kernel(channel, kernel_size, sigma, device)
+    dtype = preds.dtype
+    kernel = _gaussian_kernel(channel, kernel_size, sigma, dtype, device)
+    pad_w = (kernel_size[0] - 1) // 2
+    pad_h = (kernel_size[1] - 1) // 2
+
+    preds = F.pad(preds, (pad_w, pad_w, pad_h, pad_h), mode='reflect')
+    target = F.pad(target, (pad_w, pad_w, pad_h, pad_h), mode='reflect')
 
-    input_list = torch.cat([preds, target, preds * preds, target * target, preds * target])  # (5 * B, C, H, W)
+    input_list = torch.cat((preds, target, preds * preds, target * target, preds * target))  # (5 * B, C, H, W)
     outputs = F.conv2d(input_list, kernel, groups=channel)
     output_list = [outputs[x * preds.size(0): (x + 1) * preds.size(0)] for x in range(len(outputs))]
 
@@ -100,6 +106,7 @@ def _ssim_compute(
     lower = sigma_pred_sq + sigma_target_sq + c2
 
     ssim_idx = ((2 * mu_pred_target + c1) * upper) / ((mu_pred_sq + mu_target_sq + c1) * lower)
+    ssim_idx = ssim_idx[..., pad_h:-pad_h, pad_w:-pad_w]
 
     return reduce(ssim_idx, reduction)
 

tests/metrics/regression/test_ssim.py

@@ -16,13 +16,13 @@
 
 
 _inputs = []
-for size, channel, coef, multichannel in [
-    (16, 1, 0.9, False),
-    (32, 3, 0.8, True),
-    (48, 4, 0.7, True),
-    (64, 5, 0.6, True),
+for size, channel, coef, multichannel, dtype in [
+    (12, 3, 0.9, True, torch.float),
+    (13, 1, 0.8, False, torch.float32),
+    (14, 1, 0.7, False, torch.double),
+    (15, 3, 0.6, True, torch.float64),
 ]:
-    preds = torch.rand(NUM_BATCHES, BATCH_SIZE, channel, size, size)
+    preds = torch.rand(NUM_BATCHES, BATCH_SIZE, channel, size, size, dtype=dtype)
     _inputs.append(
         Input(
             preds=preds,
@@ -41,7 +41,8 @@ def _sk_metric(preds, target, data_range, multichannel):
         sk_target = sk_target[:, :, :, 0]
 
     return structural_similarity(
-        sk_target, sk_preds, data_range=data_range, multichannel=multichannel, gaussian_weights=True, win_size=11
+        sk_target, sk_preds, data_range=data_range, multichannel=multichannel,
+        gaussian_weights=True, win_size=11, sigma=1.5, use_sample_covariance=False
     )
 
 
@@ -50,7 +51,7 @@ def _sk_metric(preds, target, data_range, multichannel):
     [(i.preds, i.target, i.multichannel) for i in _inputs],
 )
 class TestSSIM(MetricTester):
-    atol = 1e-3  # TODO: ideally tests should pass with lower tolerance
+    atol = 6e-5
 
     # TODO: for some reason this test hangs with ddp=True
     # @pytest.mark.parametrize("ddp", [True, False])

tests/metrics/utils.py

@@ -24,7 +24,7 @@ def setup_ddp(rank, world_size):
     os.environ["MASTER_ADDR"] = 'localhost'
     os.environ['MASTER_PORT'] = '8088'
 
-    if torch.distributed.is_available() and sys.platform not in ['win32', 'cygwin']:
+    if torch.distributed.is_available() and sys.platform not in ('win32', 'cygwin'):
         torch.distributed.init_process_group("gloo", rank=rank, world_size=world_size)