Add pairwise function for 'gaussian' and 'concatenation' mode in NonLocal. (#424)

* add pairwise function for 'gaussian' and 'concatenation' mode

* rename test function

* decrease the complexity of nonlocal unittest

* fix typo and make unittest more complete

* add unittest when zero_init is False

* minor fix

* pack theta and phi
Co-authored-by: Jiarui XU <xvjiarui0826@gmail.com>

mmcv/cnn/bricks/non_local.py

@@ -14,6 +14,7 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
     This module is proposed in
     "Non-local Neural Networks"
     Paper reference: https://arxiv.org/abs/1711.07971
+    Code reference: https://github.com/AlexHex7/Non-local_pytorch
 
     Args:
         in_channels (int): Channels of the input feature map.
@@ -26,8 +27,8 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
             Default: None.
         norm_cfg (None | dict): The config dict for normalization layers.
             Default: None. (This parameter is only applicable to conv_out.)
-        mode (str): Options are `embedded_gaussian` and `dot_product`.
-            Default: embedded_gaussian.
+        mode (str): Options are `gaussian`, `concatenation`,
+            `embedded_gaussian` and `dot_product`. Default: embedded_gaussian.
     """
 
     def __init__(self,
@@ -42,13 +43,15 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
         self.in_channels = in_channels
         self.reduction = reduction
         self.use_scale = use_scale
-        self.inter_channels = in_channels // reduction
+        self.inter_channels = max(in_channels // reduction, 1)
         self.mode = mode
 
-        if mode not in ['embedded_gaussian', 'dot_product']:
-            raise ValueError(
-                "Mode should be in 'embedded_gaussian' or 'dot_product', "
-                f'but got {mode} instead.')
+        if mode not in [
+                'gaussian', 'embedded_gaussian', 'dot_product', 'concatenation'
+        ]:
+            raise ValueError("Mode should be in 'gaussian', 'concatenation', "
+                             f"'embedded_gaussian' or 'dot_product', but got "
+                             f'{mode} instead.')
 
         # g, theta, phi are defaulted as `nn.ConvNd`.
         # Here we use ConvModule for potential usage.
@@ -58,18 +61,6 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
             kernel_size=1,
             conv_cfg=conv_cfg,
             act_cfg=None)
-        self.theta = ConvModule(
-            self.in_channels,
-            self.inter_channels,
-            kernel_size=1,
-            conv_cfg=conv_cfg,
-            act_cfg=None)
-        self.phi = ConvModule(
-            self.in_channels,
-            self.inter_channels,
-            kernel_size=1,
-            conv_cfg=conv_cfg,
-            act_cfg=None)
         self.conv_out = ConvModule(
             self.inter_channels,
             self.in_channels,
@@ -78,11 +69,38 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
             norm_cfg=norm_cfg,
             act_cfg=None)
 
+        if self.mode != 'gaussian':
+            self.theta = ConvModule(
+                self.in_channels,
+                self.inter_channels,
+                kernel_size=1,
+                conv_cfg=conv_cfg,
+                act_cfg=None)
+            self.phi = ConvModule(
+                self.in_channels,
+                self.inter_channels,
+                kernel_size=1,
+                conv_cfg=conv_cfg,
+                act_cfg=None)
+
+        if self.mode == 'concatenation':
+            self.concat_project = ConvModule(
+                self.inter_channels * 2,
+                1,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=False,
+                act_cfg=dict(type='ReLU'))
+
         self.init_weights(**kwargs)
 
     def init_weights(self, std=0.01, zeros_init=True):
-        for m in [self.g, self.theta, self.phi]:
-            normal_init(m.conv, std=std)
+        if self.mode != 'gaussian':
+            for m in [self.g, self.theta, self.phi]:
+                normal_init(m.conv, std=std)
+        else:
+            normal_init(self.g.conv, std=std)
         if zeros_init:
             if self.conv_out.norm_cfg is None:
                 constant_init(self.conv_out.conv, 0)
@@ -94,6 +112,14 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
             else:
                 normal_init(self.conv_out.norm, std=std)
 
