Merge branch 'develop' of https://github.com/baidu/Paddle into inference

Author: hedaoyuan

.travis.yml

@@ -30,6 +30,7 @@ addons:
       - automake
       - libtool
       - ccache
+  ssh_known_hosts: 52.76.173.135
 before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
@@ -42,6 +43,14 @@ script:
   - |
     timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
     RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
+  - |
+    if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
+    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
+    if [[ "$TRAVIS_BRANCH" != "develop"  && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
+    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
+    export DOCS_DIR=`pwd`
+    cd ..
+    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc   
 notifications:
   email:
     on_success: change

CMakeLists.txt

@@ -126,7 +126,7 @@ include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
-include(external/pybind11)    # download pybind11
+include(external/pybind11)  # download pybind11
 include(external/nccl)
 
 include(cudnn)              # set cudnn libraries, must before configure

benchmark/IntelOptimizedPaddle.md

@@ -0,0 +1,48 @@
+# Benchmark
+
+Machine:
+
+- Server
+ 	- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
+- Laptop
+ 	- DELL XPS15-9560-R1745: i7-7700HQ 8G 256GSSD
+ 	- i5 MacBook Pro (Retina, 13-inch, Early 2015)
+- Desktop
+ 	- i7-6700k
+
+System: CentOS release 6.3 (Final), Docker 1.12.1.
+
+PaddlePaddle: paddlepaddle/paddle:latest (TODO: will rerun after 0.11.0)
+
+- MKL-DNN tag v0.10
+- MKLML 2018.0.20170720
+- OpenBLAS v0.2.20
+	 
+On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
+
+## Benchmark Model
+
+### Server
+Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
+
+Input image size - 3 * 224 * 224, Time: images/second
+
+- VGG-19
+
+| BatchSize    | 64    | 128  | 256     |
+|--------------|-------| -----| --------|
+| OpenBLAS     | 7.82  | 8.62  | 10.34  | 
+| MKLML        | 11.02 | 12.86 | 15.33  |
+| MKL-DNN      | 27.69 | 28.8 | 29.27  |
+
+
+chart on batch size 128
+TBD
+
+ - ResNet
+ - GoogLeNet
+
+### Laptop
+TBD
+### Desktop
+TBD

