Merge remote-tracking branch 'origin/develop' into memory/stable

cmake/tensorrt.cmake

@@ -16,7 +16,9 @@ find_library(TENSORRT_LIBRARY NAMES libnvinfer.so libnvinfer.a
     DOC "Path to TensorRT library.")
 
 if(TENSORRT_INCLUDE_DIR AND TENSORRT_LIBRARY)
+  if(WITH_DSO)
     set(TENSORRT_FOUND ON)
+  endif(WITH DSO)
 else()
     set(TENSORRT_FOUND OFF)
 endif()

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -429,7 +429,7 @@ struct LSTM : public PatternBase {
 
 struct GRU : public PatternBase {
   GRU(PDPattern* pattern, const std::string& name_scope)
-      : PatternBase(pattern, name_scope, "lstm") {}
+      : PatternBase(pattern, name_scope, "gru") {}
 
   PDNode* operator()(PDNode* x);
 

paddle/fluid/operators/distributed/grpc_client.cc

@@ -125,7 +125,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
   VarHandlePtr h(new VarHandle(ep, "Get", var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO([var_name_val, p_scope, p_ctx, s, this] {
+  framework::AsyncIO([var_name_val, s, this] {
     // prepare input
     sendrecv::VariableMessage req;
     req.set_varname(var_name_val);
@@ -166,7 +166,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
   s->Prepare(h, time_out);
 
   framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx,
-                      time_out, s, this] {
+                      s, this] {
     auto* var = p_scope->FindVar(in_var_name_val);
 
     ::grpc::ByteBuffer req;

paddle/fluid/operators/math/sequence_pooling.cc

@@ -103,6 +103,58 @@ class MaxSeqPoolGradFunctor {
   }
 };
 
+template <typename T>
+class LastSeqPoolFunctor {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::LoDTensor& input,
+                  framework::Tensor* output) {
+    // Create pointers to input and output data
+    auto* in_data = input.data<T>();
+    auto* out_data = output->data<T>();
+
+    // Calculate the size of each item in sequence
+    int64_t item_size = input.numel() / input.dims()[0];
+    auto lod = input.lod()[0];
+    int seq_num = static_cast<int>(lod.size()) - 1;
+    for (int i = 0; i < seq_num; ++i) {
+      // Calculate the length of each sequence
+      int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
+      // Point to the begin of next sequence
+      in_data += seq_len * item_size;
+      // Copy the last item of sequence to output
+      std::memcpy(out_data, (in_data - item_size), item_size * sizeof(T));
+      out_data += item_size;
+    }
+  }
+};
+
+template <typename T>
+class FirstSeqPoolFunctor {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::LoDTensor& input,
+                  framework::Tensor* output) {
+    // Create pointers to input and output data
+    auto* in_data = input.data<T>();
+    auto* out_data = output->data<T>();
+
+    // Calculate the size of each item in sequence
+    int64_t item_size = input.numel() / input.dims()[0];
+    auto lod = input.lod()[0];
+    int seq_num = static_cast<int>(lod.size()) - 1;
+    for (int i = 0; i < seq_num; ++i) {
+      // Calculate the length of each sequence
+      int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
+      // Copy the first item of sequence to output
+      std::memcpy(out_data, in_data, item_size * sizeof(T));
+      // Point to the next sequence
+      in_data += seq_len * item_size;
+      out_data += item_size;
+    }
+  }
+};
+
 template <typename T>
 class SequencePoolFunctor<platform::CPUDeviceContext, T> {
  public:
@@ -116,6 +168,16 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
       max_pool(context, input, output, index);
       return;
     }
+    if (pooltype == "LAST") {
+      math::LastSeqPoolFunctor<T> last_pool;
+      last_pool(context, input, output);
+      return;
+    }
+    if (pooltype == "FIRST") {
+      math::FirstSeqPoolFunctor<T> first_pool;
+      first_pool(context, input, output);
+      return;
+    }
     auto lod = input.lod()[0];
     auto& place = *context.eigen_device();
     for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
@@ -133,10 +195,6 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
       } else if (pooltype == "SQRT") {
         out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}})) /
                               std::sqrt(static_cast<T>(h));
-      } else if (pooltype == "LAST") {
-        out_e.device(place) = in_e.chip(h - 1, 0);
-      } else if (pooltype == "FIRST") {
-        out_e.device(place) = in_e.chip(0, 0);
       } else {
         PADDLE_THROW("unsupported pooling pooltype");
       }

