Merge pull request #4310 from tensor-tang/vgg

enable VGG with MKLDNN layers

benchmark/paddle/image/provider.py

@@ -22,5 +22,5 @@ def initHook(settings, height, width, color, num_class, **kwargs):
 def process(settings, file_list):
     for i in xrange(1024):
         img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten()
-        lab = random.randint(0, settings.num_class)
+        lab = random.randint(0, settings.num_class - 1)
         yield img.astype('float32'), int(lab)

benchmark/paddle/image/run_mkldnn.sh

@@ -0,0 +1,51 @@
+set -e
+
+unset OMP_NUM_THREADS MKL_NUM_THREADS
+export OMP_DYNAMIC="FALSE"
+export KMP_AFFINITY="granularity=fine,compact,0,0"
+
+function train() {
+  topology=$1
+  bs=$2
+  use_mkldnn=$3
+  if [ $3 == "True" ]; then
+    use_mkldnn=$3
+    thread=1
+    log="logs/${topology}-mkldnn-${bs}.log"
+  elif [ $3 == "False" ]; then
+    use_mkldnn=$3
+    thread=`nproc`
+    log="logs/${topology}-${thread}mklml-${bs}.log"
+  else
+    echo "Wrong input $3, use True or False."
+  fi
+  args="batch_size=${bs}"
+  config="${topology}.py"
+  paddle train --job=time \
+    --config=$config \
+    --use_mkldnn=$use_mkldnn \
+    --use_gpu=False \
+    --trainer_count=$thread \
+    --log_period=10 \
+    --test_period=100 \
+    --config_args=$args \
+    2>&1 | tee ${log} 
+}
+
+if [ ! -d "train.list" ]; then
+  echo " " > train.list
+fi
+if [ ! -d "logs" ]; then
+  mkdir logs
+fi
+
+#========= mkldnn =========#
+# vgg
+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

benchmark/paddle/image/vgg.py

@@ -0,0 +1,103 @@
+#!/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, 19)
+
+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))
+
+img = data_layer(name='image', size=height * width * 3)
+
+
+def vgg_network(vgg_num=3):
+    tmp = img_conv_group(
+        input=img,
+        num_channels=3,
+        conv_padding=1,
+        conv_num_filter=[64, 64],
+        conv_filter_size=3,
+        conv_act=ReluActivation(),
+        pool_size=2,
+        pool_stride=2,
+        pool_type=MaxPooling())
+
+    tmp = img_conv_group(
+        input=tmp,
+        conv_num_filter=[128, 128],
+        conv_padding=1,
+        conv_filter_size=3,
+        conv_act=ReluActivation(),
+        pool_stride=2,
+        pool_type=MaxPooling(),
+        pool_size=2)
+
+    channels = []
+    for i in range(vgg_num):
+        channels.append(256)
+    tmp = img_conv_group(
+        input=tmp,
+        conv_num_filter=channels,
+        conv_padding=1,
+        conv_filter_size=3,
+        conv_act=ReluActivation(),
+        pool_stride=2,
+        pool_type=MaxPooling(),
+        pool_size=2)
+    channels = []
+    for i in range(vgg_num):
+        channels.append(512)
+    tmp = img_conv_group(
+        input=tmp,
+        conv_num_filter=channels,
+        conv_padding=1,
+        conv_filter_size=3,
+        conv_act=ReluActivation(),
+        pool_stride=2,
+        pool_type=MaxPooling(),
+        pool_size=2)
+    tmp = img_conv_group(
+        input=tmp,
+        conv_num_filter=channels,
+        conv_padding=1,
+        conv_filter_size=3,
+        conv_act=ReluActivation(),
+        pool_stride=2,
+        pool_type=MaxPooling(),
+        pool_size=2)
+
+    tmp = fc_layer(
+        input=tmp,
+        size=4096,
+        act=ReluActivation(),
+        layer_attr=ExtraAttr(drop_rate=0.5))
+
+    tmp = fc_layer(
+        input=tmp,
+        size=4096,
+        act=ReluActivation(),
+        layer_attr=ExtraAttr(drop_rate=0.5))
+
+    return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
+
+
+if layer_num == 16:
+    vgg = vgg_network(3)
+elif layer_num == 19:
+    vgg = vgg_network(4)
+else:
+    print("Wrong layer number.")
+
+lab = data_layer('label', num_class)
+loss = cross_entropy(input=vgg, label=lab)
+outputs(loss)

cmake/util.cmake

@@ -97,6 +97,10 @@ function(link_paddle_exe TARGET_NAME)
         target_link_libraries(${TARGET_NAME} log)
     endif(ANDROID)
 
