Add dilate and bn_kwargs to Bottleneck and ResBlock

chainercv/links/connection/conv_2d_activ.py

@@ -35,21 +35,21 @@ class Conv2DActiv(chainer.Chain):
             If :obj:`None`, parameter initialization will be deferred until the
             first forward data pass at which time the size will be determined.
         out_channels (int): Number of channels of output arrays.
-        ksize (int or pair of ints): Size of filters (a.k.a. kernels).
+        ksize (int or tuple of ints): Size of filters (a.k.a. kernels).
             :obj:`ksize=k` and :obj:`ksize=(k, k)` are equivalent.
-        stride (int or pair of ints): Stride of filter applications.
+        stride (int or tuple of ints): Stride of filter applications.
             :obj:`stride=s` and :obj:`stride=(s, s)` are equivalent.
-        pad (int or pair of ints): Spatial padding width for input arrays.
+        pad (int or tuple of ints): Spatial padding width for input arrays.
             :obj:`pad=p` and :obj:`pad=(p, p)` are equivalent.
-        dilate (int or pair of ints): Dilation factor of filter applications.
+        dilate (int or tuple of ints): Dilation factor of filter applications.
             :obj:`dilate=d` and :obj:`dilate=(d, d)` are equivalent.
         nobias (bool): If :obj:`True`,
             then this link does not use the bias term.
-        initialW (4-D array): Initial weight value. If :obj:`None`, the default
+        initialW (callable): Initial weight value. If :obj:`None`, the default
             initializer is used.
             May also be a callable that takes :obj:`numpy.ndarray` or
             :obj:`cupy.ndarray` and edits its value.
-        initial_bias (1-D array): Initial bias value. If :obj:`None`, the bias
+        initial_bias (callable): Initial bias value. If :obj:`None`, the bias
             is set to 0.
             May also be a callable that takes :obj:`numpy.ndarray` or
             :obj:`cupy.ndarray` and edits its value.

chainercv/links/connection/conv_2d_bn_activ.py

@@ -49,21 +49,21 @@ class Conv2DBNActiv(chainer.Chain):
             If :obj:`None`, parameter initialization will be deferred until the
             first forward data pass at which time the size will be determined.
         out_channels (int): Number of channels of output arrays.
-        ksize (int or pair of ints): Size of filters (a.k.a. kernels).
+        ksize (int or tuple of ints): Size of filters (a.k.a. kernels).
             :obj:`ksize=k` and :obj:`ksize=(k, k)` are equivalent.
-        stride (int or pair of ints): Stride of filter applications.
+        stride (int or tuple of ints): Stride of filter applications.
             :obj:`stride=s` and :obj:`stride=(s, s)` are equivalent.
-        pad (int or pair of ints): Spatial padding width for input arrays.
+        pad (int or tuple of ints): Spatial padding width for input arrays.
             :obj:`pad=p` and :obj:`pad=(p, p)` are equivalent.
-        dilate (int or pair of ints): Dilation factor of filter applications.
+        dilate (int or tuple of ints): Dilation factor of filter applications.
             :obj:`dilate=d` and :obj:`dilate=(d, d)` are equivalent.
         nobias (bool): If :obj:`True`,
             then this link does not use the bias term.
-        initialW (4-D array): Initial weight value. If :obj:`None`, the default
+        initialW (callable): Initial weight value. If :obj:`None`, the default
             initializer is used.
             May also be a callable that takes :obj:`numpy.ndarray` or
             :obj:`cupy.ndarray` and edits its value.
-        initial_bias (1-D array): Initial bias value. If :obj:`None`, the bias
+        initial_bias (callable): Initial bias value. If :obj:`None`, the bias
             is set to 0.
             May also be a callable that takes :obj:`numpy.ndarray` or
             :obj:`cupy.ndarray` and edits its value.

