Merge pull request #1648 from Intsigstephon/develop

support cls inference for onnx mode

deploy/paddle2onnx/readme.md

+# paddle2onnx 模型转化与预测
+
+本章节介绍 ResNet50_vd 模型如何转化为 ONNX 模型，并基于 ONNX 引擎预测。
+
+## 1. 环境准备
+
+需要准备 Paddle2ONNX 模型转化环境，和 ONNX 模型预测环境。
+
+Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式，算子目前稳定支持导出 ONNX Opset 9~11，部分Paddle算子支持更低的ONNX Opset转换。
+更多细节可参考 [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX/blob/develop/README_zh.md)
+
+- 安装 Paddle2ONNX
+```
+python3.7 -m pip install paddle2onnx
+```
+
+- 安装 ONNX 运行时
+```
+python3.7 -m pip install onnxruntime
+```
+
+## 2. 模型转换
+
+- ResNet50_vd inference模型下载
+
+```
+cd deploy
+mkdir models && cd models
+wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNet50_vd_infer.tar && tar xf ResNet50_vd_infer.tar
+cd ..
+```
+
+- 模型转换
+
+使用 Paddle2ONNX 将 Paddle 静态图模型转换为 ONNX 模型格式：
+```
+paddle2onnx --model_dir=./models/ResNet50_vd_infer/ \
+--model_filename=inference.pdmodel \
+--params_filename=inference.pdiparams \
+--save_file=./models/ResNet50_vd_infer/inference.onnx \
+--opset_version=10 \
+--enable_onnx_checker=True
+```
+
+执行完毕后，ONNX 模型 `inference.onnx` 会被保存在 `./models/ResNet50_vd_infer/` 路径下
+
+## 3. onnx 预测
+
+执行如下命令：
+```
+python3.7 python/predict_cls.py \
+-c configs/inference_cls.yaml \
+-o Global.use_onnx=True \
+-o Global.use_gpu=False \
+-o Global.inference_model_dir=./models/ResNet50_vd_infer \
+```
+
+结果如下：
+```
+ILSVRC2012_val_00000010.jpeg:   class id(s): [153, 204, 229, 332, 155], score(s): [0.69, 0.10, 0.02, 0.01, 0.01], label_name(s): ['Maltese dog, Maltese terrier, Maltese', 'Lhasa, Lhasa apso', 'Old English sheepdog, bobtail', 'Angora, Angora rabbit', 'Shih-Tzu']
+```

deploy/python/predict_cls.py

@@ -67,12 +67,17 @@ class ClsPredictor(Predictor):
                 warmup=2)
 
     def predict(self, images):
-        input_names = self.paddle_predictor.get_input_names()
-        input_tensor = self.paddle_predictor.get_input_handle(input_names[0])
+        use_onnx = self.args.get("use_onnx", False)
+        if not use_onnx:
+            input_names = self.predictor.get_input_names()
+            input_tensor = self.predictor.get_input_handle(input_names[0])
+
+            output_names = self.predictor.get_output_names()
+            output_tensor = self.predictor.get_output_handle(output_names[0])
+        else:
+            input_names = self.predictor.get_inputs()[0].name
+            output_names = self.predictor.get_outputs()[0].name
 
-        output_names = self.paddle_predictor.get_output_names()
-        output_tensor = self.paddle_predictor.get_output_handle(output_names[
-            0])
         if self.benchmark:
             self.auto_logger.times.start()
         if not isinstance(images, (list, )):
@@ -84,9 +89,15 @@ class ClsPredictor(Predictor):
         if self.benchmark:
             self.auto_logger.times.stamp()
 
+        if not use_onnx:
             input_tensor.copy_from_cpu(image)
-        self.paddle_predictor.run()
+            self.predictor.run()
             batch_output = output_tensor.copy_to_cpu()
+        else:
+            batch_output = self.predictor.run(
+                output_names=[output_names],
+                input_feed={input_names: image})[0]
+
         if self.benchmark:
             self.auto_logger.times.stamp()
         if self.postprocess is not None:

deploy/python/predict_det.py

@@ -109,17 +109,16 @@ class DetPredictor(Predictor):
         '''
         inputs = self.preprocess(image)
         np_boxes = None
-        input_names = self.paddle_predictor.get_input_names()
+        input_names = self.predictor.get_input_names()
 
