Fix prelu python api

python/paddle/trainer/config_parser.py

@@ -2066,13 +2066,20 @@ class ParameterReluLayer(LayerBase):
     def __init__(self, name, inputs, partial_sum=1, **args):
         super(ParameterReluLayer, self).__init__(
             name, self.layer_type, 0, inputs=inputs, **args)
+
         input_layer = self.get_input_layer(0)
         config_assert(len(self.inputs) == 1, "prelu layer has only one input.")
         config_assert(input_layer.size % partial_sum == 0,
                       "a wrong setting for partial_sum")
+
+        dims = [1, input_layer.size / partial_sum]
         self.set_layer_size(input_layer.size)
         self.config.partial_sum = partial_sum
-        self.create_input_parameter(0, input_layer.size / partial_sum)
+        self.create_input_parameter(0, input_layer.size / partial_sum, dims)
+
+        self.set_layer_height_width(self.get_input_layer(0).height, \
+                                        self.get_input_layer(0).width)
+        self.set_layer_depth(self.get_input_layer(0).depth)
 
 
 @config_layer('conv')

python/paddle/trainer_config_helpers/layers.py

@@ -6479,10 +6479,11 @@ def row_conv_layer(input,
 
 @layer_support()
 @wrap_name_default()
-@wrap_param_attr_default()
 def prelu_layer(input,
                 name=None,
                 partial_sum=1,
+                channel_shared=None,
+                num_channels=None,
                 param_attr=None,
                 layer_attr=None):
     """
@@ -6513,6 +6514,12 @@ def prelu_layer(input,
         - partial_sum = number of outputs, indicates all elements share the same weight.
 
     :type partial_sum: int
+    :param channel_shared: whether or not the parameter are shared across channels.
+        - channel_shared = True, we set the partial_sum to the number of outputs.
+        - channel_shared = False, we set the partial_sum to the number of elements in one channel.
+    :type channel_shared: bool
+    :param num_channels: number of input channel.
+    :type num_channels: int
     :param param_attr: The parameter attribute. See ParameterAttribute for details.
     :type param_attr: ParameterAttribute
     :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
@@ -6523,7 +6530,25 @@ def prelu_layer(input,
     """
 
     assert isinstance(input, LayerOutput), 'prelu_layer accepts only one input.'
-    assert isinstance(param_attr, ParameterAttribute)
+
+    if not param_attr:
+        param_attr = ParamAttr(initial_mean=0.25, initial_std=0.0)
+    else:
+        assert isinstance(param_attr, ParameterAttribute)
+
+    if num_channels is None:
+        assert input.num_filters is not None, \
+                'the input channel cannot be detected, please specify the num_channels parameter'
+        num_channels = input.num_filters
+
+    if channel_shared is not None:
+        assert isinstance(channel_shared, bool)
+        assert (input.height != 0 and input.width != 0), \
+            'input height and widht must be setted'
+        if channel_shared:
+            partial_sum = input.height * input.width * num_channels
+        else:
+            partial_sum = input.height * input.width
 
     l = Layer(
         name=name,
@@ -6535,6 +6560,7 @@ def prelu_layer(input,
         name=name,
         layer_type=LayerType.PRELU,
         parents=input,
+        num_filters=num_channels,
         size=l.config.size)
 
 

python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr

@@ -4,6 +4,8 @@ layers {
   type: "data"
   size: 300
   active_type: ""
+  height: 10
+  width: 10
 }
 layers {
   name: "__prelu_layer_0__"
@@ -15,6 +17,9 @@ layers {
     input_parameter_name: "___prelu_layer_0__.w0"
   }
   partial_sum: 1
+  height: 10
+  width: 10
+  depth: 1
 }
 layers {
   name: "__prelu_layer_1__"
@@ -26,6 +31,9 @@ layers {
     input_parameter_name: "___prelu_layer_1__.w0"
   }
   partial_sum: 1
+  height: 10
+  width: 10
+  depth: 1
 }
 layers {
   name: "__prelu_layer_2__"
@@ -37,41 +45,100 @@ layers {
     input_parameter_name: "___prelu_layer_2__.w0"
   }
   partial_sum: 5
+  height: 10
+  width: 10
+  depth: 1
+}
+layers {
+  name: "__prelu_layer_3__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_3__.w0"
+  }
+  partial_sum: 300
+  height: 10
+  width: 10
+  depth: 1
+}
+layers {
+  name: "__prelu_layer_4__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_4__.w0"
+  }
+  partial_sum: 100
+  height: 10
+  width: 10
+  depth: 1
 }
 parameters {
   name: "___prelu_layer_0__.w0"
   size: 300
-  initial_mean: 0.0
-  initial_std: 0.057735026919
+  initial_mean: 0.25
+  initial_std: 0.0
+  dims: 1
+  dims: 300
   initial_strategy: 0
-  initial_smart: true
+  initial_smart: false
 }
 parameters {
   name: "___prelu_layer_1__.w0"
   size: 300
-  initial_mean: 0.0
-  initial_std: 0.057735026919
+  initial_mean: 0.25
+  initial_std: 0.0
+  dims: 1
+  dims: 300
   initial_strategy: 0
-  initial_smart: true
+  initial_smart: false
 }
 parameters {
   name: "___prelu_layer_2__.w0"
   size: 60
-  initial_mean: 0.0
-  initial_std: 0.129099444874
+  initial_mean: 0.25
+  initial_std: 0.0
+  dims: 1
+  dims: 60
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "___prelu_layer_3__.w0"
+  size: 1
+  initial_mean: 0.25
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "___prelu_layer_4__.w0"
+  size: 3
+  initial_mean: 0.25
+  initial_std: 0.0
+  dims: 1
+  dims: 3
   initial_strategy: 0
-  initial_smart: true
+  initial_smart: false
 }
 input_layer_names: "input"
-output_layer_names: "__prelu_layer_2__"
+output_layer_names: "__prelu_layer_4__"
 sub_models {
   name: "root"
   layer_names: "input"
   layer_names: "__prelu_layer_0__"
   layer_names: "__prelu_layer_1__"
   layer_names: "__prelu_layer_2__"
+  layer_names: "__prelu_layer_3__"
+  layer_names: "__prelu_layer_4__"
   input_layer_names: "input"
-  output_layer_names: "__prelu_layer_2__"
+  output_layer_names: "__prelu_layer_4__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py

@@ -1,8 +1,10 @@
 from paddle.trainer_config_helpers import *
 
-data = data_layer(name='input', size=300)
-prelu = prelu_layer(input=data)
-prelu = prelu_layer(input=data, partial_sum=1)
-prelu = prelu_layer(input=data, partial_sum=5)
+data = data_layer(name='input', size=300, height=10, width=10)
+prelu = prelu_layer(input=data, num_channels=3)
+prelu = prelu_layer(input=data, partial_sum=1, num_channels=3)
+prelu = prelu_layer(input=data, partial_sum=5, num_channels=3)
+prelu = prelu_layer(input=data, channel_shared=True, num_channels=3)
+prelu = prelu_layer(input=data, channel_shared=False, num_channels=3)
 
 outputs(prelu)