python interface for convTransProjection and convTransOperator

wangyang59 · wangyang59 · commit 07c1ea258f8b · 2017-03-20T16:49:03.000-07:00
diff --git a/paddle/gserver/layers/ConvBaseOperator.h b/paddle/gserver/layers/ConvBaseOperator.h
@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#pragma once
 
 #include "Operator.h"
 #include "paddle/math/MathUtils.h"
@@ -44,10 +45,8 @@ class ConvBaseOperator : public Operator {
     hl_destroy_filter_descriptor(filterDesc_);
     hl_destroy_convolution_descriptor(convDesc_);
   }
-  virtual void forward();
-  virtual void backward();
 
-private:
+protected:
   /**
    * Get convolution parameters from layer config and
    * initialize member variables.
diff --git a/paddle/gserver/layers/ConvOperator.h b/paddle/gserver/layers/ConvOperator.h
@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#pragma once
 
 #include "ConvBaseOperator.h"
 #include "paddle/math/MathUtils.h"
@@ -35,8 +36,8 @@ class ConvOperator : public ConvBaseOperator {
    * Free workspace in device and destroy cudnn tensor descriptor.
    */
   virtual ~ConvOperator() {}
-  virtual void forward();
-  virtual void backward();
+  void forward() override;
+  void backward() override;
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/ConvTransOperator.h b/paddle/gserver/layers/ConvTransOperator.h
@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#pragma once
 
 #include "ConvBaseOperator.h"
 #include "paddle/math/MathUtils.h"
@@ -35,8 +36,8 @@ class ConvTransOperator : public ConvBaseOperator {
    * Free workspace in device and destroy cudnn tensor descriptor.
    */
   virtual ~ConvTransOperator() {}
-  virtual void forward();
-  virtual void backward();
+  void forward() override;
+  void backward() override;
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -1503,16 +1503,20 @@ TEST(Layer, BatchNormalizationLayer) {
 #endif
 }
 
-TEST(Operator, conv) {
+void testConvOperator(bool isDeconv) {
   TestConfig config;
   const int NUM_FILTERS = 16;
   const int FILTER_SIZE = 2;
   const int FILTER_SIZE_Y = 3;
   const int CHANNELS = 3;
   const int IMAGE_SIZE = 16;
-  const int IMAGE_SIZE_Y = 8;
+  const int IMAGE_SIZE_Y = 9;
   OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs();
-  operatorConf.set_type("conv");
+  if (isDeconv) {
+    operatorConf.set_type("convt");
+  } else {
+    operatorConf.set_type("conv");
+  }
   ConvConfig* conv = operatorConf.mutable_conv_conf();
   operatorConf.set_num_filters(NUM_FILTERS);
   conv->set_filter_size(FILTER_SIZE);
@@ -1523,7 +1527,6 @@ TEST(Operator, conv) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_groups(1);
-  conv->set_filter_channels(conv->channels() / conv->groups());
   conv->set_img_size(IMAGE_SIZE);
   conv->set_img_size_y(IMAGE_SIZE_Y);
   conv->set_output_x(outputSize(conv->img_size(),
@@ -1536,11 +1539,22 @@ TEST(Operator, conv) {
                                 conv->padding_y(),
                                 conv->stride_y(),
                                 /*  caffeMode */ true));
-  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
-                              NUM_FILTERS);
 
-  config.inputDefs.push_back(
-      {INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0});
+  if (isDeconv) {
+    conv->set_filter_channels(NUM_FILTERS / conv->groups());
+    config.inputDefs.push_back({INPUT_DATA,
+                                "layer_0",
+                                conv->output_x() * conv->output_y() * CHANNELS,
+                                0});
+    config.layerConfig.set_size(IMAGE_SIZE * IMAGE_SIZE_Y * NUM_FILTERS);
+  } else {
+    conv->set_filter_channels(conv->channels() / conv->groups());
+    config.inputDefs.push_back(
+        {INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0});
+    config.layerConfig.set_size(conv->output_x() * conv->output_y() *
+                                NUM_FILTERS);
+  }
+
   config.inputDefs.push_back(
       {INPUT_DATA,
        "layer_1",
@@ -1552,6 +1566,11 @@ TEST(Operator, conv) {
   testOperatorGrad(config, operatorConf, 100, /*useGpu*/ true, false);
 }
 
+TEST(Operator, conv) {
+  testConvOperator(/*isDeconv*/ true);
+  testConvOperator(/*isDeconv*/ false);
+}
+
 TEST(Layer, FeatureMapExpandLayer) {
   TestConfig config;
   config.layerConfig.set_type("featmap_expand");
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
@@ -686,25 +686,17 @@ def calc_parameter_dims(self, input_size, output_size):
 
 
 @config_class
-class ConvProjection(Projection):
-    type = 'conv'
-
+class ConvBaseProjection(Projection):
     def __init__(self,
                  input_layer_name,
                  num_filters=None,
                  conv_conf=None,
                  **xargs):
-        super(ConvProjection, self).__init__(input_layer_name, **xargs)
+        super(ConvBaseProjection, self).__init__(input_layer_name, **xargs)
 
         if num_filters is not None:
             self.proj_conf.num_filters = num_filters
 
-        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
-                   num_filters)
-        self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
-                                     self.proj_conf.conv_conf.output_y * \
-                                     num_filters
-
     def calc_output_size(self, input_layer_config):
         return self.proj_conf.output_size
 
@@ -723,6 +715,46 @@ def calc_parameter_dims(self, input_size, output_size):
         return None
 