paddle/scripts/paddle_build.sh

@@ -33,6 +33,7 @@ function print_usage() {
     ${BLUE}single_test${NONE}: run a single unit test
     ${BLUE}bind_test${NONE}: parallel tests bind to different GPU
     ${BLUE}doc${NONE}: generate paddle documents
+    ${BLUE}gen_doc_lib${NONE}: generate paddle documents library
     ${BLUE}html${NONE}: convert C++ source code into HTML
     ${BLUE}dockerfile${NONE}: generate paddle release dockerfile
     ${BLUE}capi${NONE}: generate paddle CAPI package
@@ -431,24 +432,60 @@ EOF
     linkchecker doc/v2/cn/html/index.html
     linkchecker doc/v2/api/en/html/index.html
 
-    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
+#    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
+#
+#    # Deploy to the the content server if its a "develop" or "release/version" branch
+#    # The "develop_doc" branch is reserved to test full deploy process without impacting the real content.
+#    if [ "$TRAVIS_BRANCH" == "develop_doc" ]; then
+#        PPO_SCRIPT_BRANCH=develop
+#    elif [[ "$TRAVIS_BRANCH" == "develop"  ||  "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then
+#        PPO_SCRIPT_BRANCH=master
+#    else
+#        # Early exit, this branch doesn't require documentation build
+#        return 0;
+#    fi
+#     # Fetch the paddlepaddle.org deploy_docs.sh from the appopriate branch
+#    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/$PPO_SCRIPT_BRANCH/scripts/deploy/deploy_docs.sh
+#    export PYTHONPATH=$PYTHONPATH:${PADDLE_ROOT}/build/python:/paddle/build/python
+#    cd ..
+#    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH ${PADDLE_ROOT} ${PADDLE_ROOT}/build/doc/ ${PPO_SCRIPT_BRANCH}
+#    cd -
+}
 
-    # Deploy to the the content server if its a "develop" or "release/version" branch
-    # The "develop_doc" branch is reserved to test full deploy process without impacting the real content.
-    if [ "$TRAVIS_BRANCH" == "develop_doc" ]; then
-        PPO_SCRIPT_BRANCH=develop
-    elif [[ "$TRAVIS_BRANCH" == "develop"  ||  "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then
-        PPO_SCRIPT_BRANCH=master
-    else
-        # Early exit, this branch doesn't require documentation build
-        return 0;
-    fi
-     # Fetch the paddlepaddle.org deploy_docs.sh from the appopriate branch
-    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/$PPO_SCRIPT_BRANCH/scripts/deploy/deploy_docs.sh
-    export PYTHONPATH=$PYTHONPATH:${PADDLE_ROOT}/build/python:/paddle/build/python
-    cd ..
-    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH ${PADDLE_ROOT} ${PADDLE_ROOT}/build/doc/ ${PPO_SCRIPT_BRANCH}
-    cd -
+function gen_doc_lib() {
+    mkdir -p ${PADDLE_ROOT}/build
+    cd ${PADDLE_ROOT}/build
+    cat <<EOF
+    ========================================
+    Building documentation library ...
+    In /paddle/build
+    ========================================
+EOF
+    cmake .. \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DWITH_DOC=ON \
+        -DWITH_GPU=OFF \
+        -DWITH_MKL=OFF \
+        -DWITH_FLUID_ONLY=ON
+
+    local LIB_TYPE=$1
+    case $LIB_TYPE in
+      full)
+        # Build full Paddle Python module. Will timeout without caching 'copy_paddle_pybind' first
+        make -j `nproc` gen_proto_py framework_py_proto copy_paddle_pybind paddle_python
+        ;;
+      pybind)
+        # Build paddle pybind library. Takes 49 minutes to build. Might timeout
+        make -j `nproc` copy_paddle_pybind
+        ;;
+      proto)
+        # Even smaller library.
+        make -j `nproc` framework_py_proto
+        ;;
+      *)
+        exit 0
+        ;;
+      esac
 }
 
 function gen_html() {
@@ -608,6 +645,9 @@ function main() {
       doc)
         gen_docs
         ;;
+      gen_doc_lib)
+        gen_doc_lib $2
+        ;;
       html)
         gen_html
         ;;

python/paddle/fluid/tests/unittests/dist_transformer.py

@@ -92,7 +92,7 @@ class TrainTaskConfig(object):
     src_vocab_fpath = data_path + "vocab.bpe.32000"
     trg_vocab_fpath = data_path + "vocab.bpe.32000"
     train_file_pattern = data_path + "train.tok.clean.bpe.32000.en-de"
-    val_file_pattern = data_path + "newstest2013.tok.bpe.32000.en-de"
+    val_file_pattern = data_path + "newstest2013.tok.bpe.32000.en-de.cut"
     pool_size = 2000
     sort_type = None
     local = True
@@ -624,6 +624,7 @@ def train_loop(exe, train_progm, dev_count, sum_cost, avg_cost, lr_scheduler,
             init = True
 
