Merge pull request #13100 from luotao1/ner_ut2

add unit-test for chinese_ner

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -40,23 +40,20 @@ function (inference_analysis_test TARGET)
     endif(WITH_TESTING)
 endfunction(inference_analysis_test)
 
-set(DITU_RNN_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fmodel.tar.gz")
-set(DITU_RNN_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fdata.txt.tar.gz")
-set(DITU_INSTALL_DIR "${THIRD_PARTY_PATH}/install/ditu_rnn" CACHE PATH "Ditu RNN model and data root." FORCE)
-set(DITU_RNN_MODEL ${DITU_INSTALL_DIR}/model)
-set(DITU_RNN_DATA ${DITU_INSTALL_DIR}/data.txt)
-
-function (inference_download_and_uncompress target url gz_filename)
+function (inference_download_and_uncompress install_dir url gz_filename)
     message(STATUS "Download inference test stuff ${gz_filename} from ${url}")
-    execute_process(COMMAND bash -c "mkdir -p ${DITU_INSTALL_DIR}")
-    execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && wget -q ${url}")
-    execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && tar xzf ${gz_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} && tar xzf ${gz_filename}")
     message(STATUS "finish downloading ${gz_filename}")
 endfunction(inference_download_and_uncompress)
 
+set(DITU_RNN_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fmodel.tar.gz")
+set(DITU_RNN_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fdata.txt.tar.gz")
+set(DITU_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/ditu_rnn" CACHE PATH "Ditu RNN model and data root." FORCE)
 if (NOT EXISTS ${DITU_INSTALL_DIR})
-    inference_download_and_uncompress(ditu_rnn_model ${DITU_RNN_MODEL_URL} "ditu_rnn_fluid%2Fmodel.tar.gz")
-    inference_download_and_uncompress(ditu_rnn_data ${DITU_RNN_DATA_URL} "ditu_rnn_fluid%2Fdata.txt.tar.gz")
+  inference_download_and_uncompress(${DITU_INSTALL_DIR} ${DITU_RNN_MODEL_URL} "ditu_rnn_fluid%2Fmodel.tar.gz")
+  inference_download_and_uncompress(${DITU_INSTALL_DIR} ${DITU_RNN_DATA_URL} "ditu_rnn_fluid%2Fdata.txt.tar.gz")
 endif()
 
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
@@ -87,3 +84,17 @@ inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_
 inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
 inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc)
 inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)
+
+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" CACHE PATH "Chinese ner model and data root." FORCE)
+if (NOT EXISTS ${CHINESE_NER_INSTALL_DIR})
+  inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_MODEL_URL} "chinese_ner_model.tar.gz")
+  inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_DATA_URL} "chinese_ner-data.txt.tar.gz")
+endif()
+
+inference_analysis_test(test_chinese_ner SRCS chinese_ner_tester.cc
+    EXTRA_DEPS paddle_inference_api paddle_fluid_api
+    ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model
+        --infer_model=${CHINESE_NER_INSTALL_DIR}/model
+        --infer_data=${CHINESE_NER_INSTALL_DIR}/data.txt)

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -201,13 +201,13 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   minute_tensor.lod.assign({one_batch.lod3});
   // clang-format on
   // assign data
-  TensorAssignData(&lod_attention_tensor,
-                   std::vector<std::vector<float>>({{0, 0}}));
+  TensorAssignData<float>(&lod_attention_tensor,
+                          std::vector<std::vector<float>>({{0, 0}}));
   std::vector<float> tmp_zeros(batch_size * 15, 0.);
-  TensorAssignData(&init_zero_tensor, {tmp_zeros});
-  TensorAssignData(&lod_tensor_tensor, one_batch.rnn_link_data);
-  TensorAssignData(&week_tensor, one_batch.rnn_week_datas);
-  TensorAssignData(&minute_tensor, one_batch.rnn_minute_datas);
+  TensorAssignData<float>(&init_zero_tensor, {tmp_zeros});
+  TensorAssignData<float>(&lod_tensor_tensor, one_batch.rnn_link_data);
+  TensorAssignData<float>(&week_tensor, one_batch.rnn_week_datas);
+  TensorAssignData<float>(&minute_tensor, one_batch.rnn_minute_datas);
   // Set inputs.
   auto init_zero_tensor1 = init_zero_tensor;
   init_zero_tensor1.name = "hidden_init";

