Added a test in test_NetworkCompare to verify cudnn and exconv, fixed a bug in...

Added a test in test_NetworkCompare to verify cudnn and exconv, fixed a bug in cudnn_conv in dealing with groups

paddle/gserver/layers/ConvProjection.cpp

@@ -59,7 +59,8 @@ void ConvProjection::getConvParams() {
 
 void ConvProjection::initCudnn() {
   hl_create_filter_descriptor(
-      &filterDesc_, channels_, numFilters_, filterH_, filterW_);
+      &filterDesc_, channels_ / groups_, numFilters_ / groups_,
+      filterH_, filterW_);
   hl_create_tensor_descriptor(&inputDesc_);
   hl_create_tensor_descriptor(&outputDesc_);
   hl_create_convolution_descriptor(&convDesc_,
@@ -86,7 +87,7 @@ void ConvProjection::initCudnn() {
 void ConvProjection::reshapeTensorDesc(int batchSize) {
   hl_tensor_reshape(inputDesc_,
                     batchSize,
-                    channels_,
+                    channels_ / groups_,
                     imageH_,
                     imageW_,
                     channels_ * imageH_ * imageW_,
@@ -115,7 +116,7 @@ void ConvProjection::reshapeTensorDesc(int batchSize) {
 
   hl_tensor_reshape(outputDesc_,
                     batchSize,
-                    numFilters_,
+                    numFilters_ / groups_,
                     outputH_,
                     outputW_,
                     nStride,

paddle/gserver/tests/img_conv2_a.conf

+#edit-mode: -*- python -*-
+# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=10)
+data = data_layer(name ="input", size=8*16*16)
+
+conv = img_conv_layer(input=data, filter_size=2, filter_size_y=2,
+                      num_channels=8,
+                      num_filters=16, stride=1,
+                      bias_attr=True,
+                      act=LinearActivation(),
+                      layer_type="exconv")
+
+conv2 = img_conv_layer(input=data, filter_size=3, filter_size_y=3,
+                      num_channels=8,
+                      num_filters=16, stride=2,
+                      bias_attr=True,
+                      act=LinearActivation(),
+                      groups=1,
+                      layer_type="exconv")
+                      
+outputs(conv, conv2)

paddle/gserver/tests/img_conv2_b.conf

+#edit-mode: -*- python -*-
+# Copyright (c) 2016 Baidu, Inc. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=10)
+data = data_layer(name ="input", size=8*16*16)
+
+conv = img_conv_layer(input=data, filter_size=2, filter_size_y=2,
+                      num_channels=8,
+                      num_filters=16, stride=1,
+                      bias_attr=True,
+                      act=LinearActivation(),
+                      layer_type="cudnn_conv")
+
+conv2 = img_conv_layer(input=data, filter_size=3, filter_size_y=3,
+                      num_channels=8,
+                      num_filters=16, stride=2,
+                      bias_attr=True,
+                      act=LinearActivation(),
+                      groups=1,
+                      layer_type="cudnn_conv")
+                      
+outputs(conv, conv2)

paddle/gserver/tests/test_ConvUnify.cpp

@@ -38,7 +38,7 @@ P_DECLARE_bool(prev_batch_state);
 // matches the given result
 MatrixPtr doOneConvTest(size_t imgSize, size_t output_x, size_t stride,
                     size_t padding, size_t filter_size, size_t channel,
-                    size_t numfilters, MatrixPtr& inputData,
+                    size_t numfilters, size_t groups, MatrixPtr& inputData,
                     real* param, bool useGpu) {
   TestConfig config;
   config.biasSize = numfilters;
@@ -51,9 +51,11 @@ MatrixPtr doOneConvTest(size_t imgSize, size_t output_x, size_t stride,
   config.layerConfig.set_partial_sum(1);
   config.layerConfig.set_shared_biases(true);
 