benchmark/paddle/image/resnet.py

@@ -0,0 +1,213 @@
+#!/usr/bin/env python
+from paddle.trainer_config_helpers import *
+
+height = 224
+width = 224
+num_class = 1000
+batch_size = get_config_arg('batch_size', int, 64)
+layer_num = get_config_arg("layer_num", int, 50)
+is_test = get_config_arg("is_test", bool, False)
+
+args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+define_py_data_sources2(
+    "train.list", None, module="provider", obj="process", args=args)
+
+settings(
+    batch_size=batch_size,
+    learning_rate=0.01 / batch_size,
+    learning_method=MomentumOptimizer(0.9),
+    regularization=L2Regularization(0.0005 * batch_size))
+
+
+#######################Network Configuration #############
+def conv_bn_layer(name,
+                  input,
+                  filter_size,
+                  num_filters,
+                  stride,
+                  padding,
+                  channels=None,
+                  active_type=ReluActivation()):
+    """
+    A wrapper for conv layer with batch normalization layers.
+    Note:
+    conv layer has no activation.
+    """
+
+    tmp = img_conv_layer(
+        name=name + "_conv",
+        input=input,
+        filter_size=filter_size,
+        num_channels=channels,
+        num_filters=num_filters,
+        stride=stride,
+        padding=padding,
+        act=LinearActivation(),
+        bias_attr=False)
+    return batch_norm_layer(
+        name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test)
+
+
+def bottleneck_block(name, input, num_filters1, num_filters2):
+    """
+    A wrapper for bottlenect building block in ResNet.
+    Last conv_bn_layer has no activation.
+    Addto layer has activation of relu.
+    """
+    last_name = conv_bn_layer(
+        name=name + '_branch2a',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters1,
+        stride=1,
+        padding=0)
+    last_name = conv_bn_layer(
+        name=name + '_branch2b',
+        input=last_name,
+        filter_size=3,
+        num_filters=num_filters1,
+        stride=1,
+        padding=1)
+    last_name = conv_bn_layer(
+        name=name + '_branch2c',
+        input=last_name,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=1,
+        padding=0,
+        active_type=LinearActivation())
+
+    return addto_layer(
+        name=name + "_addto", input=[input, last_name], act=ReluActivation())
+
+
+def mid_projection(name, input, num_filters1, num_filters2, stride=2):
+    """
+    A wrapper for middile projection in ResNet.
+    projection shortcuts are used for increasing dimensions,
+    and other shortcuts are identity
+    branch1: projection shortcuts are used for increasing
+    dimensions, has no activation.
+    branch2x: bottleneck building block, shortcuts are identity.
+    """
+    # stride = 2
+    branch1 = conv_bn_layer(
+        name=name + '_branch1',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=stride,
+        padding=0,
+        active_type=LinearActivation())
+
+    last_name = conv_bn_layer(
+        name=name + '_branch2a',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters1,
+        stride=stride,
+        padding=0)
+    last_name = conv_bn_layer(
+        name=name + '_branch2b',
+        input=last_name,
+        filter_size=3,
+        num_filters=num_filters1,
+        stride=1,
+        padding=1)
+
+    last_name = conv_bn_layer(
+        name=name + '_branch2c',
+        input=last_name,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=1,
+        padding=0,
+        active_type=LinearActivation())
+
+    return addto_layer(
+        name=name + "_addto", input=[branch1, last_name], act=ReluActivation())
+
+
+img = data_layer(name='image', size=height * width * 3)
+
+
+def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3):
+    """
+    A wrapper for 50,101,152 layers of ResNet.
+    res2_num: number of blocks stacked in conv2_x
+    res3_num: number of blocks stacked in conv3_x
+    res4_num: number of blocks stacked in conv4_x
+    res5_num: number of blocks stacked in conv5_x
+    """
+    # For ImageNet
+    # conv1: 112x112
+    tmp = conv_bn_layer(
+        "conv1",
+        input=img,
+        filter_size=7,
+        channels=3,
+        num_filters=64,
+        stride=2,
+        padding=3)
+    tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2)
+
+    # conv2_x: 56x56
+    tmp = mid_projection(
+        name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1)
+    for i in xrange(2, res2_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256)
+
+    # conv3_x: 28x28
+    tmp = mid_projection(
+        name="res3_1", input=tmp, num_filters1=128, num_filters2=512)
+    for i in xrange(2, res3_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res3_" + str(i),
+            input=tmp,
+            num_filters1=128,
+            num_filters2=512)
+
+    # conv4_x: 14x14
+    tmp = mid_projection(
+        name="res4_1", input=tmp, num_filters1=256, num_filters2=1024)
+    for i in xrange(2, res4_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res4_" + str(i),
+            input=tmp,
+            num_filters1=256,
+            num_filters2=1024)
+
+    # conv5_x: 7x7
+    tmp = mid_projection(
+        name="res5_1", input=tmp, num_filters1=512, num_filters2=2048)
+    for i in xrange(2, res5_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res5_" + str(i),
+            input=tmp,
+            num_filters1=512,
+            num_filters2=2048)
+
+    tmp = img_pool_layer(
+        name='avgpool',
+        input=tmp,
+        pool_size=7,
+        stride=1,
+        pool_type=AvgPooling())
+
+    return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
+
+
+if layer_num == 50:
+    resnet = deep_res_net(3, 4, 6, 3)
+elif layer_num == 101:
+    resnet = deep_res_net(3, 4, 23, 3)
+elif layer_num == 152:
+    resnet = deep_res_net(3, 8, 36, 3)
+else:
+    print("Wrong layer number.")
+
+lbl = data_layer(name="label", size=num_class)
+loss = cross_entropy(name='loss', input=resnet, label=lbl)
+inputs(img, lbl)
+outputs(loss)

benchmark/paddle/image/run_mkldnn.sh

@@ -5,22 +5,23 @@ function train() {
   export OMP_DYNAMIC="FALSE"
   export KMP_AFFINITY="granularity=fine,compact,0,0"
   topology=$1
-  bs=$2
-  use_mkldnn=$3
-  if [ $3 == "True" ]; then
+  layer_num=$2
+  bs=$3
+  use_mkldnn=$4
+  if [ $4 == "True" ]; then
     thread=1
-    log="logs/${topology}-mkldnn-${bs}.log"
-  elif [ $3 == "False" ]; then
+    log="logs/${topology}-${layer_num}-mkldnn-${bs}.log"
+  elif [ $4 == "False" ]; then
     thread=`nproc`
     # each trainer_count use only 1 core to avoid conflict
     export OMP_NUM_THREADS=1
     export MKL_NUM_THREADS=1
-    log="logs/${topology}-${thread}mklml-${bs}.log"
+    log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
   else
     echo "Wrong input $3, use True or False."
     exit 0
   fi
-  args="batch_size=${bs}"
+  args="batch_size=${bs},layer_num=${layer_num}"
   config="${topology}.py"
   paddle train --job=time \
     --config=$config \
@@ -40,12 +41,9 @@ if [ ! -d "logs" ]; then
   mkdir logs
 fi
 
-#========== mkldnn ==========#
-train vgg 64 True
-train vgg 128 True
-train vgg 256 True
-
-#========== mklml ===========#
-train vgg 64 False
-train vgg 128 False
-train vgg 256 False
+for use_mkldnn in True False; do
+  for batchsize in 64 128 256; do
+    train vgg 19 $batchsize $use_mkldnn
+    train resnet 50  $batchsize $use_mkldnn
+  done
+done

benchmark/paddle/image/vgg.py

@@ -13,7 +13,7 @@
 
 settings(
     batch_size=batch_size,
-    learning_rate=0.01 / batch_size,
+    learning_rate=0.001 / batch_size,
     learning_method=MomentumOptimizer(0.9),
     regularization=L2Regularization(0.0005 * batch_size))