paddle/fluid/inference/analysis/chinese_ner_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 <google/protobuf/text_format.h>
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/ir/pass.h"
+#include "paddle/fluid/inference/analysis/analyzer.h"
+#include "paddle/fluid/inference/analysis/ut_helper.h"
+#include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/platform/profiler.h"
+
+DEFINE_string(infer_model, "", "model path");
+DEFINE_string(infer_data, "", "data path");
+DEFINE_int32(batch_size, 10, "batch size.");
+DEFINE_int32(repeat, 1, "Running the inference program repeat times.");
+
+namespace paddle {
+namespace inference {
+
+struct DataRecord {
+  std::vector<std::vector<int64_t>> word_data_all, mention_data_all;
+  std::vector<std::vector<int64_t>> rnn_word_datas, rnn_mention_datas;
+  std::vector<size_t> lod;  // two inputs have the same lod info.
+  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 <= word_data_all.size()) {
+      data.word_data_all.assign(word_data_all.begin() + batch_iter,
+                                word_data_all.begin() + batch_end);
+      data.mention_data_all.assign(mention_data_all.begin() + batch_iter,
+                                   mention_data_all.begin() + batch_end);
+      // Prepare LoDs
+      data.lod.push_back(0);
+      CHECK(!data.word_data_all.empty());
+      CHECK(!data.mention_data_all.empty());
+      CHECK_EQ(data.word_data_all.size(), data.mention_data_all.size());
+      for (size_t j = 0; j < data.word_data_all.size(); j++) {
+        data.rnn_word_datas.push_back(data.word_data_all[j]);
+        data.rnn_mention_datas.push_back(data.mention_data_all[j]);
+        // calculate lod
+        data.lod.push_back(data.lod.back() + data.word_data_all[j].size());
+      }
+    }
+    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);
+      // load word data
+      std::vector<int64_t> word_data;
+      split_to_int64(data[1], ' ', &word_data);
+      // load mention data
+      std::vector<int64_t> mention_data;
+      split_to_int64(data[3], ' ', &mention_data);
+      word_data_all.push_back(std::move(word_data));
+      mention_data_all.push_back(std::move(mention_data));
+    }
+  }
+};
+
+void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
+                   int batch_size) {
+  PaddleTensor lod_word_tensor, lod_mention_tensor;
+  lod_word_tensor.name = "word";
+  lod_mention_tensor.name = "mention";
+  auto one_batch = data->NextBatch();
+  int size = one_batch.lod[one_batch.lod.size() - 1];  // token batch size
+  lod_word_tensor.shape.assign({size, 1});
+  lod_word_tensor.lod.assign({one_batch.lod});
+  lod_mention_tensor.shape.assign({size, 1});
+  lod_mention_tensor.lod.assign({one_batch.lod});
+  // assign data
+  TensorAssignData<int64_t>(&lod_word_tensor, one_batch.rnn_word_datas);
+  TensorAssignData<int64_t>(&lod_mention_tensor, one_batch.rnn_mention_datas);
+  // Set inputs.
+  input_slots->assign({lod_word_tensor, lod_mention_tensor});
+  for (auto &tensor : *input_slots) {
+    tensor.dtype = PaddleDType::INT64;
+  }
+}
+
+// the first inference result
+const int chinese_ner_result_data[] = {30, 45, 41, 48, 17, 26,
+                                       48, 39, 38, 16, 25};
+
+void TestChineseNERPrediction() {
+  NativeConfig 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;
+
+  auto predictor =
+      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  std::vector<PaddleTensor> input_slots;
+  DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
+  // Prepare inputs.
+  PrepareInputs(&input_slots, &data, FLAGS_batch_size);
+  std::vector<PaddleTensor> outputs;
+
+  Timer timer;
+  timer.tic();
+  for (int i = 0; i < FLAGS_repeat; i++) {
+    predictor->Run(input_slots, &outputs);
+  }
+  LOG(INFO) << "===========profile result===========";
+  LOG(INFO) << "batch_size: " << FLAGS_batch_size
+            << ", repeat: " << FLAGS_repeat
+            << ", latency: " << timer.toc() / FLAGS_repeat << "ms";
+  LOG(INFO) << "=====================================";
+
+  PADDLE_ENFORCE(outputs.size(), 1UL);
+  auto &out = outputs[0];
+  size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
+                                [](int a, int b) { return a * b; });
+  PADDLE_ENFORCE_GT(size, 0);
+  int64_t *result = static_cast<int64_t *>(out.data.data());
+  for (size_t i = 0; i < std::min(11UL, size); i++) {
+    PADDLE_ENFORCE(result[i], chinese_ner_result_data[i]);
+  }
+}
+
+// Directly infer with the original model.
+TEST(Analyzer, Chinese_ner) { TestChineseNERPrediction(); }
+
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/api/api_impl.cc