+  size_t weightSize = channel* filter_size * filter_size *
+      config.layerConfig.num_filters() / groups;
   config.inputDefs.push_back({INPUT_DATA, "layer_0",
                               imgSize * imgSize * channel,
-      channel* filter_size * filter_size * config.layerConfig.num_filters()});
+                              weightSize});
   LayerInputConfig* input = config.layerConfig.add_inputs();
   ConvConfig* conv = input->mutable_conv_conf();
   conv->set_filter_size(filter_size);
@@ -63,8 +65,8 @@ MatrixPtr doOneConvTest(size_t imgSize, size_t output_x, size_t stride,
   conv->set_padding_y(padding);
   conv->set_stride(stride);
   conv->set_stride_y(stride);
-  conv->set_groups(1);
-  conv->set_filter_channels(channel);
+  conv->set_groups(groups);
+  conv->set_filter_channels(channel/groups);
   conv->set_img_size(imgSize);
   conv->set_output_x(output_x);
 
@@ -87,7 +89,7 @@ MatrixPtr doOneConvTest(size_t imgSize, size_t output_x, size_t stride,
   convLayer->getBiasParameter()->zeroMem();
   convLayer->getParameters()[0]->zeroMem();
   convLayer->getParameters()[0]->getBuf(PARAMETER_VALUE)->copyFrom(param,
-      channel* filter_size * filter_size * config.layerConfig.num_filters());
+      weightSize);
   convLayer->forward(PASS_GC);
 
   return convLayer->getOutputValue();
@@ -113,6 +115,7 @@ TEST(Layer, convParaUnified) {
                    /* filter_size */ 3,
                    /*channel*/ 1,
                    /*numfilters*/ 2,
+                   /*groups*/ 1,
                    input, param, false);
 
     resultGpu = doOneConvTest(/* imgSize */ 4,
@@ -122,6 +125,7 @@ TEST(Layer, convParaUnified) {
                        /* filter_size */ 3,
                        /*channel*/ 1,
                        /*numfilters*/ 2,
+                       /*groups*/ 1,
                        input, param, true);
     checkMatrixEqual(resultCpu, resultGpu);
 
@@ -145,6 +149,7 @@ TEST(Layer, convParaUnified) {
                    /* filter_size */ 2,
                    /*channel*/ 2,
                    /*numfilters*/ 2,
+                   /*groups*/ 1,
                    input, param2, false);
 
     resultGpu = doOneConvTest(/* imgSize */ 3,
@@ -154,8 +159,34 @@ TEST(Layer, convParaUnified) {
                        /* filter_size */ 2,
                        /*channel*/ 2,
                        /*numfilters*/ 2,
+                       /*groups*/ 1,
                        input, param2, true);
     checkMatrixEqual(resultCpu, resultGpu);
+
+
+    float param3[] = {1, 2, 3, 4,
+                      4, 3, 2, 1};
+
+    resultCpu = doOneConvTest(/* imgSize */ 3,
+                   /* output_x */ 2,
+                   /* stride */ 1,
+                   /* padding */ 0,
+                   /* filter_size */ 2,
+                   /*channel*/ 2,
+                   /*numfilters*/ 2,
+                   /*groups*/ 2,
+                   input, param3, false);
+
+    resultGpu = doOneConvTest(/* imgSize */ 3,
+                       /* output_x */ 2,
+                       /* stride */ 1,
+                       /* padding */ 0,
+                       /* filter_size */ 2,
+                       /*channel*/ 2,
+                       /*numfilters*/ 2,
+                       /*groups*/ 2,
+                       input, param3, true);
+    checkMatrixEqual(resultCpu, resultGpu);
   #endif
 }
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -213,6 +213,53 @@ TEST(Projection, conv) {
       NUM_FILTERS,
       true);
 }
+
+TEST(Projection, conv2) {
+  const int NUM_FILTERS = 18;
+  const int FILTER_SIZE = 2;
+  const int FILTER_SIZE_Y = 3;
+  const int CHANNELS = 3;
+  const int IMAGE_SIZE = 16;
+
+  ProjectionConfig conf;
+  conf.set_type("conv");
+  conf.set_num_filters(NUM_FILTERS);
+
+  ConvConfig* conv = conf.mutable_conv_conf();
+  conv->set_filter_size(FILTER_SIZE);
+  conv->set_filter_size_y(FILTER_SIZE_Y);
+  conv->set_channels(CHANNELS);
+  conv->set_padding(0);
+  conv->set_padding_y(1);
+  conv->set_stride(2);
+  conv->set_stride_y(2);
+  conv->set_groups(3);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  conv->set_img_size(IMAGE_SIZE);
+  int output_x = outputSize(conv->img_size(),
+                            conv->filter_size(),
+                            conv->padding(),
+                            conv->stride(),
+                            /* caffeMode */ true);
+  int output_y = outputSize(conv->img_size(),
+                            conv->filter_size_y(),
+                            conv->padding_y(),
+                            conv->stride_y(),
+                            /* caffeMode */ true);
+  conv->set_output_x(output_x);
+  conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS);
+  conf.set_output_size(output_x * output_y * NUM_FILTERS);
+
+  testProjectionGrad(
+      conf,
+      INPUT_DATA,
+      /* parameterSize */ NUM_FILTERS * FILTER_SIZE * FILTER_SIZE_Y,
+      /* batchSize */ 100,
+      true,
+      false,
+      NUM_FILTERS,
+      true);
+}
 #endif
 
