Merge branch 'develop' into nlp_multi_thread

paddle/fluid/inference/tests/api/CMakeLists.txt

-function (inference_download_and_uncompress install_dir url)
-    get_filename_component(filename ${url} NAME)
-    message(STATUS "Download inference test stuff ${filename} from ${url}")
+set(INFERENCE_URL "http://paddle-inference-dist.bj.bcebos.com")
+set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo")
+set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
+function (inference_download_and_uncompress install_dir filename)
+    message(STATUS "Download inference test stuff from ${INFERENCE_URL}/${filename}")
     execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
-    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}")
+    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${INFERENCE_URL}/${filename}")
     execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
     message(STATUS "finish downloading ${filename}")
 endfunction(inference_download_and_uncompress)
 
-function(download_model_and_data install_dir model_url data_url)
+function(download_model_and_data install_dir model_name data_name)
     if (NOT EXISTS ${install_dir} AND WITH_INFERENCE)
-        inference_download_and_uncompress(${install_dir} ${model_url})
-        inference_download_and_uncompress(${install_dir} ${data_url})
+        inference_download_and_uncompress(${install_dir} ${model_name})
+        inference_download_and_uncompress(${install_dir} ${data_name})
     endif()
 endfunction()
 
 # RNN1
-set(RNN1_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fmodel.tar.gz")
-set(RNN1_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fdata.txt.tar.gz")
-set(RNN1_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/rnn1")
-download_model_and_data(${RNN1_INSTALL_DIR} ${RNN1_MODEL_URL} ${RNN1_DATA_URL})
-inference_analysis_test(test_analyzer_rnn1 SRCS analyzer_rnn1_tester.cc
-    EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
+set(RNN1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn1")
+download_model_and_data(${RNN1_INSTALL_DIR} "rnn1%2Fmodel.tar.gz" "rnn1%2Fdata.txt.tar.gz")
+inference_analysis_test(test_analyzer_rnn1 SRCS analyzer_rnn1_tester.cc 
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
     ARGS --infer_model=${RNN1_INSTALL_DIR}/model
          --infer_data=${RNN1_INSTALL_DIR}/data.txt)
 
+# RNN2
+set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2")
+download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz")
+inference_analysis_test(test_analyzer_rnn2 SRCS analyzer_rnn2_tester.cc
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
+    ARGS --infer_model=${RNN2_INSTALL_DIR}/model
+         --infer_data=${RNN2_INSTALL_DIR}/data.txt)
+
 # chinese_ner
-set(CHINESE_NER_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner_model.tar.gz")
-set(CHINESE_NER_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner-data.txt.tar.gz")
-set(CHINESE_NER_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/chinese_ner")
-download_model_and_data(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_MODEL_URL} ${CHINESE_NER_DATA_URL})
+set(CHINESE_NER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/chinese_ner")
+download_model_and_data(${CHINESE_NER_INSTALL_DIR} "chinese_ner_model.tar.gz" "chinese_ner-data.txt.tar.gz")
 inference_analysis_test(test_analyzer_ner SRCS analyzer_ner_tester.cc
-    EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
     ARGS --infer_model=${CHINESE_NER_INSTALL_DIR}/model
         --infer_data=${CHINESE_NER_INSTALL_DIR}/data.txt)
 
 # lac
-set(LAC_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/lac_model.tar.gz")
-set(LAC_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/lac_data.txt.tar.gz")
-set(LAC_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/lac")
-download_model_and_data(${LAC_INSTALL_DIR} ${LAC_MODEL_URL} ${LAC_DATA_URL})
+set(LAC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/lac")
+download_model_and_data(${LAC_INSTALL_DIR} "lac_model.tar.gz" "lac_data.txt.tar.gz")
 inference_analysis_test(test_analyzer_lac SRCS analyzer_lac_tester.cc
-    EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
     ARGS --infer_model=${LAC_INSTALL_DIR}/model
         --infer_data=${LAC_INSTALL_DIR}/data.txt)
 