+    def gaussian(self, theta_x, phi_x):
+        # NonLocal1d pairwise_weight: [N, H, H]
+        # NonLocal2d pairwise_weight: [N, HxW, HxW]
+        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
+        pairwise_weight = torch.matmul(theta_x, phi_x)
+        pairwise_weight = pairwise_weight.softmax(dim=-1)
+        return pairwise_weight
+
     def embedded_gaussian(self, theta_x, phi_x):
         # NonLocal1d pairwise_weight: [N, H, H]
         # NonLocal2d pairwise_weight: [N, HxW, HxW]
@@ -113,8 +139,27 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
         pairwise_weight /= pairwise_weight.shape[-1]
         return pairwise_weight
 
+    def concatenation(self, theta_x, phi_x):
+        # NonLocal1d pairwise_weight: [N, H, H]
+        # NonLocal2d pairwise_weight: [N, HxW, HxW]
+        # NonLocal3d pairwise_weight: [N, TxHxW, TxHxW]
+        h = theta_x.size(2)
+        w = phi_x.size(3)
+        theta_x = theta_x.repeat(1, 1, 1, w)
+        phi_x = phi_x.repeat(1, 1, h, 1)
+
+        concat_feature = torch.cat([theta_x, phi_x], dim=1)
+        pairwise_weight = self.concat_project(concat_feature)
+        n, _, h, w = pairwise_weight.size()
+        pairwise_weight = pairwise_weight.view(n, h, w)
+        pairwise_weight /= pairwise_weight.shape[-1]
+
+        return pairwise_weight
+
     def forward(self, x):
         # Assume `reduction = 1`, then `inter_channels = C`
+        # or `inter_channels = C` when `mode="gaussian"`
+
         # NonLocal1d x: [N, C, H]
         # NonLocal2d x: [N, C, H, W]
         # NonLocal3d x: [N, C, T, H, W]
@@ -126,16 +171,23 @@ class _NonLocalNd(nn.Module, metaclass=ABCMeta):
         g_x = self.g(x).view(n, self.inter_channels, -1)
         g_x = g_x.permute(0, 2, 1)
 
-        # NonLocal1d theta_x: [N, H, C]
-        # NonLocal2d theta_x: [N, HxW, C]
-        # NonLocal3d theta_x: [N, TxHxW, C]
-        theta_x = self.theta(x).view(n, self.inter_channels, -1)
-        theta_x = theta_x.permute(0, 2, 1)
-
-        # NonLocal1d phi_x: [N, C, H]
-        # NonLocal2d phi_x: [N, C, HxW]
-        # NonLocal3d phi_x: [N, C, TxHxW]
-        phi_x = self.phi(x).view(n, self.inter_channels, -1)
+        # NonLocal1d theta_x: [N, H, C], phi_x: [N, C, H]
+        # NonLocal2d theta_x: [N, HxW, C], phi_x: [N, C, HxW]
+        # NonLocal3d theta_x: [N, TxHxW, C], phi_x: [N, C, TxHxW]
+        if self.mode == 'gaussian':
+            theta_x = x.view(n, self.in_channels, -1)
+            theta_x = theta_x.permute(0, 2, 1)
+            if self.sub_sample:
+                phi_x = self.phi(x).view(n, self.in_channels, -1)
+            else:
+                phi_x = x.view(n, self.in_channels, -1)
+        elif self.mode == 'concatenation':
+            theta_x = self.theta(x).view(n, self.inter_channels, -1, 1)
+            phi_x = self.phi(x).view(n, self.inter_channels, 1, -1)
+        else:
+            theta_x = self.theta(x).view(n, self.inter_channels, -1)
+            theta_x = theta_x.permute(0, 2, 1)
+            phi_x = self.phi(x).view(n, self.inter_channels, -1)
 
         pairwise_func = getattr(self, self.mode)
         # NonLocal1d pairwise_weight: [N, H, H]
@@ -183,7 +235,10 @@ class NonLocal1d(_NonLocalNd):
         if sub_sample:
             max_pool_layer = nn.MaxPool1d(kernel_size=2)
             self.g = nn.Sequential(self.g, max_pool_layer)
-            self.phi = nn.Sequential(self.phi, max_pool_layer)
+            if self.mode != 'gaussian':
+                self.phi = nn.Sequential(self.phi, max_pool_layer)
+            else:
+                self.phi = max_pool_layer
 