 TEST(Layer, BilinearInterpLayer) {

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -255,6 +255,15 @@ TEST(Compare, img_conv) {
   compareNetwork(config_file_a, config_file_b);
   FLAGS_use_gpu = useGpu;
 }
+
+TEST(Compare, img_conv2) {
+  std::string config_file_a = "./gserver/tests/img_conv2_a.conf";
+  std::string config_file_b = "./gserver/tests/img_conv2_b.conf";
+  bool useGpu = FLAGS_use_gpu;
+  FLAGS_use_gpu = true;
+  compareNetwork(config_file_a, config_file_b);
+  FLAGS_use_gpu = useGpu;
+}
 #endif
 
 P_DEFINE_string(config_file_a, "", "config of one network to compare");

python/paddle/trainer_config_helpers/layers.py

@@ -129,6 +129,9 @@ class LayerType(object):
     HSIGMOID = 'hsigmoid'
     CONV_LAYER = "conv"
     CONVTRANS_LAYER = "convt"
+    EXCONV_LAYER = "exconv"
+    EXCONVTRANS_LAYER = "exconvt"
+    CUDNNCONV_LAYER = "cudnn_conv"
     POOL_LAYER = "pool"
     BATCH_NORM_LAYER = 'batch_norm'
     NORM_LAYER = 'norm'
@@ -1762,7 +1765,8 @@ def img_conv_layer(input,
                    filter_size_y=None,
                    stride_y=None,
                    padding_y=None,
-                   trans=False):
+                   trans=False,
+                   layer_type=None):
     """
     Convolution layer for image. Paddle only support square input currently and
     thus input image's width equals height.
@@ -1829,6 +1833,10 @@ def img_conv_layer(input,
     :type layer_attr: ExtraLayerAttribute
     :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"
+    :type layer_type: String
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1864,8 +1872,15 @@ def img_conv_layer(input,
         param_attr.attr["initial_std"] = init_w
         param_attr.attr["initial_strategy"] = 0
         param_attr.attr["initial_smart"] = False
-
-    lt = LayerType.CONVTRANS_LAYER if trans else LayerType.CONV_LAYER
+    
+    if layer_type:
+        if trans:
+            assert layer_type in ["exconvt"]
+        else:
+            assert layer_type in ["exconv", "cudnn_conv"]
+        lt = layer_type
+    else:
+        lt = LayerType.CONVTRANS_LAYER if trans else LayerType.CONV_LAYER
 
     l = Layer(
         name=name,