 # text_classification
-set(TEXT_CLASSIFICATION_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/text-classification-Senta.tar.gz")
-set(TEXT_CLASSIFICATION_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/text_classification_data.txt.tar.gz")
-set(TEXT_CLASSIFICATION_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/text_classification")
-download_model_and_data(${TEXT_CLASSIFICATION_INSTALL_DIR} ${TEXT_CLASSIFICATION_MODEL_URL} ${TEXT_CLASSIFICATION_DATA_URL})
+set(TEXT_CLASSIFICATION_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/text_classification")
+download_model_and_data(${TEXT_CLASSIFICATION_INSTALL_DIR} "text-classification-Senta.tar.gz" "text_classification_data.txt.tar.gz")
 inference_analysis_test(test_analyzer_text_classification SRCS analyzer_text_classification_tester.cc
-    EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
     ARGS --infer_model=${TEXT_CLASSIFICATION_INSTALL_DIR}/text-classification-Senta
          --infer_data=${TEXT_CLASSIFICATION_INSTALL_DIR}/data.txt)

paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc

+// Copyright (c) 2018 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.
+
+#include "paddle/fluid/inference/analysis/analyzer.h"
+
+#include <google/protobuf/text_format.h>
+#include <gtest/gtest.h>
+#include <thread>  // NOLINT
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/pass.h"
+#include "paddle/fluid/inference/analysis/ut_helper.h"
+#include "paddle/fluid/inference/api/analysis_predictor.h"
+#include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/inference/api/paddle_inference_pass.h"
+
+DEFINE_string(infer_model, "", "model path");
+DEFINE_string(infer_data, "", "data path");
+DEFINE_int32(batch_size, 1, "batch size.");
+DEFINE_int32(repeat, 1, "Running the inference program repeat times.");
+DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads.");
+
+namespace paddle {
+namespace inference {
+
+using namespace framework;  // NOLINT
+
+struct DataRecord {
+  std::vector<std::vector<std::vector<float>>> link_step_data_all;
+  std::vector<size_t> lod;
+  std::vector<std::vector<float>> rnn_link_data;
+  std::vector<float> result_data;
+  size_t batch_iter{0};
+  size_t batch_size{1};
+  DataRecord() = default;
+  explicit DataRecord(const std::string &path, int batch_size = 1)
+      : batch_size(batch_size) {
+    Load(path);
+  }
+  DataRecord NextBatch() {
+    DataRecord data;
+    size_t batch_end = batch_iter + batch_size;
+    // NOTE skip the final batch, if no enough data is provided.
+    if (batch_end <= link_step_data_all.size()) {
+      data.link_step_data_all.assign(link_step_data_all.begin() + batch_iter,
+                                     link_step_data_all.begin() + batch_end);
+      // Prepare LoDs
+      data.lod.push_back(0);
+      CHECK(!data.link_step_data_all.empty()) << "empty";
+      for (size_t j = 0; j < data.link_step_data_all.size(); j++) {
+        for (const auto &d : data.link_step_data_all[j]) {
+          data.rnn_link_data.push_back(d);
+          // calculate lod
+          data.lod.push_back(data.lod.back() + 11);
+        }
+      }
+    }
+    batch_iter += batch_size;
+    return data;
+  }
+  void Load(const std::string &path) {
+    std::ifstream file(path);
+    std::string line;
+    int num_lines = 0;
+    while (std::getline(file, line)) {
+      num_lines++;
+      std::vector<std::string> data;
+      split(line, ':', &data);
+      if (num_lines % 2) {  // feature
+        std::vector<std::string> feature_data;
+        split(data[1], ' ', &feature_data);
+        std::vector<std::vector<float>> link_step_data;
+        int feature_count = 1;
+        std::vector<float> feature;
+        for (auto &step_data : feature_data) {
+          std::vector<float> tmp;
+          split_to_float(step_data, ',', &tmp);
+          feature.insert(feature.end(), tmp.begin(), tmp.end());
+          if (feature_count % 11 == 0) {  // each sample has 11 features
+            link_step_data.push_back(feature);
+            feature.clear();
+          }
+          feature_count++;
+        }
+        link_step_data_all.push_back(std::move(link_step_data));
+      } else {  // result
+        std::vector<float> tmp;
+        split_to_float(data[1], ',', &tmp);
+        result_data.insert(result_data.end(), tmp.begin(), tmp.end());
+      }
+    }
+  }
+};
+void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
+                   int batch_size) {
+  PaddleTensor feed_tensor;
+  feed_tensor.name = "feed";
+  auto one_batch = data->NextBatch();
+  int token_size = one_batch.rnn_link_data.size();
+  // each token has 11 features, each feature's dim is 54.
+  std::vector<int> rnn_link_data_shape({token_size * 11, 54});
+  feed_tensor.shape = rnn_link_data_shape;
+  feed_tensor.lod.assign({one_batch.lod});
+  feed_tensor.dtype = PaddleDType::FLOAT32;
+  TensorAssignData<float>(&feed_tensor, one_batch.rnn_link_data);
+  // Set inputs.
+  input_slots->assign({feed_tensor});
+}
+
+void CompareResult(const std::vector<PaddleTensor> &outputs,
+                   const std::vector<float> &base_result) {
+  PADDLE_ENFORCE_GT(outputs.size(), 0);
+  for (size_t i = 0; i < outputs.size(); i++) {
+    auto &out = outputs[i];
+    size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
+                                  [](int a, int b) { return a * b; });
+    PADDLE_ENFORCE_GT(size, 0);
+    float *data = static_cast<float *>(out.data.data());
+    for (size_t i = 0; i < size; i++) {
+      EXPECT_NEAR(data[i], base_result[i], 1e-3);
+    }
+  }
+}
+// Test with a really complicate model.
+void TestRNN2Prediction() {
+  AnalysisConfig config;
+  config.prog_file = FLAGS_infer_model + "/__model__";
+  config.param_file = FLAGS_infer_model + "/param";
+  config.use_gpu = false;
+  config.device = 0;
+  config.specify_input_name = true;
+  config.enable_ir_optim = true;
+  PADDLE_ENFORCE(config.ir_mode ==
+                 AnalysisConfig::IrPassMode::kExclude);  // default
+
+  int batch_size = FLAGS_batch_size;
+  int num_times = FLAGS_repeat;
+
+  auto base_predictor =
+      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto predictor =
+      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
+          config);
+  std::vector<PaddleTensor> input_slots;
+  DataRecord data(FLAGS_infer_data, batch_size);
+  PrepareInputs(&input_slots, &data, batch_size);
+  std::vector<PaddleTensor> outputs, base_outputs;
+
+  Timer timer1;
+  timer1.tic();
+  for (int i = 0; i < num_times; i++) {
+    base_predictor->Run(input_slots, &base_outputs);
+  }
+  PrintTime(batch_size, num_times, 1, 0, timer1.toc() / num_times);
+
+  Timer timer2;
+  timer2.tic();
+  for (int i = 0; i < num_times; i++) {
+    predictor->Run(input_slots, &outputs);
+  }
+  PrintTime(batch_size, num_times, 1, 0, timer2.toc() / num_times);
+
+  CompareResult(base_outputs, data.result_data);
+  CompareResult(outputs, data.result_data);
+}
+
+TEST(Analyzer, rnn2) { TestRNN2Prediction(); }
+
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/operators/distributed/grpc_client.cc

