Resolve conflict with upstream/develop

doc_cn/build/docker/build_docker_image.rst

-构建PaddlePaddle Docker Image
-===========================
-
-PaddlePaddle的Docker Image构建源码放置在 :code:`${源码根目录}/paddle/scripts/docker/`目录下。
-该Image基于ubuntu 14.04。该目录下有两个文件，Dockerfile和build.sh。其中:
-
-*  Dockerfile是docker image的主要描述文件。描述了Docker image的构建步骤、各种参数和维护
-   人员等等。
-*  build.sh是docker image的主要构建步骤。
-
-该image的构建在docker 1.12版本测试通过, 低于docker 1.12版本的情况下并没有测试。主要由于旧版本
-的docker可能缺乏 :code:`--build-arg` 参数，从而不能在运行编译命令的时候接受参数。
-
-同时，该构建脚本充分考虑了网络不稳定的情况，对于cuda的Toolkit有断点续传和传输速度过小重启下载的
-简单优化。
-
-使用脚本构建PaddlePaddle Docker Image
--------------------------------------------
-
-该脚本的使用方法是，进入该源码目录，执行 :code:`docker build .` 命令。可以使用
- :code:`--build-arg` 传入的配置参数包括:
-
-*  LOWEST\_DL\_SPEED\: 多线程下载过程中，最低线程的下载速度(默认单位是Bytes，可以传入10K, 
-   10M，或者10G这样的单位)。如果小于这个下载速度，那么这个下载线程将会关闭。所有的下载线程关闭时，
-   下载进程会重启。
-*  WITH\_GPU\: ON or OFF。是否开启GPU功能。注意，编译PaddlePaddle的GPU版本并不需要一定在具有GPU
-   的机器上进行。但是，运行PaddlePaddle的GPU版本一定要在具有CUDA的机器上运行。
-
-简单的使用样例为\:
-
-..  code-block:: bash
-
-    cd ${源码根目录}/paddle/scripts/docker/
-    docker build --build-arg LOWEST_DL_SPEED=50K\
-                 --build-arg WITH_GPU=ON \
-                 --tag  paddle_gpu:latest .
-
-即可在本地编译出PaddlePaddle的镜像。

doc_cn/howto/build_docker_image.rst

+构建PaddlePaddle的Docker Image
+==============================
+PaddlePaddle的Docker Image构建源码放置在 ``${源码根目录}/paddle/scripts/docker/`` 目录下。该目录有三类文件：
+
+- Dockerfile：Docker Image的描述文件，包括构建步骤、各种参数和维护人员等。
+  
+  - 一共维护了12个Dockerfile，Dockerfile.m4是它们的模板。
+  - PaddlePaddle中所有的Image都基于ubuntu 14.04。
+
+- build.sh：Docker Image的构建脚本，使用方式见下一小节。
+- generate.sh：通过Dockerfile.m4模板生成不同的Dockerfile。
+
+使用脚本构建Docker Image
+------------------------
+
+进入源码目录，执行 ``docker build`` 命令，即可在本地编译出PaddlePaddle的镜像。简单的使用样例为
+
+..  code-block:: bash
+
+    cd ${源码根目录}/paddle/scripts/docker/
+    docker build --build-arg LOWEST_DL_SPEED=50K\
+                 --build-arg WITH_GPU=ON \
+                 --tag  paddle_gpu:latest .
+
+其中，``--build-arg`` 传入的配置参数包括:
+
+- LOWEST\_DL\_SPEED\: 在多线程下载过程中，设置下线线程的最低速度。
+
+  - 默认单位是Bytes，但可以传入10K、10M、或10G等这样的单位。
+  - 如果小于这个速度，那么这个线程将会关闭。当所有的线程都关闭了，那么下载进程将会重启。
+-  WITH\_GPU\: ON or OFF，是否开启GPU功能。注意，
+  - **编译** PaddlePaddle的GPU版本 **不一定** 要在具有GPU的机器上进行。
+  - **运行** PaddlePaddle的GPU版本 **一定** 要在具有GPU的机器上运行。
+
+注意：所有Image的构建在Docker 1.12版本测试通过, 低于1.12的版本并没有测试。原因是旧版本可能缺乏 ``--build-arg`` 参数，从而不能在运行编译命令的时候接受参数。

doc_cn/index.rst

@@ -16,6 +16,7 @@ PaddlePaddle文档
 --------
 * `新写Layer <../doc/dev/new_layer/index.html>`_
 * `如何贡献文档 <howto/how_to_write_docs/index.html>`_
+* `如何构建Docker Image <howto/build_docker_image.html>`_
 
