[Dygraph] Fix deploy python/cpp infer bugs (#2234)

* use scripts to build on windows flatform, test=dygraph

* Use new preproces OPs in deploy/infer.py, test=dygraph

* fix bug of image_shape can not set 2-dims, test=dygraph

* fix cpp build, test=dygraph

* set param target_size must be [h, w] in Resize function, test=dygraph

* fix bug, test=dygraph

* fix ssd output format (bbox: [N, 6], bbox_num: bbox_num), test=dygraph

* fix ssd prediction in cpp mode, test=dygraph

dygraph/deploy/cpp/docs/windows_vs2019_build.md

@@ -47,24 +47,14 @@ fluid_inference
 
 ### Step4: 编译
 
-#### 通过图形化操作编译CMake
-
-1. 打开Visual Studio 2019 Community，点击`继续但无需代码`
-![step2](https://paddleseg.bj.bcebos.com/inference/vs2019_step1.png)
-2. 点击： `文件`->`打开`->`CMake`
-![step2.1](https://paddleseg.bj.bcebos.com/inference/vs2019_step2.png)
-
-选择项目代码所在路径，并打开`CMakeList.txt`：
-
-![step2.2](https://paddleseg.bj.bcebos.com/inference/vs2019_step3.png)
-
-3. 点击：`项目`->`cpp_inference_demo的CMake设置`
-
-![step3](https://paddleseg.bj.bcebos.com/inference/vs2019_step4.png)
+1. 进入到`cpp`文件夹
+```
+cd D:\projects\PaddleDetection\deploy\cpp
+```
 
-4. 点击`浏览`，分别设置编译选项指定`CUDA`、`CUDNN_LIB`、`OpenCV`、`Paddle预测库`的路径
+2. 使用CMake生成项目文件
 
-三个编译参数的含义说明如下（带*表示仅在使用**GPU版本**预测库时指定, 其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1等版本CUDA库**）：
+编译参数的含义说明如下（带*表示仅在使用**GPU版本**预测库时指定, 其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1等版本CUDA库**）：
 
 |  参数名   | 含义  |
 |  ----  | ----  |
@@ -75,23 +65,8 @@ fluid_inference
 | USE_PADDLE_20RC1 | 是否使用2.0rc1预测库。如果使用2.0rc1，在windows环境下预测库名称发生变化，且仅支持动态库方式编译 |
 
 **注意：** 1. 使用`CPU`版预测库，请把`WITH_GPU`的勾去掉 2. 如果使用的是`openblas`版本，请把`WITH_MKL`勾去掉
-![step4](https://paddleseg.bj.bcebos.com/inference/vs2019_step5.png)
-
-**设置完成后**, 点击上图中`保存并生成CMake缓存以加载变量`。
 