@@ -290,12 +290,18 @@ void GRPCClient::Proceed() {
       c->Finish(false);
     }
 
-    delete c;
+    bool notify = false;
     {
       std::lock_guard<std::mutex> lk(sync_mutex_);
       req_count_--;
+      notify = (req_count_ <= 0 || !c->status_.ok());
+    }
+
+    delete c;
+
+    if (notify) {
+      sync_cond_.notify_all();
     }
-    sync_cond_.notify_all();
   }
   VLOG(3) << "GRPCClient Proceed end";
 }

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -28,6 +28,10 @@ list(REMOVE_ITEM TEST_OPS test_cond_op) # FIXME(qijun): https://github.com/Paddl
 
 list(REMOVE_ITEM TEST_OPS op_test) # op_test is a helper python file, not a test
 list(REMOVE_ITEM TEST_OPS decorators) # decorators is a helper python file, not a test
+if(APPLE)
+    # this op is not support on mac
+    list(REMOVE_ITEM TEST_OPS test_fusion_seqexpand_concat_fc_op)
+endif()
 
 function(py_test_modules TARGET_NAME)
   if(WITH_TESTING)

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

@@ -84,7 +84,7 @@ class TestDataBalance(unittest.TestCase):
         self.data_file_name = './data_balance_test.recordio'
         self.lod_data_file_name = './data_balance_with_lod_test.recordio'
         self.total_ins_num = 50
-        self.batch_size = 10
+        self.batch_size = 12
         self.prepare_data()
         self.prepare_lod_data()
 

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

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from __future__ import print_function
-
+import os
 import paddle.fluid as fluid
 import paddle
 import numpy as np
@@ -41,6 +41,8 @@ class TestReaderReset(unittest.TestCase):
                 self.data_file_name, reader, feeder)
 
     def setUp(self):
+        # set parallel threads to fit 20 batches in line 49
+        os.environ['CPU_NUM'] = str(20)
         self.use_cuda = fluid.core.is_compiled_with_cuda()
         self.data_file_name = './reader_reset_test.recordio'
         self.ins_shape = [3]