             # Validate and save the model for inference.
+            if batch_id == 0 or batch_id == 4:
                 if TrainTaskConfig.val_file_pattern is not None:
                     val_avg_cost, val_ppl = test()
                     print("[%f]" % val_avg_cost)
@@ -1701,8 +1702,9 @@ class DistTransformer2x2(TestDistRunnerBase):
         exe.run(startup_prog)
         exe.run(pserver_prog)
 
-    def run_trainer(self, place, args):
-
+    def run_trainer(self, use_cuda, args):
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+        TrainTaskConfig.use_gpu = use_cuda
         sum_cost, avg_cost, predict, token_num, local_lr_scheduler = get_model(
             args.is_dist, not args.sync_mode)
 

python/paddle/fluid/tests/unittests/test_dist_base.py

@@ -61,9 +61,10 @@ class TestDistRunnerBase(object):
         exe.run(startup_prog)
         exe.run(pserver_prog)
 
-    def run_trainer(self, place, args):
+    def run_trainer(self, use_cuda, args):
         import paddle
         import paddle.fluid as fluid
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
         test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
             self.get_model(batch_size=2)
         if args.mem_opt:
@@ -91,7 +92,7 @@ class TestDistRunnerBase(object):
             build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
 
         exe = fluid.ParallelExecutor(
-            True,
+            use_cuda,
             loss_name=avg_cost.name,
             exec_strategy=strategy,
             build_strategy=build_stra)
@@ -142,9 +143,8 @@ def runtime_main(test_class):
     if args.role == "pserver" and args.is_dist:
         model.run_pserver(args)
     else:
-        p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
-        ) else fluid.CPUPlace()
-        model.run_trainer(p, args)
+        use_cuda = True if core.is_compiled_with_cuda() else False
+        model.run_trainer(use_cuda, args)
 
 
 import paddle.compat as cpt
@@ -225,11 +225,12 @@ class TestDistBase(unittest.TestCase):
     def check_with_place(self, model_file, delta=1e-3, check_error_log=False):
         # TODO(typhoonzero): should auto adapt GPU count on the machine.
         required_envs = {
-            "PATH": os.getenv("PATH"),
-            "PYTHONPATH": os.getenv("PYTHONPATH"),
-            "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH"),
+            "PATH": os.getenv("PATH", ""),
+            "PYTHONPATH": os.getenv("PYTHONPATH", ""),
+            "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""),
             "FLAGS_fraction_of_gpu_memory_to_use": "0.15",
-            "FLAGS_cudnn_deterministic": "1"
+            "FLAGS_cudnn_deterministic": "1",
+            "CPU_NUM": "1"
         }
 
         if check_error_log:

python/paddle/fluid/tests/unittests/test_dist_transformer.py

@@ -14,6 +14,7 @@
 
 from __future__ import print_function
 
+import os
 import unittest
 import paddle
 from test_dist_base import TestDistBase
@@ -44,6 +45,14 @@ def download_files():
     test_url = url_prefix + 'newstest2013.tok.bpe.32000.en-de'
     test_md5 = '9dd74a266dbdb25314183899f269b4a2'
     paddle.dataset.common.download(test_url, 'test_dist_transformer', test_md5)
+    # cut test data for faster CI
+    orig_path = os.path.join(paddle.dataset.common.DATA_HOME,
+                             "test_dist_transformer",
+                             "newstest2013.tok.bpe.32000.en-de")
+    head_path = os.path.join(paddle.dataset.common.DATA_HOME,
+                             "test_dist_transformer",
+                             "newstest2013.tok.bpe.32000.en-de.cut")
+    os.system("head -n10 %s > %s" % (orig_path, head_path))
 
 
 class TestDistTransformer2x2Sync(TestDistBase):