@@ -62,14 +62,14 @@ void NativePaddlePredictor::PrepareFeedFetch() {
   for (auto *op : inference_program_->Block(0).AllOps()) {
     if (op->Type() == "feed") {
       int idx = boost::get<int>(op->GetAttr("col"));
-      if (feeds_.size() <= static_cast<size_t>(idx)) {
+      if (feeds_.size() <= (size_t)idx) {
         feeds_.resize(idx + 1);
       }
       feeds_[idx] = op;
       feed_names_[op->Output("Out")[0]] = idx;
     } else if (op->Type() == "fetch") {
       int idx = boost::get<int>(op->GetAttr("col"));
-      if (fetchs_.size() <= idx) {
+      if (fetchs_.size() <= (size_t)idx) {
         fetchs_.resize(idx + 1);
       }
       fetchs_[idx] = op;
@@ -222,6 +222,62 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
   }
   return true;
 }
+template <typename T>
+void NativePaddlePredictor::GetFetchOne(const framework::LoDTensor &fetch,
+                                        PaddleTensor *output) {
+  std::vector<int> shape;
+  auto dims_i = fetch.dims();
+  auto lod = fetch.lod();
+  const T *output_ptr = fetch.data<T>();
+  auto num = fetch.numel();
+  std::vector<T> data;
+  if (0 == lod.size()) {
+    std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
+    for (int j = 0; j < dims_i.size(); ++j) {
+      shape.push_back(dims_i[j]);
+    }
+  } else {
+    // for batch detection
+    // image[0] -> output[0] shape {145, 6}
+    // image[1] -> output[1] shape {176, 6}
+    // then,
+    // the batch output shape {321, 6}
+    // the lod {{0, 145, 321}}
+    // so we should append output[0] to {176, 6}
+    size_t max_dim = 0;
+    for (size_t j = 1; j < lod[0].size(); j++) {
+      max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
+    }
+    size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
+    if (max_dim > 0) {
+      data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
+    }
+    for (size_t j = 1; j < lod[0].size(); j++) {
+      size_t start = lod[0][j - 1] * common_dim;
+      size_t end = lod[0][j] * common_dim;
+      if (end > start) {
+        std::copy(output_ptr + start, output_ptr + end,
+                  data.begin() + (j - 1) * max_dim * common_dim);
+      }
+    }
+    shape.push_back(lod[0].size() - 1);
+    shape.push_back(max_dim);
+    for (int j = 1; j < dims_i.size(); ++j) {
+      shape.push_back(dims_i[j]);
+    }
+  }
+
+  output->shape = shape;
+  auto &buffer = output->data;
+  if (buffer.empty() || buffer.length() < sizeof(T) * data.size()) {
+    buffer.Resize(sizeof(T) * data.size());
+  }
+  std::memcpy(buffer.data(), data.data(), buffer.length());
+  // copy LoD
+  for (const auto &level : fetch.lod()) {
+    output->lod.emplace_back(level);
+  }
+}
 
 bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
                                      framework::Scope *scope) {
@@ -229,70 +285,20 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
   outputs->resize(fetchs_.size());
   for (size_t i = 0; i < fetchs_.size(); ++i) {
     int idx = boost::get<int>(fetchs_[i]->GetAttr("col"));
-    PADDLE_ENFORCE(idx == i);
-    framework::LoDTensor &output =
+    PADDLE_ENFORCE((size_t)idx == i);
+    framework::LoDTensor &fetch =
         framework::GetFetchVariable(*scope, "fetch", idx);
-    // TODO(panyx0718): Support fetch of other types.
-    if (output.type() != typeid(float)) {
-      LOG(ERROR) << "only support fetching float now.";
-      return false;
-    }
-
-    std::vector<int> shape;
-    auto dims_i = output.dims();
-    auto lod = output.lod();
-    const float *output_ptr = output.data<float>();
-    // const int64_t* output_ptr = fetchs[i].data<int64_t>();
-    auto num = output.numel();
-    std::vector<float> data;
-    if (0 == lod.size()) {
-      std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
-      for (int j = 0; j < dims_i.size(); ++j) {
-        shape.push_back(dims_i[j]);
-      }
+    auto type = fetch.type();
+    auto output = &(outputs->at(i));
+    if (type == typeid(float)) {
+      GetFetchOne<float>(fetch, output);
+      output->dtype = PaddleDType::FLOAT32;
+    } else if (type == typeid(int64_t)) {
+      GetFetchOne<int64_t>(fetch, output);
+      output->dtype = PaddleDType::INT64;
     } else {
-      // for batch detection
-      // image[0] -> output[0] shape {145, 6}
-      // image[1] -> output[1] shape {176, 6}
-      // then,
-      // the batch output shape {321, 6}
-      // the lod {{0, 145, 321}}
-      // so we should append output[0] to {176, 6}
-      size_t max_dim = 0;
-      for (size_t j = 1; j < lod[0].size(); j++) {
-        max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
-      }
-      size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
-      if (max_dim > 0) {
-        data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
-      }
-      for (size_t j = 1; j < lod[0].size(); j++) {
-        size_t start = lod[0][j - 1] * common_dim;
-        size_t end = lod[0][j] * common_dim;
-        if (end > start) {
-          std::copy(output_ptr + start, output_ptr + end,
-                    data.begin() + (j - 1) * max_dim * common_dim);
-        }
-      }
-      shape.push_back(lod[0].size() - 1);
-      shape.push_back(max_dim);
-      for (int j = 1; j < dims_i.size(); ++j) {
-        shape.push_back(dims_i[j]);
-      }
-    }
-
-    outputs->at(i).shape = shape;
-    auto &buffer = outputs->at(i).data;
-    if (buffer.empty() || buffer.length() < sizeof(float) * data.size()) {
-      buffer.Resize(sizeof(float) * data.size());
-    }
-    std::memcpy(buffer.data(), data.data(), buffer.length());
-    // copy LoD
-    for (const auto &level : output.lod()) {
-      outputs->at(i).lod.emplace_back(level);
+      LOG(ERROR) << "unknown type, only support float32 and int64 now.";
     }
-    outputs->at(i).dtype = PaddleDType::FLOAT32;
-    // TODO(panyx0718): support other types? fill tensor name? avoid a copy.
   }
   return true;
 }

paddle/fluid/inference/api/api_impl.h

@@ -51,7 +51,9 @@ class NativePaddlePredictor : public PaddlePredictor {
                framework::Scope *scope);
   bool GetFetch(std::vector<PaddleTensor> *output_data,
                 framework::Scope *scope);
-
+  template <typename T>
+  void GetFetchOne(const framework::LoDTensor &fetchs,
+                   PaddleTensor *output_data);
   void PrepareFeedFetch();
 
   NativeConfig config_;

paddle/fluid/inference/api/helper.h

@@ -68,6 +68,13 @@ static void split_to_float(const std::string &str, char sep,
   std::transform(pieces.begin(), pieces.end(), std::back_inserter(*fs),
                  [](const std::string &v) { return std::stof(v); });
 }
+static void split_to_int64(const std::string &str, char sep,
+                           std::vector<int64_t> *is) {
+  std::vector<std::string> pieces;
+  split(str, sep, &pieces);
+  std::transform(pieces.begin(), pieces.end(), std::back_inserter(*is),
+                 [](const std::string &v) { return std::stoi(v); });
+}
 template <typename T>
 std::string to_string(const std::vector<T> &vec) {
   std::stringstream ss;
@@ -84,14 +91,18 @@ template <>
 std::string to_string<std::vector<std::vector<float>>>(
     const std::vector<std::vector<std::vector<float>>> &vec);
 
-// clang-format off
-static void TensorAssignData(PaddleTensor *tensor, const std::vector<std::vector<float>> &data) {
+template <typename T>
+static void TensorAssignData(PaddleTensor *tensor,
+                             const std::vector<std::vector<T>> &data) {
   // Assign buffer
-  int dim = std::accumulate(tensor->shape.begin(), tensor->shape.end(), 1, [](int a, int b) { return a * b; });
-  tensor->data.Resize(sizeof(float) * dim);
+  int dim = std::accumulate(tensor->shape.begin(), tensor->shape.end(), 1,
+                            [](int a, int b) { return a * b; });
+  tensor->data.Resize(sizeof(T) * dim);
   int c = 0;
   for (const auto &f : data) {
-    for (float v : f) { static_cast<float *>(tensor->data.data())[c++] = v; }
+    for (T v : f) {
+      static_cast<T *>(tensor->data.data())[c++] = v;
+    }
   }
 }