+    if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR)
+      target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed")
+    endif()
+
     add_dependencies(${TARGET_NAME} ${external_project_dependencies})
 endfunction()
 

paddle/gserver/activations/MKLDNNActivation.h

@@ -100,6 +100,7 @@ public:
     if (cnt_ == act.value->getElementCnt()) {
       return;
     }
+    VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
     cnt_ = act.value->getElementCnt();
     stream_.reset(new MKLDNNStream());
     auto eng = CPUEngine::Instance().getEngine();
@@ -110,7 +111,6 @@ public:
     float alpha = getAlpha();
     float beta = getBeta();
 
-    /// forward
     pipelineFwd_.clear();
     val_ = std::dynamic_pointer_cast<MKLDNNMatrix>(act.value);
     if (val_ == nullptr) {
@@ -152,6 +152,7 @@ public:
     if (!needResetBwd_) {
       return;
     }
+    VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
     needResetBwd_ = false;
     mkldnn::algorithm algo = getAlgo(this->getName());
     float alpha = getBwdAlpha();

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -64,7 +64,7 @@ bool MKLDNNConvLayer::init(const LayerMap& layerMap,
 
   // create biases
   if (biasParameter_.get() != NULL) {
-    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_, 0));
   }
   return true;
 }
@@ -251,22 +251,31 @@ void MKLDNNConvLayer::resetInValue(
   // create buffer and reorder if input value do not match
   cpuInVal_ = nullptr;
   cvtInVal_ = nullptr;
-  if (inputIsOnlyMKLDNN()) {
+
-    MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+  MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
-    CHECK(dnnIn) << "Input should be MKLDNNMatrix";
+  CHECK_EQ(inputIsOnlyMKLDNN(), dnnIn != nullptr);
-    if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+  if (dnnIn != nullptr && dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
-      CHECK_EQ(dnnIn->getFormat(), format::nc);
+    in = dnnIn;
+    return;
+  }
+  if (dnnIn) {
+    if (dnnIn->getFormat() == format::nc) {
       CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
       // create a new one with nchw format and same data
       memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
       dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
-      CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
     }
-    in = dnnIn;
+    if (dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
+      in = dnnIn;
+      return;
+    }
+    cpuInVal_ = dnnIn;
+    in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
+    cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
+    CHECK(cvtInVal_) << "should not be emptry";
   } else {
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
-    cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
+    cpuInVal_ = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
     if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
       // create new mkldnn matrix
       in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
@@ -535,7 +544,7 @@ void MKLDNNConvLayer::resetWgtValBwdData(
   } else {
     wgtValBwdData_ = wgtVal_;
   }
-  VLOG(MKLDNN_FMTS) << "weight value format for backward data"
+  VLOG(MKLDNN_FMTS) << "weight value format for backward data: "
                     << wgtValBwdData_->getFormat();
 }
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -49,7 +49,7 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
 
   // create biases
   if (biasParameter_.get() != NULL) {
-    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_, 0));
   }
   return true;
 }
@@ -161,9 +161,16 @@ void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
 
 void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
                                       MKLDNNMatrixPtr& bias) {
+  format wgtFmt = format::oihw;
+  if (inVal_->getFormat() == format::nChw8c) {
+    wgtFmt = format::oIhw8i;
+  } else if (inVal_->getFormat() == format::nChw16c) {
+    wgtFmt = format::oIhw16i;
+  }
   wgt = MKLDNNMatrix::create(
-      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
+      weight_->getW(), {oc_, ic_, ih_, iw_}, wgtFmt, engine_);
   wgt->downSpatial();
+  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
 
   bias = (biases_ && biases_->getW())
              ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)

paddle/gserver/layers/MKLDNNLayer.h

@@ -115,6 +115,7 @@ public:
       copySeqInfoToOutputs();
       size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
       if (inputElemenCnt_ != elemenCnt) {
+        VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
         // reset when input total sizes changed, not only the batchsize
         inputElemenCnt_ = elemenCnt;
         reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
@@ -142,6 +143,7 @@ public:
 