-5. 点击`生成`->`全部生成`
-
-![step6](https://paddleseg.bj.bcebos.com/inference/vs2019_step6.png)
-
-
-#### 通过命令行操作编译CMake
-
-1. 进入到`cpp`文件夹
-```
-cd D:\projects\PaddleDetection\deploy\cpp
-```
-
-2. 使用CMake生成项目文件
+执行如下命令项目文件：
 ```
 cmake . -G "Visual Studio 16 2019" -A x64 -T host=x64 -DWITH_GPU=ON -DWITH_MKL=ON -DCMAKE_BUILD_TYPE=Release -DCUDA_LIB=path_to_cuda_lib -DCUDNN_LIB=path_to_cudnn_lib -DPADDLE_DIR=path_to_paddle_lib -DOPENCV_DIR=path_to_opencv
 ```
@@ -102,7 +77,8 @@ cmake . -G "Visual Studio 16 2019" -A x64 -T host=x64 -DWITH_GPU=ON -DWITH_MKL=O
 ```
 
 3. 编译
-用`Visual Studio 16 2019`打开`cpp`文件夹下的`PaddleObjectDetector.sln`，点击`生成`->`全部生成`
+用`Visual Studio 16 2019`打开`cpp`文件夹下的`PaddleObjectDetector.sln`，将编译模式设置为`Release`，点击`生成`->`全部生成
+
 
 ### Step5: 预测及可视化
 

dygraph/deploy/cpp/include/preprocess_op.h

@@ -58,7 +58,7 @@ class InitInfo : public PreprocessOp{
   virtual void Run(cv::Mat* im, ImageBlob* data);
 };
 
-class Normalize : public PreprocessOp {
+class NormalizeImage : public PreprocessOp {
  public:
   virtual void Init(const YAML::Node& item, const std::vector<int> image_shape) {
     mean_ = item["mean"].as<std::vector<float>>();
@@ -133,13 +133,14 @@ class Preprocessor {
   }
 
   std::shared_ptr<PreprocessOp> CreateOp(const std::string& name) {
-    if (name == "ResizeOp") {
+    if (name == "Resize") {
       return std::make_shared<Resize>();
-    } else if (name == "PermuteOp") {
+    } else if (name == "Permute") {
       return std::make_shared<Permute>();
-    } else if (name == "NormalizeImageOp") {
-      return std::make_shared<Normalize>();
-    } else if (name == "PadBatchOp" || name == "PadStride") {
+    } else if (name == "NormalizeImage") {
+      return std::make_shared<NormalizeImage>();
+    } else if (name == "PadStride") {
+      // use PadStride instead of PadBatch
       return std::make_shared<PadStride>();
     }
     std::cerr << "can not find function of OP: " << name << " and return: nullptr" << std::endl;

dygraph/deploy/cpp/src/object_detector.cc

@@ -131,7 +131,7 @@ void ObjectDetector::Postprocess(
   result->clear();
   int rh = 1;
   int rw = 1;
-  if (config_.arch_ == "SSD" || config_.arch_ == "Face") {
+  if (config_.arch_ == "Face") {
     rh = raw_mat.rows;
     rw = raw_mat.cols;
   }

dygraph/deploy/cpp/src/preprocess_op.cc

@@ -31,7 +31,7 @@ void InitInfo::Run(cv::Mat* im, ImageBlob* data) {
   };
 }
 
-void Normalize::Run(cv::Mat* im, ImageBlob* data) {
+void NormalizeImage::Run(cv::Mat* im, ImageBlob* data) {
   double e = 1.0;
   if (is_scale_) {
     e /= 255.0;
@@ -62,6 +62,10 @@ void Permute::Run(cv::Mat* im, ImageBlob* data) {
 
 void Resize::Run(cv::Mat* im, ImageBlob* data) {
   auto resize_scale = GenerateScale(*im);
+  data->input_shape_ = {
+      static_cast<int>(im->cols * resize_scale.first),
+      static_cast<int>(im->rows * resize_scale.second)
+  };
   cv::resize(
       *im, *im, cv::Size(), resize_scale.first, resize_scale.second, interp_);
   data->im_shape_ = {
@@ -72,24 +76,6 @@ void Resize::Run(cv::Mat* im, ImageBlob* data) {
     resize_scale.second,
     resize_scale.first,
   };
-
-  if (keep_ratio_) {
-    int max_size = input_shape_[1];
-    // Padding the image with 0 border
-    cv::copyMakeBorder(
-      *im,
-      *im,
-      0,
-      max_size - im->rows,
-      0,
-      max_size - im->cols,
-      cv::BORDER_CONSTANT,
-      cv::Scalar(0));
-  }
-  data->input_shape_ = {
-    static_cast<int>(im->rows),
-    static_cast<int>(im->cols),
-  };
 }
 
 std::pair<float, float> Resize::GenerateScale(const cv::Mat& im) {
@@ -145,7 +131,7 @@ void PadStride::Run(cv::Mat* im, ImageBlob* data) {
 
 // Preprocessor op running order
 const std::vector<std::string> Preprocessor::RUN_ORDER = {
-  "InitInfo", "ResizeOp", "NormalizeImageOp", "PadStrideOp", "PermuteOp"
+  "InitInfo", "Resize", "NormalizeImage", "PadStride", "Permute"
 };
 
 void Preprocessor::Run(cv::Mat* im, ImageBlob* data) {

dygraph/deploy/python/preprocess.py

@@ -35,6 +35,7 @@ def decode_image(im_file, im_info):
     else:
         im = im_file
