completed implementation of cudnn_convt convTransProjection and convTransOperator

python/paddle/trainer/config_parser.py

@@ -726,7 +726,7 @@ class ConvProjection(ConvBaseProjection):
                  **xargs):
         super(ConvProjection, self).__init__(input_layer_name, **xargs)
 
-        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
+        parse_conv(conv_conf, self.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 * \
@@ -746,7 +746,7 @@ class ConvTransProjection(ConvBaseProjection):
 
         parse_conv(
             conv_conf,
-            input_layer_name,
+            self.input_layer_name,
             self.proj_conf.conv_conf,
             num_filters,
             trans=True)
@@ -1834,8 +1834,17 @@ class ConvTransLayerBase(LayerBase):
         use_gpu = int(g_command_config_args.get("use_gpu", 0))
         parallel_nn = int(g_command_config_args.get("parallel_nn", 0))
 
-        # cudnn_convt has not been implemented so use exconvt only
-        self.layer_type = "exconvt"
+        # Automatically select cudnn_type for GPU and exconvt for CPU
+        # if set type=exconvt, but still reserve the way user specify
+        # exconvt or cudnn_convt manually.
+        if self.layer_type == "cudnn_convt":
+            config_assert(use_gpu, "cudnn_convt only support GPU")
+
+        if (use_gpu == 1 and self.layer_type != "exconvt" and
+            (parallel_nn == 0 or self.config.device > -1)):
+            self.layer_type = "cudnn_convt"
+        else:
+            self.layer_type = "exconvt"
         # need to specify layer in config
         self.config.type = self.layer_type
 
@@ -1852,10 +1861,9 @@ class ConvTransLayerBase(LayerBase):
                 trans=True)
             conv_conf = self.config.inputs[input_index].conv_conf
             psize = self.calc_parameter_size(conv_conf)
-            print("output size for %s is %d " % (name, conv_conf.output_x))
             self.create_input_parameter(input_index, psize)
-            self.set_layer_size(
-                (conv_conf.img_size**2) * self.config.num_filters)
+            self.set_cnn_layer(name, conv_conf.img_size_y, conv_conf.img_size,
+                               self.config.num_filters)
 
         psize = self.config.size
         if shared_biases:
@@ -1872,6 +1880,11 @@ class ConvTransLayer(ConvTransLayerBase):
     layer_type = 'exconvt'
 
 
+@config_layer('cudnn_convt')
+class ConvTransLayer(ConvTransLayerBase):
+    layer_type = 'cudnn_convt'
+
+
 @config_layer('norm')
 class NormLayer(LayerBase):
     def __init__(self, name, inputs, **xargs):

python/paddle/trainer_config_helpers/layers.py

@@ -2046,8 +2046,9 @@ def img_conv_layer(input,
     :param trans: true if it is a convTransLayer, false if it is a convLayer
     :type trans: bool
     :param layer_type: specify the layer_type, default is None. If trans=True,
-                       layer_type has to be "exconvt", otherwise layer_type
-                       has to be either "exconv" or "cudnn_conv"
+                       layer_type has to be "exconvt" or "cudnn_convt", 
+                       otherwise layer_type has to be either "exconv" or 
+                       "cudnn_conv"
     :type layer_type: String
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2087,7 +2088,7 @@ def img_conv_layer(input,
 
     if layer_type:
         if trans:
-            assert layer_type in ["exconvt"]
+            assert layer_type in ["exconvt", "cudnn_convt"]
         else:
             assert layer_type in ["exconv", "cudnn_conv"]
         lt = layer_type

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

@@ -33,6 +33,8 @@ layers {
   bias_parameter_name: "___conv_0__.wbias"
   num_filters: 64
   shared_biases: true
+  height: 256
+  width: 256
 }
 layers {
   name: "__batch_norm_0__"
@@ -58,6 +60,8 @@ layers {
   }
   bias_parameter_name: "___batch_norm_0__.wbias"
   moving_average_fraction: 0.9
+  height: 256
+  width: 256
 }
 layers {
   name: "__crmnorm_0__"