   void backward(const UpdateCallback& callback) override {
     if (needResetBwd_) {
+      VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
       resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
       needResetBwd_ = false;
     }

paddle/trainer/tests/CMakeLists.txt

@@ -37,6 +37,19 @@ add_test(NAME test_CompareTwoNets
             --config_file_a=trainer/tests/sample_trainer_config_qb_rnn.conf --config_file_b=trainer/tests/sample_trainer_config_rnn.conf
     WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
+################ test_CompareMKLDNNandCPU ######################
+if(WITH_MKLDNN)
+  add_unittest_without_exec(test_CompareMKLDNNandCPU
+      test_CompareTwoNets.cpp)
+  add_test(NAME test_CompareMKLDNNandCPU
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+          ${CMAKE_CURRENT_BINARY_DIR}/test_CompareMKLDNNandCPU
+              --config_file_a=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_a=True
+              --config_file_b=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_b=False
+              --use_gpu=False
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+endif()
+
 ############### test_CompareTwoOpts ###################
 add_unittest_without_exec(test_CompareTwoOpts
     test_CompareTwoOpts.cpp)

paddle/trainer/tests/sample_trainer_config_simple_net.conf

@@ -0,0 +1,63 @@
+# Copyright (c) 2017 PaddlePaddle Authors. 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 *
+
+################################### Data Configuration ###################################
+TrainData(ProtoData(files = "trainer/tests/mnist.list"))
+################################### Algorithm Configuration ###################################
+settings(batch_size = 1000,
+         learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
+################################### Network Configuration ###################################
+data = data_layer(name ="input", size=784)
+
+tmp = img_conv_layer(input=data,
+            num_channels=1,
+            filter_size=3,
+            num_filters=32,
+            padding=1,
+            shared_biases=True,
+            act=ReluActivation())
+
+tmp = img_pool_layer(input=tmp,
+            pool_size=3,
+            stride=2,
+            padding=1,
+            pool_type=AvgPooling())
+            
+tmp = img_conv_layer(input=tmp,
+            filter_size=3,
+            num_filters=64,
+            padding=1,
+            shared_biases=True,
+            act=ReluActivation())
+
+tmp = img_pool_layer(input=tmp,
+            pool_size=3,
+            stride=2,
+            padding=1,
+            pool_type=MaxPooling())
+            
+tmp = fc_layer(input=tmp, size=64,
+               bias_attr=True,
+               act=ReluActivation())
+
+output = fc_layer(input=tmp, size=10,
+                  bias_attr=True,
+                  act=SoftmaxActivation())
+
+lbl = data_layer(name ="label", size=10)
+
+cost = classification_cost(input=output, label=lbl)
+outputs(cost)

paddle/trainer/tests/test_CompareTwoNets.cpp

@@ -26,12 +26,15 @@ DECLARE_int32(gpu_id);
 
 DECLARE_bool(local);
 DECLARE_bool(use_gpu);
+DECLARE_bool(use_mkldnn);
 
 DECLARE_string(config);
 DECLARE_string(nics);
 
 DEFINE_string(config_file_a, "", "config of one network to compare");
 DEFINE_string(config_file_b, "", "config of another network to compare");
+DEFINE_bool(use_mkldnn_a, false, "whether to use mkldnn to run config_file_a");
+DEFINE_bool(use_mkldnn_b, false, "whether to use mkldnn to run config_file_b");
 DEFINE_bool(need_high_accuracy,
             false,
             "whether need to run in double accuracy");
@@ -128,6 +131,12 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
                 matA.getWidth());
   }
 
+  if (FLAGS_use_mkldnn_a || FLAGS_use_mkldnn_b) {
+    // some format of mkldnn parameter is different with cpu
+    // test_MKLDNN will check the parameters
+    return;
+  }
+
   vector<ParameterPtr>& parametersA = comDataA.parameters;
   vector<ParameterPtr>& parametersB = comDataB.parameters;
 
@@ -167,10 +176,12 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
 
 TEST(Trainer, create) {
   ComData dataA;
+  FLAGS_use_mkldnn = FLAGS_use_mkldnn_a;
   calcGradient(dataA, FLAGS_config_file_a);
   LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n";
 
   ComData dataB;
+  FLAGS_use_mkldnn = FLAGS_use_mkldnn_b;
   calcGradient(dataB, FLAGS_config_file_b);
   LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n";