chainercv/links/model/resnet/resblock.py

@@ -17,8 +17,12 @@ class ResBlock(PickableSequentialChain):
         mid_channels (int): The number of channels of intermediate arrays.
         out_channels (int): The number of channels of the output array.
         stride (int or tuple of ints): Stride of filter application.
-        initialW (4-D array): Initial weight value used in
+        dilate (int or tuple of ints): Dilation factor of filter applications.
+            :obj:`dilate=d` and :obj:`dilate=(d, d)` are equivalent.
+        initialW (callable): Initial weight value used in
             the convolutional layers.
+        bn_kwargs (dict): Keyword arguments passed to initialize
+            :class:`chainer.links.BatchNormalization`.
         stride_first (bool): This determines the behavior of the
             bottleneck with a shortcut. If :obj:`True`, apply strided
             convolution with the first convolution layer.
@@ -28,17 +32,21 @@ class ResBlock(PickableSequentialChain):
     """
 
     def __init__(self, n_layer, in_channels, mid_channels,
-                 out_channels, stride, initialW=None, stride_first=False):
+                 out_channels, stride, dilate=1, initialW=None,
+                 bn_kwargs={}, stride_first=False):
         super(ResBlock, self).__init__()
+        # Dilate option is applied to all bottlenecks.
         with self.init_scope():
             self.a = Bottleneck(
-                in_channels, mid_channels, out_channels, stride,
-                initialW, residual_conv=True, stride_first=stride_first)
+                in_channels, mid_channels, out_channels, stride, dilate,
+                initialW, bn_kwargs=bn_kwargs, residual_conv=True,
+                stride_first=stride_first)
             for i in range(n_layer - 1):
                 name = 'b{}'.format(i + 1)
                 bottleneck = Bottleneck(
                     out_channels, mid_channels, out_channels, stride=1,
-                    initialW=initialW, residual_conv=False)
+                    dilate=dilate, initialW=initialW, bn_kwargs=bn_kwargs,
+                    residual_conv=False)
                 self.add_link(name, bottleneck)
 
 
@@ -51,8 +59,12 @@ class Bottleneck(chainer.Chain):
         mid_channels (int): The number of channels of intermediate arrays.
         out_channels (int): The number of channels of the output array.
         stride (int or tuple of ints): Stride of filter application.
-        initialW (4-D array): Initial weight value used in
+        dilate (int or tuple of ints): Dilation factor of filter applications.
+            :obj:`dilate=d` and :obj:`dilate=(d, d)` are equivalent.
+        initialW (callable): Initial weight value used in
             the convolutional layers.
+        bn_kwargs (dict): Keyword arguments passed to initialize
+            :class:`chainer.links.BatchNormalization`.
         residual_conv (bool): If :obj:`True`, apply a 1x1 convolution
             to the residual.
         stride_first (bool): If :obj:`True`, apply strided convolution
@@ -62,8 +74,8 @@ class Bottleneck(chainer.Chain):
     """
 
     def __init__(self, in_channels, mid_channels, out_channels,
-                 stride=1, initialW=None, residual_conv=False,
-                 stride_first=False):
+                 stride=1, dilate=1, initialW=None, bn_kwargs={},
+                 residual_conv=False, stride_first=False):
         if stride_first:
             first_stride = stride
             second_stride = 1
@@ -72,19 +84,23 @@ def __init__(self, in_channels, mid_channels, out_channels,
             second_stride = stride
         super(Bottleneck, self).__init__()
         with self.init_scope():
-            self.conv1 = Conv2DBNActiv(in_channels, mid_channels, 1,
-                                       first_stride, 0, initialW=initialW,
-                                       nobias=True)
-            self.conv2 = Conv2DBNActiv(mid_channels, mid_channels, 3,
-                                       second_stride, 1, initialW=initialW,
-                                       nobias=True)
+            self.conv1 = Conv2DBNActiv(in_channels, mid_channels,
+                                       1, first_stride, 0,
+                                       nobias=True, initialW=initialW,
+                                       bn_kwargs=bn_kwargs)
+            # pad = dilate
+            self.conv2 = Conv2DBNActiv(mid_channels, mid_channels,
+                                       3, second_stride, dilate, dilate,
+                                       nobias=True, initialW=initialW,
+                                       bn_kwargs=bn_kwargs)
             self.conv3 = Conv2DBNActiv(mid_channels, out_channels, 1, 1, 0,
-                                       initialW=initialW, nobias=True,
-                                       activ=None)
+                                       nobias=True, initialW=initialW,
+                                       activ=None, bn_kwargs=bn_kwargs)
             if residual_conv:
                 self.residual_conv = Conv2DBNActiv(
                     in_channels, out_channels, 1, stride, 0,
-                    nobias=True, initialW=initialW, activ=None)
+                    nobias=True, initialW=initialW,
+                    activ=None, bn_kwargs=bn_kwargs)
 
     def __call__(self, x):
         h = self.conv1(x)