         for i in range(len(input_names)):
-            input_tensor = self.paddle_predictor.get_input_handle(input_names[
-                i])
+            input_tensor = self.predictor.get_input_handle(input_names[i])
             input_tensor.copy_from_cpu(inputs[input_names[i]])
 
         t1 = time.time()
-        self.paddle_predictor.run()
-        output_names = self.paddle_predictor.get_output_names()
-        boxes_tensor = self.paddle_predictor.get_output_handle(output_names[0])
+        self.predictor.run()
+        output_names = self.predictor.get_output_names()
+        boxes_tensor = self.predictor.get_output_handle(output_names[0])
         np_boxes = boxes_tensor.copy_to_cpu()
         t2 = time.time()
 

deploy/python/predict_rec.py

@@ -58,12 +58,16 @@ class RecPredictor(Predictor):
                 warmup=2)
 
     def predict(self, images, feature_normalize=True):
-        input_names = self.paddle_predictor.get_input_names()
-        input_tensor = self.paddle_predictor.get_input_handle(input_names[0])
+        use_onnx = self.args.get("use_onnx", False)
+        if not use_onnx:
+            input_names = self.predictor.get_input_names()
+            input_tensor = self.predictor.get_input_handle(input_names[0])
 
-        output_names = self.paddle_predictor.get_output_names()
-        output_tensor = self.paddle_predictor.get_output_handle(output_names[
-            0])
+            output_names = self.predictor.get_output_names()
+            output_tensor = self.predictor.get_output_handle(output_names[0])
+        else:
+            input_names = self.predictor.get_inputs()[0].name
+            output_names = self.predictor.get_outputs()[0].name
 
         if self.benchmark:
             self.auto_logger.times.start()
@@ -76,9 +80,15 @@ class RecPredictor(Predictor):
         if self.benchmark:
             self.auto_logger.times.stamp()
 
+        if not use_onnx:
             input_tensor.copy_from_cpu(image)
-        self.paddle_predictor.run()
+            self.predictor.run()
             batch_output = output_tensor.copy_to_cpu()
+        else:
+            batch_output = self.predictor.run(
+                output_names=[output_names],
+                input_feed={input_names: image})[0]
+
         if self.benchmark:
             self.auto_logger.times.stamp()
 

deploy/utils/predictor.py

@@ -28,7 +28,11 @@ class Predictor(object):
         if args.use_fp16 is True:
             assert args.use_tensorrt is True
         self.args = args
-        self.paddle_predictor, self.config = self.create_paddle_predictor(
+        if self.args.get("use_onnx", False):
+            self.predictor, self.config = self.create_onnx_predictor(
+                args, inference_model_dir)
+        else:
+            self.predictor, self.config = self.create_paddle_predictor(
                 args, inference_model_dir)
 
     def predict(self, image):
@@ -69,3 +73,20 @@ class Predictor(object):
         predictor = create_predictor(config)
 
         return predictor, config
+
+    def create_onnx_predictor(self, args, inference_model_dir=None):
+        import onnxruntime as ort
+        if inference_model_dir is None:
+            inference_model_dir = args.inference_model_dir
+        model_file = os.path.join(inference_model_dir, "inference.onnx")
+        config = ort.SessionOptions()
+        if args.use_gpu:
+            raise ValueError(
+                "onnx inference now only supports cpu! please specify use_gpu false."
+            )
+        else:
+            config.intra_op_num_threads = args.cpu_num_threads
+            if args.ir_optim:
+                config.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+        predictor = ort.InferenceSession(model_file, sess_options=config)
+        return predictor, config

test_tipc/config/ResNet/ResNet50_vd_linux_gpu_normal_normal_paddle2onnx_python_linux_cpu.txt

