Support conv layers' own `init_weights` method (#278)

* Support conv layers' own `init_weights` method

* Add related unittest about ConvModule init_weight

* make the comments more specific

mmcv/cnn/bricks/conv_module.py

@@ -150,13 +150,21 @@ class ConvModule(nn.Module):
         return getattr(self, self.norm_name)
 
     def init_weights(self):
+        # 1. It is mainly for customized conv layers with their own
+        #    initialization manners, and we do not want ConvModule to
+        #    overrides the initialization.
+        # 2. For customized conv layers without their own initialization
+        #    manners, they will be initialized by this method with default
+        #    `kaiming_init`.
+        # 3. For PyTorch's conv layers, they will be initialized anyway by
+        #    their own `reset_parameters` methods.
+        if not hasattr(self.conv, 'init_weights'):
             if self.with_activation and self.act_cfg['type'] == 'LeakyReLU':
                 nonlinearity = 'leaky_relu'
                 a = self.act_cfg.get('negative_slope', 0.01)
             else:
                 nonlinearity = 'relu'
                 a = 0
-
             kaiming_init(self.conv, a=a, nonlinearity=nonlinearity)
         if self.with_norm:
             constant_init(self.norm, 1, bias=0)

tests/test_cnn/test_conv_module.py

@@ -4,7 +4,44 @@ import pytest
 import torch
 import torch.nn as nn
 
-from mmcv.cnn.bricks import ConvModule
+from mmcv.cnn.bricks import CONV_LAYERS, ConvModule
+
+
+@CONV_LAYERS.register_module()
+class ExampleConv(nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1,
+                 bias=True,
+                 norm_cfg=None):
+        super(ExampleConv, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        self.dilation = dilation
+        self.groups = groups
+        self.bias = bias
+        self.norm_cfg = norm_cfg
+        self.output_padding = (0, 0, 0)
+        self.transposed = False
+
+        self.conv0 = nn.Conv2d(in_channels, out_channels, kernel_size)
+        self.init_weights()
+
+    def forward(self, x):
+        x = self.conv0(x)
+        return x
+
+    def init_weights(self):
+        nn.init.constant_(self.conv0.weight, 0)
 
 
 def test_conv_module():
@@ -53,6 +90,11 @@ def test_conv_module():
     output = conv(x)
     assert output.shape == (1, 8, 255, 255)
 
+    # conv with its own `init_weights` method
+    conv_module = ConvModule(
+        3, 8, 2, conv_cfg=dict(type='ExampleConv'), act_cfg=None)
+    assert torch.equal(conv_module.conv.conv0.weight, torch.zeros(8, 3, 2, 2))
+
     # with_spectral_norm=True
     conv = ConvModule(3, 8, 3, padding=1, with_spectral_norm=True)
     assert hasattr(conv.conv, 'weight_orig')