 
 @PLUGIN_LAYERS.register_module()
@@ -214,7 +269,10 @@ class NonLocal2d(_NonLocalNd):
         if sub_sample:
             max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
             self.g = nn.Sequential(self.g, max_pool_layer)
-            self.phi = nn.Sequential(self.phi, max_pool_layer)
+            if self.mode != 'gaussian':
+                self.phi = nn.Sequential(self.phi, max_pool_layer)
+            else:
+                self.phi = max_pool_layer
 
 
 class NonLocal3d(_NonLocalNd):
@@ -241,4 +299,7 @@ class NonLocal3d(_NonLocalNd):
         if sub_sample:
             max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
             self.g = nn.Sequential(self.g, max_pool_layer)
-            self.phi = nn.Sequential(self.phi, max_pool_layer)
+            if self.mode != 'gaussian':
+                self.phi = nn.Sequential(self.phi, max_pool_layer)
+            else:
+                self.phi = max_pool_layer

tests/test_cnn/test_non_local.py

@@ -11,12 +11,17 @@ def test_nonlocal():
         # mode should be in ['embedded_gaussian', 'dot_product']
         _NonLocalNd(3, mode='unsupport_mode')
 
-    # _NonLocalNd
+    # _NonLocalNd with zero initialization
+    _NonLocalNd(3)
     _NonLocalNd(3, norm_cfg=dict(type='BN'))
-    # Not Zero initialization
-    _NonLocalNd(3, norm_cfg=dict(type='BN'), zeros_init=True)
 
-    # NonLocal3d
+    # _NonLocalNd without zero initialization
+    _NonLocalNd(3, zeros_init=False)
+    _NonLocalNd(3, norm_cfg=dict(type='BN'), zeros_init=False)
+
+
+def test_nonlocal3d():
+    # NonLocal3d with 'embedded_gaussian' mode
     imgs = torch.randn(2, 3, 10, 20, 20)
     nonlocal_3d = NonLocal3d(3)
     if torch.__version__ == 'parrots':
@@ -27,6 +32,7 @@ def test_nonlocal():
     out = nonlocal_3d(imgs)
     assert out.shape == imgs.shape
 
+    # NonLocal3d with 'dot_product' mode
     nonlocal_3d = NonLocal3d(3, mode='dot_product')
     assert nonlocal_3d.mode == 'dot_product'
     if torch.__version__ == 'parrots':
@@ -35,6 +41,38 @@ def test_nonlocal():
     out = nonlocal_3d(imgs)
     assert out.shape == imgs.shape
 
+    # NonLocal3d with 'concatenation' mode
+    nonlocal_3d = NonLocal3d(3, mode='concatenation')
+    assert nonlocal_3d.mode == 'concatenation'
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            nonlocal_3d.cuda()
+    out = nonlocal_3d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal3d with 'gaussian' mode
+    nonlocal_3d = NonLocal3d(3, mode='gaussian')
+    assert not hasattr(nonlocal_3d, 'phi')
+    assert nonlocal_3d.mode == 'gaussian'
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            nonlocal_3d.cuda()
+    out = nonlocal_3d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal3d with 'gaussian' mode and sub_sample
+    nonlocal_3d = NonLocal3d(3, mode='gaussian', sub_sample=True)
+    assert isinstance(nonlocal_3d.g, nn.Sequential) and len(nonlocal_3d.g) == 2
+    assert isinstance(nonlocal_3d.g[1], nn.MaxPool3d)
+    assert nonlocal_3d.g[1].kernel_size == (1, 2, 2)
+    assert isinstance(nonlocal_3d.phi, nn.MaxPool3d)
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            nonlocal_3d.cuda()
+    out = nonlocal_3d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal3d with 'dot_product' mode and sub_sample
     nonlocal_3d = NonLocal3d(3, mode='dot_product', sub_sample=True)
     for m in [nonlocal_3d.g, nonlocal_3d.phi]:
         assert isinstance(m, nn.Sequential) and len(m) == 2
@@ -46,7 +84,9 @@ def test_nonlocal():
     out = nonlocal_3d(imgs)
     assert out.shape == imgs.shape
 