 算法教程
 --------

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/layers/ExpandConvBaseLayer.cpp

@@ -145,7 +145,7 @@ void ExpandConvBaseLayer::expandFwdOnce(MatrixPtr image,
   real *expInData = expandInput_->getData();
   for (int g = 0; g < groups_[inIdx]; ++g) {
     MatrixPtr A =
-        Matrix::create(wgtData, subK, subM, true, useGpu_);  // mark transpose
+        Matrix::create(wgtData, subM, subK, false, useGpu_);  // mark transpose
     MatrixPtr B = Matrix::create(expInData, subK, subN, false, useGpu_);
     MatrixPtr C = Matrix::create(outData, subM, subN, false, useGpu_);
     C->mul(A, B, 1, 1);
@@ -182,7 +182,7 @@ void ExpandConvBaseLayer::bpropActs(MatrixPtr out,
       // create temporary matrix
       MatrixPtr C = Matrix::create(expandInData, subK, subN, false, useGpu_);
       MatrixPtr B = Matrix::create(localGradData, subM, subN, false, useGpu_);
-      MatrixPtr A = Matrix::create(wgtData, subK, subM, false, useGpu_);
+      MatrixPtr A = Matrix::create(wgtData, subM, subK, true, useGpu_);
       C->mul(A, B);  // mul
 
       // clear the temporary matrix
@@ -247,10 +247,10 @@ void ExpandConvBaseLayer::bpropWeights(MatrixPtr image,
 
     // expand-mul one-group by one
     for (int g = 0; g < groups_[inpIdx]; g++) {
-      MatrixPtr A = Matrix::create(expandInData, subK, subN, false, useGpu_);
-      MatrixPtr B = Matrix::create(gradData, subM, subN, true, useGpu_);
-      MatrixPtr C = Matrix::create(wGradData, subK, subM, false, useGpu_);
-      C->mul(A, B, 1, 1);
+      MatrixPtr A = Matrix::create(expandInData, subK, subN, true, useGpu_);
+      MatrixPtr B = Matrix::create(gradData, subM, subN, false, useGpu_);
+      MatrixPtr C = Matrix::create(wGradData, subM, subK, false, useGpu_);
+      C->mul(B, A, 1, 1);
 
       A->clear();
       B->clear();

paddle/gserver/tests/CMakeLists.txt

@@ -34,7 +34,14 @@ add_unittest_without_exec(test_ConvTrans
 
 add_test(NAME test_ConvTrans
     COMMAND test_ConvTrans)
+################# test_ConvUnify #######################
+add_unittest_without_exec(test_ConvUnify
+    test_ConvUnify.cpp
+    LayerGradUtil.cpp
+    TestUtil.cpp)
 
+add_test(NAME test_ConvUnify
+    COMMAND test_ConvUnify)
 ################## test_Evaluator #######################
 add_unittest(test_Evaluator
     test_Evaluator.cpp

paddle/gserver/tests/img_conv_a.conf

@@ -34,6 +34,7 @@ conv = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
                       num_channels=8,
                       num_filters=16, stride=1,
                       bias_attr=True,
-                      act=LinearActivation())
+                      act=LinearActivation(),
+                      groups=2)
 
 outputs(concat, conv)

paddle/gserver/tests/img_conv_b.conf

@@ -24,7 +24,7 @@ proj2 = conv_projection(input=data, filter_size=1, filter_size_y=1,
 concat = concat_layer(input=[proj1, proj2], bias_attr=False, act=ReluActivation())
 
 proj = conv_projection(input=data, filter_size=1, filter_size_y=1,
-                       num_channels=8, num_filters=16, stride=1)
+                       num_channels=8, num_filters=16, stride=1, groups=2)
 
 with mixed_layer(bias_attr=True, act=LinearActivation()) as conv:
     conv += proj

