fix inference batch_size configuration (PaddlePaddle#57)

* gpu_graph_infer

* simplify infer

* fix

* remove logs

* remove logs

* change logs

* speed up dumpfile

* reduce inference time

* replace to_string

* fix inference

* change log

* fix infer batch_size

* remove log

* check tensor initialization

paddle/fluid/framework/data_feed.cu

@@ -554,7 +554,7 @@ int GraphDataGenerator::GenerateBatch() {
   }
 
   cudaStreamSynchronize(stream_);
-  if (!gpu_graph_training_) return total_instance / 2;
+  if (!gpu_graph_training_) return 1;
   ins_buf_pair_len_ -= total_instance / 2;
   if (debug_mode_) {
     uint64_t h_slot_tensor[slot_num_][total_instance];
@@ -966,12 +966,12 @@ void GraphDataGenerator::SetConfig(
   window_ = graph_config.window();
   once_sample_startid_len_ = graph_config.once_sample_startid_len();
   debug_mode_ = graph_config.debug_mode();
-  if (debug_mode_) {
+  gpu_graph_training_ = graph_config.gpu_graph_training();
+  if (debug_mode_ || !gpu_graph_training_) {
     batch_size_ = graph_config.batch_size();
   } else {
     batch_size_ = once_sample_startid_len_;
   }
-  gpu_graph_training_ = graph_config.gpu_graph_training();
   repeat_time_ = graph_config.sample_times_one_chunk();
   buf_size_ =
       once_sample_startid_len_ * walk_len_ * walk_degree_ * repeat_time_;

paddle/fluid/framework/device_worker.cc

@@ -262,11 +262,35 @@ void DeviceWorker::DumpField(const Scope& scope, int dump_mode,
   size_t batch_size = device_reader_->GetCurBatchSize();
   auto& ins_id_vec = device_reader_->GetInsIdVec();
   auto& ins_content_vec = device_reader_->GetInsContentVec();
-  if (ins_id_vec.size() > 0) {
+  if (dump_mode_ == 3) {
+    batch_size = std::string::npos;
+    bool has_valid_batch = false;
+    for (auto& field : *dump_fields_) {
+      Variable* var = scope.FindVar(field);
+      if (var == nullptr) {
+        VLOG(0) << "Note: field[" << field
+                << "] cannot be find in scope, so it was skipped.";
+        continue;
+      }
+      LoDTensor* tensor = var->GetMutable<LoDTensor>();
+      if (!tensor->IsInitialized()) {
+        VLOG(0) << "Note: field[" << field
+                << "] is not initialized, so it was skipped.";
+        continue;
+      }
+      auto& dims = tensor->dims();
+      if (dims.size() == 2 && dims[0] > 0) {
+        batch_size = std::min(batch_size, static_cast<size_t>(dims[0]));
+        // VLOG(0)<<"in dump field ---> "<<field<<" dim_size = "<<dims[0]<<"
+        // "<<dims[1]<<" batch_size = "<<batch_size;
+        has_valid_batch = true;
+      }
+    }
+    if (!has_valid_batch) return;
+  } else if (ins_id_vec.size() > 0) {
     batch_size = ins_id_vec.size();
   }
   std::vector<std::string> ars(batch_size);
-  std::vector<bool> hit(batch_size, false);
   if (dump_mode_ == 3) {
     if (dump_fields_ == NULL || (*dump_fields_).size() == 0) {
       return;
@@ -308,7 +332,7 @@ void DeviceWorker::DumpField(const Scope& scope, int dump_mode,
         tensor = &cpu_tensor;
       }
       auto& dims = tensor->dims();
-      if (dims.size() != 2 || dims[0] != static_cast<int>(batch_size)) {
+      if (dims.size() != 2 || dims[0] <= 0) {
         VLOG(0) << "Note: field[" << field << "] cannot pass check, so it was "
                                               "skipped. Maybe the dimension is "
                                               "wrong ";
@@ -348,6 +372,7 @@ void DeviceWorker::DumpField(const Scope& scope, int dump_mode,
     }
     return;
   }
+  std::vector<bool> hit(batch_size, false);
   std::default_random_engine engine(0);
   std::uniform_int_distribution<size_t> dist(0U, INT_MAX);
   for (size_t i = 0; i < batch_size; i++) {