-    # NonLocal2d
+
+def test_nonlocal2d():
+    # NonLocal2d with 'embedded_gaussian' mode
     imgs = torch.randn(2, 3, 20, 20)
     nonlocal_2d = NonLocal2d(3)
     if torch.__version__ == 'parrots':
@@ -56,6 +96,50 @@ def test_nonlocal():
     out = nonlocal_2d(imgs)
     assert out.shape == imgs.shape
 
+    # NonLocal2d with 'dot_product' mode
+    imgs = torch.randn(2, 3, 20, 20)
+    nonlocal_2d = NonLocal2d(3, mode='dot_product')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_2d.cuda()
+    out = nonlocal_2d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal2d with 'concatenation' mode
+    imgs = torch.randn(2, 3, 20, 20)
+    nonlocal_2d = NonLocal2d(3, mode='concatenation')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_2d.cuda()
+    out = nonlocal_2d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal2d with 'gaussian' mode
+    imgs = torch.randn(2, 3, 20, 20)
+    nonlocal_2d = NonLocal2d(3, mode='gaussian')
+    assert not hasattr(nonlocal_2d, 'phi')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_2d.cuda()
+    out = nonlocal_2d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal2d with 'gaussian' mode and sub_sample
+    nonlocal_2d = NonLocal2d(3, mode='gaussian', sub_sample=True)
+    assert isinstance(nonlocal_2d.g, nn.Sequential) and len(nonlocal_2d.g) == 2
+    assert isinstance(nonlocal_2d.g[1], nn.MaxPool2d)
+    assert nonlocal_2d.g[1].kernel_size == (2, 2)
+    assert isinstance(nonlocal_2d.phi, nn.MaxPool2d)
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            nonlocal_2d.cuda()
+    out = nonlocal_2d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal2d with 'dot_product' mode and sub_sample
     nonlocal_2d = NonLocal2d(3, mode='dot_product', sub_sample=True)
     for m in [nonlocal_2d.g, nonlocal_2d.phi]:
         assert isinstance(m, nn.Sequential) and len(m) == 2
@@ -67,7 +151,9 @@ def test_nonlocal():
     out = nonlocal_2d(imgs)
     assert out.shape == imgs.shape
 
-    # NonLocal1d
+
+def test_nonlocal1d():
+    # NonLocal1d with 'embedded_gaussian' mode
     imgs = torch.randn(2, 3, 20)
     nonlocal_1d = NonLocal1d(3)
     if torch.__version__ == 'parrots':
@@ -77,6 +163,50 @@ def test_nonlocal():
     out = nonlocal_1d(imgs)
     assert out.shape == imgs.shape
 
+    # NonLocal1d with 'dot_product' mode
+    imgs = torch.randn(2, 3, 20)
+    nonlocal_1d = NonLocal1d(3, mode='dot_product')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_1d.cuda()
+    out = nonlocal_1d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal1d with 'concatenation' mode
+    imgs = torch.randn(2, 3, 20)
+    nonlocal_1d = NonLocal1d(3, mode='concatenation')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_1d.cuda()
+    out = nonlocal_1d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal1d with 'gaussian' mode
+    imgs = torch.randn(2, 3, 20)
+    nonlocal_1d = NonLocal1d(3, mode='gaussian')
+    assert not hasattr(nonlocal_1d, 'phi')
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            imgs = imgs.cuda()
+            nonlocal_1d.cuda()
+    out = nonlocal_1d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal1d with 'gaussian' mode and sub_sample
+    nonlocal_1d = NonLocal1d(3, mode='gaussian', sub_sample=True)
+    assert isinstance(nonlocal_1d.g, nn.Sequential) and len(nonlocal_1d.g) == 2
+    assert isinstance(nonlocal_1d.g[1], nn.MaxPool1d)
+    assert nonlocal_1d.g[1].kernel_size == 2
+    assert isinstance(nonlocal_1d.phi, nn.MaxPool1d)
+    if torch.__version__ == 'parrots':
+        if torch.cuda.is_available():
+            nonlocal_1d.cuda()
+    out = nonlocal_1d(imgs)
+    assert out.shape == imgs.shape
+
+    # NonLocal1d with 'dot_product' mode and sub_sample
     nonlocal_1d = NonLocal1d(3, mode='dot_product', sub_sample=True)
     for m in [nonlocal_1d.g, nonlocal_1d.phi]:
         assert isinstance(m, nn.Sequential) and len(m) == 2