paddle/gserver/tests/img_conv_c.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)
+conv1 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                        num_channels=8,
+                        num_filters=16, stride=1,
+                        bias_attr=False,
+                        act=ReluActivation(),
+                        layer_type="exconv")
+conv2 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                       num_channels=8,
+                       num_filters=16, stride=1,
+                       bias_attr=False,
+                       act=ReluActivation(),
+                       layer_type="exconv")
+
+concat = concat_layer(input=[conv1, conv2])
+
+conv = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                      num_channels=8,
+                      num_filters=16, stride=1,
+                      bias_attr=True,
+                      act=LinearActivation(),
+                      groups=2,
+                      layer_type="exconv")
+
+outputs(concat, conv)

paddle/gserver/tests/test_ConvUnify.cpp

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+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. */
+
+#include <gtest/gtest.h>
+#include <vector>
+#include <string>
+#include "paddle/gserver/layers/DataLayer.h"
+#include "ModelConfig.pb.h"
+#include "paddle/trainer/Trainer.h"
+#include "paddle/utils/GlobalConstants.h"
+#include "paddle/gserver/layers/ExpandConvTransLayer.h"
+#include "paddle/math/MathUtils.h"
+
+#include "TestUtil.h"
+#include "LayerGradUtil.h"
+
+using namespace paddle;  // NOLINT
+using namespace std;     // NOLINT
+
+P_DECLARE_bool(use_gpu);
+P_DECLARE_int32(gpu_id);
+P_DECLARE_double(checkgrad_eps);
+P_DECLARE_bool(thread_local_rand_use_global_seed);
+P_DECLARE_bool(prev_batch_state);
+
+// Do one forward pass of convTrans layer and check to see if its output
+// 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, size_t groups, MatrixPtr& inputData,
+                    real* param, bool useGpu) {
+  TestConfig config;
+  config.biasSize = numfilters;
+  if (useGpu) {
+    config.layerConfig.set_type("cudnn_conv");
+  } else {
+    config.layerConfig.set_type("exconv");
+  }
+  config.layerConfig.set_num_filters(numfilters);
+  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,
+                              weightSize});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+  conv->set_filter_size(filter_size);
+  conv->set_filter_size_y(filter_size);
+  conv->set_channels(channel);
+  conv->set_padding(padding);
+  conv->set_padding_y(padding);
+  conv->set_stride(stride);
+  conv->set_stride_y(stride);
+  conv->set_groups(groups);
+  conv->set_filter_channels(channel/groups);
+  conv->set_img_size(imgSize);
+  conv->set_output_x(output_x);
+
+  config.layerConfig.set_size(conv->output_x() * conv->output_x() *
+                              config.layerConfig.num_filters());
+  config.layerConfig.set_name("conv");
+
+  std::vector<DataLayerPtr> dataLayers;
+  LayerMap layerMap;
+  vector<Argument> datas;
+  initDataLayer(config, &dataLayers, &datas, &layerMap, "conv",
+                1, false, useGpu);
+  dataLayers[0]->getOutputValue()->zeroMem();
+  dataLayers[0]->getOutputValue()->copyFrom(*inputData);
+
+  // test layer initialize
+  std::vector<ParameterPtr> parameters;
+  LayerPtr convLayer;
+  initTestLayer(config, &layerMap, &parameters, &convLayer);
+  convLayer->getBiasParameter()->zeroMem();
+  convLayer->getParameters()[0]->zeroMem();
+  convLayer->getParameters()[0]->getBuf(PARAMETER_VALUE)->copyFrom(param,
+      weightSize);
+  convLayer->forward(PASS_GC);
+
+  return convLayer->getOutputValue();
+}
+
+TEST(Layer, convParaUnified) {
+  #ifndef PADDLE_ONLY_CPU
+    MatrixPtr input, resultCpu, resultGpu;
+    input = Matrix::create(1, 4 * 4, false, false);
+    float inputData[] = {1, 2, 3, 4,
+                         5, 6, 7, 8,
+                         9, 10, 11, 12,
+                         13, 14, 15, 16};
+    float param[] = {1, 2, 3, 4, 5, 6, 7, 8, 9,
+                     9, 8, 7, 6, 5, 4, 3, 2, 1};
+
+    input->setData(inputData);
+
+    resultCpu = doOneConvTest(/* imgSize */ 4,
+                   /* output_x */ 2,
+                   /* stride */ 1,
+                   /* padding */ 0,
+                   /* filter_size */ 3,
+                   /*channel*/ 1,
+                   /*numfilters*/ 2,
+                   /*groups*/ 1,
+                   input, param, false);
+
+    resultGpu = doOneConvTest(/* imgSize */ 4,
+                       /* output_x */ 2,
+                       /* stride */ 1,
+                       /* padding */ 0,
+                       /* filter_size */ 3,
+                       /*channel*/ 1,
+                       /*numfilters*/ 2,
+                       /*groups*/ 1,
+                       input, param, true);
+    checkMatrixEqual(resultCpu, resultGpu);
+
+    input = Matrix::create(1, 3 * 3 * 2, false, false);
+    float inputData2[] = {1, 2, 3,
+                          4, 5, 6,
+                          7, 8, 9,
+
+                          10, 11, 12,
+                          13, 14, 15,
+                          16, 17, 18};
+    float param2[] = {1, 2, 3, 4, 5, 6, 7, 8,
+                      8, 7, 6, 5, 4, 3, 2, 1};
+
+    input->setData(inputData2);
+
+    resultCpu = doOneConvTest(/* imgSize */ 3,
+                   /* output_x */ 2,
+                   /* stride */ 1,
+                   /* padding */ 0,
+                   /* filter_size */ 2,
+                   /*channel*/ 2,
+                   /*numfilters*/ 2,
+                   /*groups*/ 1,
+                   input, param2, false);
+
+    resultGpu = doOneConvTest(/* imgSize */ 3,
+                       /* output_x */ 2,
+                       /* stride */ 1,
+                       /* padding */ 0,
+                       /* 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
+}
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  initMain(argc, argv);
+  FLAGS_thread_local_rand_use_global_seed = true;
+  srand(1);
+  return RUN_ALL_TESTS();
+}