 
+@config_class
+class ConvProjection(ConvBaseProjection):
+    type = 'conv'
+
+    def __init__(self,
+                 input_layer_name,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvProjection, self).__init__(input_layer_name, **xargs)
+
+        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
+                   num_filters)
+        self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
+                                     self.proj_conf.conv_conf.output_y * \
+                                     num_filters
+
+
+@config_class
+class ConvTransProjection(ConvBaseProjection):
+    type = 'convt'
+
+    def __init__(self,
+                 input_layer_name,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvTransProjection, self).__init__(input_layer_name, **xargs)
+
+        parse_conv(
+            conv_conf,
+            input_layer_name,
+            self.proj_conf.conv_conf,
+            num_filters,
+            trans=True)
+        self.proj_conf.output_size = self.proj_conf.conv_conf.img_size_y * \
+                                     self.proj_conf.conv_conf.img_size * \
+                                     num_filters
+
+
 # Define a operator for mixed layer
 @config_class
 class Operator(Cfg):
@@ -789,6 +821,36 @@ def calc_output_size(self, input_sizes):
         return self.operator_conf.output_size
 
 
+@config_class
+class ConvTransOperator(Operator):
+    type = 'convt'
+
+    def __init__(self,
+                 input_layer_names,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvTransOperator, self).__init__(input_layer_names, **xargs)
+        if num_filters is not None:
+            self.operator_conf.num_filters = num_filters
+
+        parse_conv(
+            conv_conf,
+            MakeLayerNameInSubmodel(input_layer_names[0]),
+            self.operator_conf.conv_conf,
+            num_filters,
+            trans=True)
+        self.operator_conf.output_size = \
+            self.operator_conf.conv_conf.img_size * \
+            self.operator_conf.conv_conf.img_size_y * \
+            num_filters
+
+        config_assert(len(input_layer_names) == 2, "Conv is binary operator")
+
+    def calc_output_size(self, input_sizes):
+        return self.operator_conf.output_size
+
+
 # please refer to the comments in proto/ModelConfig.proto
 @config_class
 class Conv(Cfg):
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
@@ -712,8 +712,10 @@ def __enter__(self):
         assert len(self.inputs) == 0
         return self
 
-    def __exit__(self, *args, **kwargs):
-        del args, kwargs  # unused parameter to suppress warning
+    def __exit__(self, exc_type, exc_value, tb):
+        if exc_type != None:
+            traceback.print_exception(exc_type, exc_value, tb)
+            sys.exit(1)
         assert len(self.inputs) != 0
         ml = MixedLayer(
             name=self.name,
@@ -3715,7 +3717,8 @@ def conv_operator(img,
                   padding=0,
                   filter_size_y=None,
                   stride_y=None,
-                  padding_y=None):
+                  padding_y=None,
+                  trans=False):
     """
     Different from img_conv_layer, conv_op is an Operator, which can be used
     in mixed_layer. And conv_op takes two inputs to perform convolution.
@@ -3771,7 +3774,9 @@ def conv_operator(img,
     if filter.size is not None:
         filter.size = filter_size * filter_size_y * num_filters * num_channels
 
-    op = ConvOperator(
+    opCls = ConvTransOperator if trans else ConvOperator
+
+    op = opCls(
         input_layer_names=[img.name, filter.name],
         num_filters=num_filters,
         conv_conf=Conv(
@@ -3783,6 +3788,7 @@ def conv_operator(img,
             padding_y=padding_y,
             stride_y=stride_y,
             groups=1))
+
     op.origin = [img, filter]
     return op
 
@@ -3798,7 +3804,8 @@ def conv_projection(input,
                     stride_y=None,
                     padding_y=None,
                     groups=1,
-                    param_attr=None):
+                    param_attr=None,
+                    trans=False):
     """
     Different from img_conv_layer and conv_op, conv_projection is an Projection,
     which can be used in mixed_layer and conat_layer. It use cudnn to implement
@@ -3837,6 +3844,8 @@ def conv_projection(input,
     :type groups: int
     :param param_attr: Convolution param attribute. None means default attribute
     :type param_attr: ParameterAttribute
+    :param trans: whether it is convTrans or conv
+    :type trans: boolean
     :return: A DotMulProjection Object.
     :rtype: DotMulProjection
     """
@@ -3873,7 +3882,9 @@ def conv_projection(input,
         param_attr.attr["initial_strategy"] = 0
         param_attr.attr["initial_smart"] = False
 
-    proj = ConvProjection(
+    projCls = ConvTransProjection if trans else ConvProjection
+
+    proj = projCls(
         input_layer_name=input.name,
         num_filters=num_filters,
         conv_conf=Conv(
diff --git a/python/paddle/trainer_config_helpers/tests/configs/projections.py b/python/paddle/trainer_config_helpers/tests/configs/projections.py
@@ -34,11 +34,31 @@
 with mixed_layer() as m7:
     m7 += conv_operator(
         img=img, filter=flt, num_filters=64, num_channels=1, filter_size=3)
+    m7 += conv_projection(img, filter_size=3, num_filters=64, num_channels=1)
 
+with mixed_layer() as m8:
+    m8 += conv_operator(
+        img=img,
+        filter=flt,
+        num_filters=64,
+        num_channels=1,
+        filter_size=3,
+        stride=2,
+        padding=1,
+        trans=True)
+    m8 += conv_projection(
+        img,
+        filter_size=3,
+        num_filters=64,
+        num_channels=1,
+        stride=2,
+        padding=1,
+        trans=True)
 end = mixed_layer(
     input=[
         full_matrix_projection(input=m5),
-        trans_full_matrix_projection(input=m6), full_matrix_projection(input=m7)
+        trans_full_matrix_projection(input=m6),
+        full_matrix_projection(input=m7), full_matrix_projection(input=m8)
     ],
     size=100,
     layer_attr=ExtraAttr(
