Resolve conflict with upstream/develop

Author: Yi Wang

doc_cn/build/docker/build_docker_image.rst

@@ -0,0 +1,35 @@
+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/howto/build_docker_image.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
 --------

doc_cn/index.rst

@@ -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/ConvProjection.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/layers/ExpandConvBaseLayer.cpp

@@ -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/CMakeLists.txt

@@ -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_a.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_b.conf

@@ -0,0 +1,43 @@
+#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)