     im_info['im_shape'] = np.array(im.shape[:2], dtype=np.float32)
+    im_info['scale_factor'] = np.array([1., 1.], dtype=np.float32)
     return im, im_info
 
 
@@ -66,8 +67,13 @@ class Resize(object):
             im (np.ndarray):  processed image (np.ndarray)
             im_info (dict): info of processed image
         """
+        assert len(self.target_size) == 2
+        assert self.target_size[0] > 0 and self.target_size[1] > 0
         im_channel = im.shape[2]
         im_scale_y, im_scale_x = self.generate_scale(im)
+        # set image_shape
+        im_info['input_shape'][1] = int(im_scale_y * im.shape[0])
+        im_info['input_shape'][2] = int(im_scale_x * im.shape[1])
         im = cv2.resize(
             im,
             None,
@@ -78,14 +84,6 @@ class Resize(object):
         im_info['im_shape'] = np.array(im.shape[:2]).astype('float32')
         im_info['scale_factor'] = np.array(
             [im_scale_y, im_scale_x]).astype('float32')
-        # padding im when image_shape fixed by infer_cfg.yml
-        if self.keep_ratio and im_info['input_shape'][1] is not None:
-            max_size = im_info['input_shape'][1]
-            padding_im = np.zeros(
-                (max_size, max_size, im_channel), dtype=np.float32)
-            im_h, im_w = im.shape[:2]
-            padding_im[:im_h, :im_w, :] = im
-            im = padding_im
         return im, im_info
 
     def generate_scale(self, im):
@@ -174,7 +172,7 @@ class Permute(object):
 
 
 class PadStride(object):
-    """ padding image for model with FPN 
+    """ padding image for model with FPN , instead PadBatch(pad_to_stride, pad_gt) in original config
     Args:
         stride (bool): model with FPN need image shape % stride == 0 
     """

dygraph/ppdet/engine/export_utils.py

@@ -52,12 +52,6 @@ def _parse_reader(reader_cfg, dataset_cfg, metric, arch, image_shape):
     for st in sample_transforms[1:]:
         for key, value in st.items():
             p = {'type': key}
-            if key == 'Resize':
-                if value.get('keep_ratio',
-                             False) and image_shape[1] is not None:
-                    max_size = max(image_shape[1:])
-                    image_shape = [3, max_size, max_size]
-                    value['target_size'] = image_shape[1:]
             p.update(value)
             preprocess_list.append(p)
     batch_transforms = reader_cfg.get('batch_transforms', None)
@@ -65,9 +59,10 @@ def _parse_reader(reader_cfg, dataset_cfg, metric, arch, image_shape):
         methods = [list(bt.keys())[0] for bt in batch_transforms]
         for bt in batch_transforms:
             for key, value in bt.items():
+                # for deploy/infer, use PadStride(stride) instead PadBatch(pad_to_stride, pad_gt)
                 if key == 'PadBatch':
-                    preprocess_list.append({'type': 'PadStride'})
-                    preprocess_list[-1].update({
+                    preprocess_list.append({
+                        'type': 'PadStride',
                         'stride': value['pad_to_stride']
                     })
                     break

dygraph/ppdet/engine/trainer.py

@@ -360,8 +360,9 @@ class Trainer(object):
         if 'inputs_def' in self.cfg['TestReader']:
             inputs_def = self.cfg['TestReader']['inputs_def']
             image_shape = inputs_def.get('image_shape', None)
+        # set image_shape=[3, -1, -1] as default
         if image_shape is None:
-            image_shape = [3, None, None]
+            image_shape = [3, -1, -1]
 
         self.model.eval()
 

dygraph/ppdet/modeling/architectures/ssd.py

@@ -54,13 +54,8 @@ class SSD(BaseArch):
 
     def get_pred(self):
         bbox_pred, bbox_num = self._forward()
-        label = bbox_pred[:, 0]
-        score = bbox_pred[:, 1]
-        bbox = bbox_pred[:, 2:]
         output = {
-            'bbox': bbox,
-            'score': score,
-            'label': label,
-            'bbox_num': bbox_num
+            "bbox": bbox_pred,
+            "bbox_num": bbox_num,
         }
         return output