replace zero_copy_run to run for memory leak

deploy/cpp_infer/src/ocr_det.cpp

@@ -31,7 +31,8 @@ void DBDetector::LoadModel(const std::string &model_dir) {
   }
 
   // false for zero copy tensor
-  config.SwitchUseFeedFetchOps(false);
+  // true for commom tensor
+  config.SwitchUseFeedFetchOps(true);
   // true for multiple input
   config.SwitchSpecifyInputNames(true);
 
@@ -59,12 +60,13 @@ void DBDetector::Run(cv::Mat &img,
   std::vector<float> input(1 * 3 * resize_img.rows * resize_img.cols, 0.0f);
   this->permute_op_.Run(&resize_img, input.data());
 
-  auto input_names = this->predictor_->GetInputNames();
-  auto input_t = this->predictor_->GetInputTensor(input_names[0]);
-  input_t->Reshape({1, 3, resize_img.rows, resize_img.cols});
-  input_t->copy_from_cpu(input.data());
-
-  this->predictor_->ZeroCopyRun();
+  // Inference.
+  paddle::PaddleTensor input_t;
+  input_t.shape = {1, 3, resize_img.rows, resize_img.cols};
+  input_t.data = paddle::PaddleBuf(input.data(), input.size() * sizeof(float));
+  input_t.dtype = PaddleDType::FLOAT32;
+  std::vector<paddle::PaddleTensor> outputs;
+  this->predictor_->Run({input_t}, &outputs, 1);
 
   std::vector<float> out_data;
   auto output_names = this->predictor_->GetOutputNames();

deploy/cpp_infer/src/ocr_rec.cpp

@@ -39,18 +39,21 @@ void CRNNRecognizer::Run(std::vector<std::vector<std::vector<int>>> boxes,
 
     this->permute_op_.Run(&resize_img, input.data());
 
-    auto input_names = this->predictor_->GetInputNames();
-    auto input_t = this->predictor_->GetInputTensor(input_names[0]);
-    input_t->Reshape({1, 3, resize_img.rows, resize_img.cols});
-    input_t->copy_from_cpu(input.data());
-
-    this->predictor_->ZeroCopyRun();
+    // Inference.
+    paddle::PaddleTensor input_t;
+    input_t.shape = {1, 3, resize_img.rows, resize_img.cols};
+    input_t.data =
+        paddle::PaddleBuf(input.data(), input.size() * sizeof(float));
+    input_t.dtype = PaddleDType::FLOAT32;
+    std::vector<paddle::PaddleTensor> outputs;
+    this->predictor_->Run({input_t}, &outputs, 1);
 
     std::vector<int64_t> rec_idx;
     auto output_names = this->predictor_->GetOutputNames();
     auto output_t = this->predictor_->GetOutputTensor(output_names[0]);
     auto rec_idx_lod = output_t->lod();
     auto shape_out = output_t->shape();
+
     int out_num = std::accumulate(shape_out.begin(), shape_out.end(), 1,
                                   std::multiplies<int>());
 
@@ -120,7 +123,8 @@ void CRNNRecognizer::LoadModel(const std::string &model_dir) {
   }
 
   // false for zero copy tensor
-  config.SwitchUseFeedFetchOps(false);
+  // true for commom tensor
+  config.SwitchUseFeedFetchOps(true);
   // true for multiple input
   config.SwitchSpecifyInputNames(true);
 

tools/infer/predict_det.py

@@ -17,22 +17,25 @@ __dir__ = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(__dir__)
 sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
 
+import cv2
+import copy
+import numpy as np
+import math
+import time
+import sys
+
+import paddle.fluid as fluid
+
 import tools.infer.utility as utility
 from ppocr.utils.utility import initial_logger
 logger = initial_logger()
 from ppocr.utils.utility import get_image_file_list, check_and_read_gif
-import cv2
 from ppocr.data.det.sast_process import SASTProcessTest
 from ppocr.data.det.east_process import EASTProcessTest
 from ppocr.data.det.db_process import DBProcessTest
 from ppocr.postprocess.db_postprocess import DBPostProcess
 from ppocr.postprocess.east_postprocess import EASTPostPocess
 from ppocr.postprocess.sast_postprocess import SASTPostProcess
-import copy
-import numpy as np
-import math
-import time
-import sys
 
 
 class TextDetector(object):
@@ -127,7 +130,7 @@ class TextDetector(object):
             dt_boxes_new.append(box)
         dt_boxes = np.array(dt_boxes_new)
         return dt_boxes
-    
+
     def __call__(self, img):
         ori_im = img.copy()
         im, ratio_list = self.preprocess_op(img)
@@ -135,8 +138,8 @@ class TextDetector(object):
             return None, 0
         im = im.copy()
         starttime = time.time()
-        self.input_tensor.copy_from_cpu(im)
-        self.predictor.zero_copy_run()
+        im = fluid.core.PaddleTensor(im)
+        self.predictor.run([im])
         outputs = []
         for output_tensor in self.output_tensors:
             output = output_tensor.copy_to_cpu()
@@ -152,7 +155,7 @@ class TextDetector(object):
             outs_dict['f_tvo'] = outputs[3]
         else:
             outs_dict['maps'] = outputs[0]
-                
+
         dt_boxes_list = self.postprocess_op(outs_dict, [ratio_list])
         dt_boxes = dt_boxes_list[0]
         if self.det_algorithm == "SAST" and self.det_sast_polygon:

tools/infer/predict_rec.py

@@ -17,15 +17,18 @@ __dir__ = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(__dir__)
 sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
 
-import tools.infer.utility as utility
-from ppocr.utils.utility import initial_logger
-logger = initial_logger()
-from ppocr.utils.utility import get_image_file_list, check_and_read_gif
 import cv2
 import copy
 import numpy as np
 import math
 import time
+
+import paddle.fluid as fluid
+
+import tools.infer.utility as utility
+from ppocr.utils.utility import initial_logger
+logger = initial_logger()
+from ppocr.utils.utility import get_image_file_list, check_and_read_gif
 from ppocr.utils.character import CharacterOps
 
 
@@ -102,8 +105,8 @@ class TextRecognizer(object):
             norm_img_batch = np.concatenate(norm_img_batch)
             norm_img_batch = norm_img_batch.copy()
             starttime = time.time()
-            self.input_tensor.copy_from_cpu(norm_img_batch)
-            self.predictor.zero_copy_run()
+            norm_img_batch = fluid.core.PaddleTensor(norm_img_batch)
+            self.predictor.run([norm_img_batch])
 
             if self.loss_type == "ctc":
                 rec_idx_batch = self.output_tensors[0].copy_to_cpu()

tools/infer/utility.py

@@ -106,8 +106,7 @@ def create_predictor(args, mode):
     config.disable_glog_info()
 
     # use zero copy
-    config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass")
-    config.switch_use_feed_fetch_ops(False)
+    config.switch_use_feed_fetch_ops(True)
     predictor = create_paddle_predictor(config)
     input_names = predictor.get_input_names()
     input_tensor = predictor.get_input_tensor(input_names[0])