paddle/gserver/tests/test_LayerGrad.cpp

@@ -166,9 +166,8 @@ TEST(Projection, scaling) {
   }
 }
 
-#ifndef PADDLE_ONLY_CPU
-TEST(Projection, conv) {
-  const int NUM_FILTERS = 16;
+void testProjectionConv(size_t groups) {
+  const int NUM_FILTERS = 18;
   const int FILTER_SIZE = 2;
   const int FILTER_SIZE_Y = 3;
   const int CHANNELS = 3;
@@ -186,7 +185,7 @@ TEST(Projection, conv) {
   conv->set_padding_y(1);
   conv->set_stride(2);
   conv->set_stride_y(2);
-  conv->set_groups(1);
+  conv->set_groups(groups);
   conv->set_filter_channels(conv->channels() / conv->groups());
   conv->set_img_size(IMAGE_SIZE);
   int output_x = outputSize(conv->img_size(),
@@ -206,13 +205,20 @@ TEST(Projection, conv) {
   testProjectionGrad(
       conf,
       INPUT_DATA,
-      /* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE * FILTER_SIZE_Y,
+      /* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE * FILTER_SIZE_Y
+                          / groups,
       /* batchSize */ 100,
       true,
       false,
       NUM_FILTERS,
       true);
 }
+
+#ifndef PADDLE_ONLY_CPU
+TEST(Projection, conv) {
+  testProjectionConv(1);
+  testProjectionConv(3);
+}
 #endif
 
 TEST(Layer, BilinearInterpLayer) {

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -255,6 +255,16 @@ TEST(Compare, img_conv) {
   compareNetwork(config_file_a, config_file_b);
   FLAGS_use_gpu = useGpu;
 }
+
+// Test cudnn_conv and exconv give the same result
+TEST(Compare, img_conv2) {
+  std::string config_file_a = "./gserver/tests/img_conv_a.conf";
+  std::string config_file_b = "./gserver/tests/img_conv_c.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_parser.py

@@ -698,7 +698,8 @@ class ConvProjection(Projection):
         ci = self.proj_conf.conv_conf.channels
         fh = self.proj_conf.conv_conf.filter_size
         fw = self.proj_conf.conv_conf.filter_size_y
-        return co * ci * fh * fw
+        gr = self.proj_conf.conv_conf.groups
+        return co * ci * fh * fw / gr
 
     def calc_bias_size(self):
         return self.proj_conf.num_filters

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
     """
@@ -1865,6 +1873,13 @@ def img_conv_layer(input,
         param_attr.attr["initial_strategy"] = 0
         param_attr.attr["initial_smart"] = False
     
+    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(