[Bug Fix] Fix PPOCR dynamic input shape bug (PaddlePaddle#667)

* Imporve OCR Readme * Improve OCR Readme * Improve OCR Readme * Improve OCR Readme * Improve OCR Readme * Add Initialize function to PP-OCR * Add Initialize function to PP-OCR * Add Initialize function to PP-OCR * Make all the model links come from PaddleOCR * Improve OCR readme * Improve OCR readme * Improve OCR readme * Improve OCR readme * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add Readme for vision results * Add check for label file in postprocess of Rec model * Add check for label file in postprocess of Rec model * Add check for label file in postprocess of Rec model * Add check for label file in postprocess of Rec model * Add check for label file in postprocess of Rec model * Add check for label file in postprocess of Rec model * Add comments to create API docs * Improve OCR comments * Rename OCR and add comments * Make sure previous python example works * Make sure previous python example works * Fix Rec model bug * Fix Rec model bug * Fix rec model bug * Add SetTrtMaxBatchSize function for TensorRT * Add SetTrtMaxBatchSize Pybind * Add set_trt_max_batch_size python function * Set TRT dynamic shape in PPOCR examples * Set TRT dynamic shape in PPOCR examples * Set TRT dynamic shape in PPOCR examples * Fix PPOCRv2 python example * Fix PPOCR dynamic input shape bug * Remove useless code Co-authored-by: Jason <[email protected]>
xinzaifeixiang1992 · Nov 23, 2022 · 737c793 · 737c793
1 parent 712d7fd
commit 737c793
Show file tree

Hide file tree

Showing 4 changed files with 69 additions and 39 deletions.
diff --git a/examples/vision/ocr/PP-OCRv2/cpp/infer.cc b/examples/vision/ocr/PP-OCRv2/cpp/infer.cc
@@ -34,11 +34,12 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   auto rec_option = option;
 
   // If use TRT backend, the dynamic shape will be set as follow.
-  det_option.SetTrtInputShape("x", {1, 3, 50, 50}, {1, 3, 640, 640},
-                                {1, 3, 1536, 1536});
-  cls_option.SetTrtInputShape("x", {1, 3, 48, 10}, {1, 3, 48, 320}, {1, 3, 48, 1024});
-  rec_option.SetTrtInputShape("x", {1, 3, 32, 10}, {1, 3, 32, 320},
-                                {1, 3, 32, 2304});
+  // We recommend that users set the length and height of the detection model to a multiple of 32.
+  det_option.SetTrtInputShape("x", {1, 3, 64,64}, {1, 3, 640, 640},
+                                {1, 3, 960, 960});
+  cls_option.SetTrtInputShape("x", {1, 3, 48, 10}, {10, 3, 48, 320}, {64, 3, 48, 1024});
+  rec_option.SetTrtInputShape("x", {1, 3, 32, 10}, {10, 3, 32, 320},
+                                {64, 3, 32, 2304});
 
   // Users could save TRT cache file to disk as follow. 
   // det_option.SetTrtCacheFile(det_model_dir + sep + "det_trt_cache.trt");
@@ -103,6 +104,11 @@ int main(int argc, char* argv[]) {
   } else if (flag == 2) {
     option.UseGpu();
     option.UseTrtBackend();
+  } else if (flag == 3) {
+    option.UseGpu();
+    option.UseTrtBackend();
+    option.EnablePaddleTrtCollectShape();
+    option.EnablePaddleToTrt();
   }
 
   std::string det_model_dir = argv[1];

diff --git a/examples/vision/ocr/PP-OCRv2/python/infer.py b/examples/vision/ocr/PP-OCRv2/python/infer.py
@@ -72,6 +72,12 @@ def build_option(args):
         assert args.device.lower(
         ) == "gpu", "TensorRT backend require inference on device GPU."
         option.use_trt_backend()
+    elif args.backend.lower() == "pptrt":
+        assert args.device.lower(
+        ) == "gpu", "Paddle-TensorRT backend require inference on device GPU."
+        option.use_trt_backend()
+        option.enable_paddle_trt_collect_shape()
+        option.enable_paddle_to_trt()
     elif args.backend.lower() == "ort":
         option.use_ort_backend()
     elif args.backend.lower() == "paddle":
@@ -100,27 +106,30 @@ def build_option(args):
 # 用户也可根据自行需求分别配置
 runtime_option = build_option(args)
 
+# 当使用TRT时，分别给三个模型的runtime设置动态shape,并完成模型的创建.
+# 注意: 需要在检测模型创建完成后，再设置分类模型的动态输入并创建分类模型, 识别模型同理.
+# 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
 det_option = runtime_option
-cls_option = runtime_option
-rec_option = runtime_option
-
-# 当使用TRT时，分别给三个Runtime设置动态shape
-det_option.set_trt_input_shape("x", [1, 3, 50, 50], [1, 3, 640, 640],
-                               [1, 3, 1536, 1536])
-cls_option.set_trt_input_shape("x", [1, 3, 48, 10], [1, 3, 48, 320],
-                               [1, 3, 48, 1024])
-rec_option.set_trt_input_shape("x", [1, 3, 32, 10], [1, 3, 32, 320],
-                               [1, 3, 32, 2304])
-
+det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                               [1, 3, 960, 960])
 # 用户可以把TRT引擎文件保存至本地
 # det_option.set_trt_cache_file(args.det_model  + "/det_trt_cache.trt")
-# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
-# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
-
 det_model = fd.vision.ocr.DBDetector(
     det_model_file, det_params_file, runtime_option=det_option)
+
+cls_option = runtime_option
+cls_option.set_trt_input_shape("x", [1, 3, 48, 10], [10, 3, 48, 320],
+                               [64, 3, 48, 1024])
+# 用户可以把TRT引擎文件保存至本地
+# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
 cls_model = fd.vision.ocr.Classifier(
     cls_model_file, cls_params_file, runtime_option=cls_option)
+
+rec_option = runtime_option
+rec_option.set_trt_input_shape("x", [1, 3, 32, 10], [10, 3, 32, 320],
+                               [64, 3, 32, 2304])
+# 用户可以把TRT引擎文件保存至本地
+# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
 rec_model = fd.vision.ocr.Recognizer(
     rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)
 

