Merge pull request #6631 from will-am/img_separable_conv

Add separable convolution

python/paddle/trainer_config_helpers/networks.py

@@ -25,10 +25,10 @@ from paddle.trainer.config_parser import *
 __all__ = [
     'sequence_conv_pool', 'simple_lstm', "simple_img_conv_pool",
     "img_conv_bn_pool", 'lstmemory_group', 'lstmemory_unit', 'small_vgg',
-    'img_conv_group', 'vgg_16_network', 'gru_unit', 'gru_group', 'simple_gru',
-    'simple_attention', 'dot_product_attention', 'multi_head_attention',
-    'simple_gru2', 'bidirectional_gru', 'text_conv_pool', 'bidirectional_lstm',
-    'inputs', 'outputs'
+    'img_conv_group', 'img_separable_conv', 'vgg_16_network', 'gru_unit',
+    'gru_group', 'simple_gru', 'simple_attention', 'dot_product_attention',
+    'multi_head_attention', 'simple_gru2', 'bidirectional_gru',
+    'text_conv_pool', 'bidirectional_lstm', 'inputs', 'outputs'
 ]
 
 ######################################################
@@ -435,6 +435,85 @@ def img_conv_group(input,
         input=tmp, stride=pool_stride, pool_size=pool_size, pool_type=pool_type)
 
 
+@wrap_name_default("separable_conv")
+def img_separable_conv(input,
+                       num_channels,
+                       num_out_channels,
+                       filter_size,
+                       stride=1,
+                       padding=0,
+                       depth_multiplier=1,
+                       act=None,
+                       bias_attr=None,
+                       param_attr=None,
+                       shared_bias=True,
+                       layer_type='exconv',
+                       name=None):
+    """
+    Separable Convolution.
+
+    The separable convolution module is consisted of a depthwise convolution
+    that acts separately on input channels, followed by a pointwise convolution
+    with 1*1 kernels that mixes channels. It is used for Xception:
+    https://arxiv.org/pdf/1610.02357.pdf
+
+    :param input: input layer.
+    :type input: LayerOutput
+    :param num_channels: the number of input channels.
+    :type num_channels: int
+    :param num_out_channels: the number of output channels.
+    :type num_out_channels: int
+    :param filter_size: the filter size for the depthwise convolution.
+    :type filter_size: int|tuple
+    :param stride: the stride size for the depthwise convolution.
+    :type stride: int|tuple
+    :param padding: the padding size for the depthwise convolution.
+    :type padding: int|tuple
+    :param depth_multiplier: the number of filter for one channel in the
+                             depthwize convolution.
+    :type depth_multiplier: int
+    :param act: the activation function for the output.
+    :type act: BaseActivation
+    :param bias_attr: see img_conv_layer for details.
+    :type bias_attr: ParameterAttribute
+    :param param_attr: see img_conv_layer for details.
+    :type param_attr: ParameterAttribute
+    :param shared_bias: see img_conv_layer for details.
+    :type shared_bias: bool
+    :param layer_type: see img_conv_layer for details.
+    :type layer_type: bool
+    :return: layer's output
+    :rtype: LayerOutput
+    """
+    __depthwise_conv__ = img_conv_layer(
+        name="%s_depthwise_conv" % name,
+        input=input,
+        num_channels=num_channels,
+        num_filters=num_channels * depth_multiplier,
+        groups=num_channels,
+        filter_size=filter_size,
+        stride=stride,
+        padding=padding,
+        act=LinearActivation(),
+        bias_attr=bias_attr,
+        param_attr=param_attr,
+        shared_biases=shared_bias,
+        layer_type=layer_type)
+    __pointwise_conv__ = img_conv_layer(
+        name="%s_pointwise_conv" % name,
+        input=__depthwise_conv__,
+        num_channels=num_channels * depth_multiplier,
+        num_filters=num_out_channels,
+        filter_size=1,
+        stride=1,
+        padding=0,
+        act=act,
+        bias_attr=bias_attr,
+        param_attr=param_attr,
+        shared_biases=shared_bias)
+    return __pointwise_conv__
+
+
 def small_vgg(input_image, num_channels, num_classes):
     def __vgg__(ipt, num_filter, times, dropouts, num_channels_=None):
         return img_conv_group(