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

@@ -154,13 +154,38 @@ layers {
   inputs {
     input_layer_name: "img"
   }
+  inputs {
+    input_layer_name: "img"
+    proj_conf {
+      type: "conv"
+      name: "___mixed_6__.w1"
+      input_size: 1024
+      output_size: 57600
+      conv_conf {
+        filter_size: 3
+        channels: 1
+        stride: 1
+        padding: 0
+        groups: 1
+        filter_channels: 1
+        output_x: 30
+        img_size: 32
+        caffe_mode: true
+        filter_size_y: 3
+        padding_y: 0
+        stride_y: 1
+        output_y: 30
+        img_size_y: 32
+      }
+    }
+  }
   inputs {
     input_layer_name: "filter"
   }
   operator_confs {
     type: "conv"
     input_indices: 0
-    input_indices: 1
+    input_indices: 2
     input_sizes: 1024
     input_sizes: 576
     output_size: 57600
@@ -186,38 +211,110 @@ layers {
 layers {
   name: "__mixed_7__"
   type: "mixed"
+  size: 254016
+  active_type: ""
+  inputs {
+    input_layer_name: "img"
+  }
+  inputs {
+    input_layer_name: "img"
+    proj_conf {
+      type: "convt"
+      name: "___mixed_7__.w1"
+      input_size: 1024
+      output_size: 254016
+      conv_conf {
+        filter_size: 3
+        channels: 1
+        stride: 2
+        padding: 1
+        groups: 1
+        filter_channels: 64
+        output_x: 32
+        img_size: 63
+        caffe_mode: true
+        filter_size_y: 3
+        padding_y: 1
+        stride_y: 2
+        output_y: 32
+        img_size_y: 63
+      }
+    }
+  }
+  inputs {
+    input_layer_name: "filter"
+  }
+  operator_confs {
+    type: "convt"
+    input_indices: 0
+    input_indices: 2
+    input_sizes: 1024
+    input_sizes: 576
+    output_size: 254016
+    conv_conf {
+      filter_size: 3
+      channels: 1
+      stride: 2
+      padding: 1
+      groups: 1
+      filter_channels: 64
+      output_x: 32
+      img_size: 63
+      caffe_mode: true
+      filter_size_y: 3
+      padding_y: 1
+      stride_y: 2
+      output_y: 32
+      img_size_y: 63
+    }
+    num_filters: 64
+  }
+}
+layers {
+  name: "__mixed_8__"
+  type: "mixed"
   size: 100
   active_type: ""
   inputs {
     input_layer_name: "__mixed_4__"
-    input_parameter_name: "___mixed_7__.w0"
+    input_parameter_name: "___mixed_8__.w0"
     proj_conf {
       type: "fc"
-      name: "___mixed_7__.w0"
+      name: "___mixed_8__.w0"
       input_size: 300
       output_size: 100
     }
   }
   inputs {
     input_layer_name: "__mixed_5__"
-    input_parameter_name: "___mixed_7__.w1"
+    input_parameter_name: "___mixed_8__.w1"
     proj_conf {
       type: "trans_fc"
-      name: "___mixed_7__.w1"
+      name: "___mixed_8__.w1"
       input_size: 100
       output_size: 100
     }
   }
   inputs {
     input_layer_name: "__mixed_6__"
-    input_parameter_name: "___mixed_7__.w2"
+    input_parameter_name: "___mixed_8__.w2"
     proj_conf {
       type: "fc"
-      name: "___mixed_7__.w2"
+      name: "___mixed_8__.w2"
       input_size: 57600
       output_size: 100
     }
   }
+  inputs {
+    input_layer_name: "__mixed_7__"
+    input_parameter_name: "___mixed_8__.w3"
+    proj_conf {
+      type: "fc"
+      name: "___mixed_8__.w3"
+      input_size: 254016
+      output_size: 100
+    }
+  }
   drop_rate: 0.5
 }
 parameters {
@@ -281,7 +378,7 @@ parameters {
   initial_smart: true
 }
 parameters {
-  name: "___mixed_7__.w0"
+  name: "___mixed_8__.w0"
   size: 30000
   initial_mean: 0.0
   initial_std: 0.057735026919
@@ -291,7 +388,7 @@ parameters {
   initial_smart: true
 }
 parameters {
-  name: "___mixed_7__.w1"
+  name: "___mixed_8__.w1"
   size: 10000
   initial_mean: 0.0
   initial_std: 0.1
@@ -301,7 +398,7 @@ parameters {
   initial_smart: true
 }
 parameters {
-  name: "___mixed_7__.w2"
+  name: "___mixed_8__.w2"
   size: 5760000
   initial_mean: 0.0
   initial_std: 0.00416666666667
@@ -310,10 +407,20 @@ parameters {
   initial_strategy: 0
   initial_smart: true
 }
+parameters {
+  name: "___mixed_8__.w3"
+  size: 25401600
+  initial_mean: 0.0
+  initial_std: 0.00198412698413
+  dims: 254016
+  dims: 100
+  initial_strategy: 0
+  initial_smart: true
+}
 input_layer_names: "test"
 input_layer_names: "img"
 input_layer_names: "filter"
-output_layer_names: "__mixed_7__"
+output_layer_names: "__mixed_8__"
 sub_models {
   name: "root"
   layer_names: "test"
@@ -328,10 +435,11 @@ sub_models {
   layer_names: "filter"
   layer_names: "__mixed_6__"
   layer_names: "__mixed_7__"
+  layer_names: "__mixed_8__"
   input_layer_names: "test"
   input_layer_names: "img"
   input_layer_names: "filter"
-  output_layer_names: "__mixed_7__"
+  output_layer_names: "__mixed_8__"
   is_recurrent_layer_group: false
 }