diff --git a/examples/vision/ocr/PP-OCRv3/cpp/infer.cc b/examples/vision/ocr/PP-OCRv3/cpp/infer.cc
@@ -33,12 +33,13 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   auto cls_option = option;
   auto rec_option = option;
 
-  // If use TRT backend, the dynamic shape will be set as follow. 
-  det_option.SetTrtInputShape("x", {1, 3, 50, 50}, {1, 3, 640, 640},
-                                {1, 3, 1536, 1536});
-  cls_option.SetTrtInputShape("x", {1, 3, 48, 10}, {1, 3, 48, 320}, {1, 3, 48, 1024});
-  rec_option.SetTrtInputShape("x", {1, 3, 48, 10}, {1, 3, 48, 320},
-                                {1, 3, 48, 2304});
+  // If use TRT backend, the dynamic shape will be set as follow.
+  // We recommend that users set the length and height of the detection model to a multiple of 32.
+  det_option.SetTrtInputShape("x", {1, 3, 64,64}, {1, 3, 640, 640},
+                                {1, 3, 960, 960});
+  cls_option.SetTrtInputShape("x", {1, 3, 48, 10}, {10, 3, 48, 320}, {64, 3, 48, 1024});
+  rec_option.SetTrtInputShape("x", {1, 3, 48, 10}, {10, 3, 48, 320},
+                                {64, 3, 48, 2304});
 
   // Users could save TRT cache file to disk as follow. 
   // det_option.SetTrtCacheFile(det_model_dir + sep + "det_trt_cache.trt");
@@ -103,6 +104,11 @@ int main(int argc, char* argv[]) {
   } else if (flag == 2) {
     option.UseGpu();
     option.UseTrtBackend();
+  } else if (flag == 3) {
+    option.UseGpu();
+    option.UseTrtBackend();
+    option.EnablePaddleTrtCollectShape();
+    option.EnablePaddleToTrt();
   }
 
   std::string det_model_dir = argv[1];

diff --git a/examples/vision/ocr/PP-OCRv3/python/infer.py b/examples/vision/ocr/PP-OCRv3/python/infer.py
@@ -72,6 +72,12 @@ def build_option(args):
         assert args.device.lower(
         ) == "gpu", "TensorRT backend require inference on device GPU."
         option.use_trt_backend()
+    elif args.backend.lower() == "pptrt":
+        assert args.device.lower(
+        ) == "gpu", "Paddle-TensorRT backend require inference on device GPU."
+        option.use_trt_backend()
+        option.enable_paddle_trt_collect_shape()
+        option.enable_paddle_to_trt()
     elif args.backend.lower() == "ort":
         option.use_ort_backend()
     elif args.backend.lower() == "paddle":
@@ -100,27 +106,30 @@ def build_option(args):
 # 用户也可根据自行需求分别配置
 runtime_option = build_option(args)
 
+# 当使用TRT时，分别给三个模型的runtime设置动态shape,并完成模型的创建.
+# 注意: 需要在检测模型创建完成后，再设置分类模型的动态输入并创建分类模型, 识别模型同理.
+# 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
 det_option = runtime_option
-cls_option = runtime_option
-rec_option = runtime_option
-
-# 当使用TRT时，分别给三个Runtime设置动态shape
-det_option.set_trt_input_shape("x", [1, 3, 50, 50], [1, 3, 640, 640],
-                               [1, 3, 1536, 1536])
-cls_option.set_trt_input_shape("x", [1, 3, 48, 10], [1, 3, 48, 320],
-                               [1, 3, 48, 1024])
-rec_option.set_trt_input_shape("x", [1, 3, 48, 10], [1, 3, 48, 320],
-                               [1, 3, 48, 2304])
-
+det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                               [1, 3, 960, 960])
 # 用户可以把TRT引擎文件保存至本地
 # det_option.set_trt_cache_file(args.det_model  + "/det_trt_cache.trt")
-# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
-# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
-
 det_model = fd.vision.ocr.DBDetector(
     det_model_file, det_params_file, runtime_option=det_option)
+
+cls_option = runtime_option
+cls_option.set_trt_input_shape("x", [1, 3, 48, 10], [10, 3, 48, 320],
+                               [64, 3, 48, 1024])
+# 用户可以把TRT引擎文件保存至本地
+# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
 cls_model = fd.vision.ocr.Classifier(
     cls_model_file, cls_params_file, runtime_option=cls_option)
+
+rec_option = runtime_option
+rec_option.set_trt_input_shape("x", [1, 3, 48, 10], [10, 3, 48, 320],
+                               [64, 3, 48, 2304])
+# 用户可以把TRT引擎文件保存至本地
+# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
 rec_model = fd.vision.ocr.Recognizer(
     rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)