Merge pull request #8404 from Xreki/core_refine_inference

Refine the inference API and unittests

Author: luotao1

paddle/fluid/framework/lod_tensor.cc

@@ -31,8 +31,14 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
   os << "{";
   for (auto &v : lod) {
     os << "{";
+    bool is_first = true;
     for (auto &i : v) {
-      os << i << ",";
+      if (is_first) {
+        os << i;
+        is_first = false;
+      } else {
+        os << ", " << i;
+      }
     }
     os << "}";
   }

paddle/fluid/inference/io.cc

@@ -32,23 +32,11 @@ void ReadBinaryFile(const std::string& filename, std::string& contents) {
   inputfs.close();
 }
 
-bool IsParameter(const framework::VarDesc* var,
-                 const framework::ProgramDesc& main_program) {
-  if (var->Persistable()) {
-    // There are many unreachable variables in the program
-    for (size_t i = 0; i < main_program.Size(); ++i) {
-      const framework::BlockDesc& block = main_program.Block(i);
-      for (auto* op : block.AllOps()) {
-        if (op->Type() == framework::kFeedOpType) {
-          continue;
-        }
-        for (auto input_argument_name : op->InputArgumentNames()) {
-          if (input_argument_name == var->Name()) {
-            return true;
-          }
-        }
-      }
-    }
+bool IsPersistable(const framework::VarDesc* var) {
+  if (var->Persistable() &&
+      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
+      var->GetType() != framework::proto::VarType::FETCH_LIST) {
+    return true;
   }
   return false;
 }
@@ -65,8 +53,8 @@ void LoadPersistables(framework::Executor& executor,
   std::vector<std::string> paramlist;
 
   for (auto* var : global_block.AllVars()) {
-    if (IsParameter(var, main_program)) {
-      VLOG(3) << "parameter's name: " << var->Name();
+    if (IsPersistable(var)) {
+      VLOG(3) << "persistable variable's name: " << var->Name();
 
       framework::VarDesc* new_var = load_block->Var(var->Name());
       new_var->SetShape(var->GetShape());
@@ -101,7 +89,6 @@ void LoadPersistables(framework::Executor& executor,
 
   executor.Run(*load_program, &scope, 0, true, true);
 
-  VLOG(3) << "Ran loading successfully";
   delete load_program;
 }
 

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

@@ -30,5 +30,5 @@ inference_test(label_semantic_roles)
 inference_test(recognize_digits ARGS mlp conv)
 inference_test(recommender_system)
 #inference_test(rnn_encoder_decoder)
-inference_test(understand_sentiment)
+inference_test(understand_sentiment ARGS conv)
 inference_test(word2vec)

paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc

@@ -32,16 +32,42 @@ TEST(inference, label_semantic_roles) {
   paddle::framework::LoDTensor word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1,
       ctx_p2, mark;
   paddle::framework::LoD lod{{0, 4, 10}};
-
-  SetupLoDTensor(word, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(
-      predicate, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_0, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(mark, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
+  int64_t word_dict_len = 44068;
+  int64_t predicate_dict_len = 3162;
+  int64_t mark_dict_len = 2;
+
+  SetupLoDTensor(word,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(predicate,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(predicate_dict_len - 1));
+  SetupLoDTensor(ctx_n2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_n1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_0,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(mark,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(mark_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&word);

paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc

@@ -31,7 +31,12 @@ TEST(inference, understand_sentiment) {
 
   paddle::framework::LoDTensor words;
   paddle::framework::LoD lod{{0, 4, 10}};
-  SetupLoDTensor(words, lod, static_cast<int64_t>(0), static_cast<int64_t>(10));
+  int64_t word_dict_len = 5147;
+
+  SetupLoDTensor(words,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&words);

paddle/fluid/inference/tests/book/test_inference_word2vec.cc

@@ -31,12 +31,12 @@ TEST(inference, word2vec) {
 
   paddle::framework::LoDTensor first_word, second_word, third_word, fourth_word;
   paddle::framework::LoD lod{{0, 1}};
-  int64_t dict_size = 2072;  // Hard-coding the size of dictionary
+  int64_t dict_size = 2073;  // The size of dictionary
 
-  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size);
+  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size - 1);
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&first_word);

paddle/fluid/inference/tests/test_helper.h

@@ -101,8 +101,8 @@ void TestInference(const std::string& dirname,
   if (IsCombined) {
     // All parameters are saved in a single file.
     // Hard-coding the file names of program and parameters in unittest.
-    // Users are free to specify different filename
-    // (provided: the filenames are changed in the python api as well: io.py)
+    // The file names should be consistent with that used in Python API
+    //  `fluid.io.save_inference_model`.
     std::string prog_filename = "__model_combined__";
     std::string param_filename = "__params_combined__";
     inference_program = paddle::inference::Load(executor,