+===========================paddle2onnx_params===========================
+model_name:ResNet50_vd
+python:python3.7
+2onnx: paddle2onnx
+--model_dir:./deploy/models/ResNet50_vd_infer/
+--model_filename:inference.pdmodel
+--params_filename:inference.pdiparams
+--save_file:./deploy/models/ResNet50_vd_infer/inference.onnx
+--opset_version:10
+--enable_onnx_checker:True
+inference: python/predict_cls.py  -c configs/inference_cls.yaml
+Global.use_onnx:True
+Global.inference_model_dir:models/ResNet50_vd_infer/
+Global.use_gpu:False

test_tipc/prepare.sh

@@ -165,3 +165,15 @@ if [ ${MODE} = "serving_infer" ];then
     cd ./deploy/paddleserving
     wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNet50_vd_infer.tar && tar xf ResNet50_vd_infer.tar
 fi
+
+if [ ${MODE} = "paddle2onnx_infer" ];then
+    # prepare paddle2onnx env
+    python_name=$(func_parser_value "${lines[2]}")
+    ${python_name} -m pip install install paddle2onnx
+    ${python_name} -m pip install onnxruntime
+
+    # wget model
+    cd deploy && mkdir models && cd models
+    wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNet50_vd_infer.tar  && tar xf ResNet50_vd_infer.tar
+    cd ../../
+fi

test_tipc/test_paddle2onnx.sh

@@ -11,7 +11,7 @@ python=$(func_parser_value "${lines[2]}")
 
 
 # parser params
-dataline=$(awk 'NR==1, NR==12{print}'  $FILENAME)
+dataline=$(awk 'NR==1, NR==14{print}'  $FILENAME)
 IFS=$'\n'
 lines=(${dataline})
 
@@ -33,12 +33,12 @@ enable_onnx_checker_key=$(func_parser_key "${lines[9]}")
 enable_onnx_checker_value=$(func_parser_value "${lines[9]}")
 # parser onnx inference 
 inference_py=$(func_parser_value "${lines[10]}")
-use_gpu_key=$(func_parser_key "${lines[11]}")
-use_gpu_value=$(func_parser_value "${lines[11]}")
-det_model_key=$(func_parser_key "${lines[12]}")
-image_dir_key=$(func_parser_key "${lines[13]}")
-image_dir_value=$(func_parser_value "${lines[13]}")
-
+use_onnx_key=$(func_parser_key "${lines[11]}")
+use_onnx_value=$(func_parser_value "${lines[11]}")
+inference_model_dir_key=$(func_parser_key "${lines[12]}")
+inference_model_dir_value=$(func_parser_value "${lines[12]}")
+inference_hardware_key=$(func_parser_key "${lines[13]}")
+inference_hardware_value=$(func_parser_value "${lines[13]}")
 
 LOG_PATH="./test_tipc/output"
 mkdir -p ./test_tipc/output
@@ -50,7 +50,7 @@ function func_paddle2onnx(){
     _script=$1
 
     # paddle2onnx
-    _save_log_path="${LOG_PATH}/paddle2onnx_infer_cpu.log"
+    _save_log_path=".${LOG_PATH}/paddle2onnx_infer_cpu.log"
     set_dirname=$(func_set_params "${infer_model_dir_key}" "${infer_model_dir_value}")
     set_model_filename=$(func_set_params "${model_filename_key}" "${model_filename_value}")
     set_params_filename=$(func_set_params "${params_filename_key}" "${params_filename_value}")
@@ -62,10 +62,10 @@ function func_paddle2onnx(){
     last_status=${PIPESTATUS[0]}
     status_check $last_status "${trans_model_cmd}" "${status_log}"
     # python inference
-    set_gpu=$(func_set_params "${use_gpu_key}" "${use_gpu_value}")
-    set_model_dir=$(func_set_params "${det_model_key}" "${save_file_value}")
-    set_img_dir=$(func_set_params "${image_dir_key}" "${image_dir_value}")
-    infer_model_cmd="${python} ${inference_py} ${set_gpu} ${set_img_dir} ${set_model_dir} --use_onnx=True > ${_save_log_path} 2>&1 "
+    set_model_dir=$(func_set_params "${inference_model_dir_key}" "${inference_model_dir_value}")
+    set_use_onnx=$(func_set_params "${use_onnx_key}" "${use_onnx_value}")
+    set_hardware=$(func_set_params "${inference_hardware_key}" "${inference_hardware_value}")
+    infer_model_cmd="cd deploy && ${python} ${inference_py} -o ${set_model_dir} -o ${set_use_onnx} -o ${set_hardware} >${_save_log_path} 2>&1 && cd ../"
     eval $infer_model_cmd
     status_check $last_status "${infer_model_cmd}" "${status_log}"
 }