Update API usage according to 1.8 recommendations

PaddleCV/rrpn/README.md

@@ -12,7 +12,7 @@
 
 ## 安装
 
-在当前目录下运行样例代码需要PadddlePaddle Fluid的develop或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/)中的说明来更新PaddlePaddle。
+在当前目录下运行样例代码需要PadddlePaddle Fluid的1.8.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/)中的说明来更新PaddlePaddle。
 
 
 ## 简介
@@ -27,16 +27,23 @@ RRPN是在Faster RCNN基础上拓展出的两阶段目标检测器，可用于
 
 ### 编译自定义OP
 
+**注意：** 通过pip方式安装的PaddlePaddle由GCC 4.8编译得到，由于GCC 4.8和GCC 5以上C++11 ABI不兼容，您编写的自定义OP，需要通过GCC 4.8编译。若是GCC 5及以上的环境上使用自定义OP，推荐使用Docker安装PaddlePaddle，使得编Paddle和编译自定义OP的GCC版本相同。
+
 自定义OP编译方式如下：
 
     进入 `models/ext_op/src` 目录，执行编译脚本
     ```
     cd models/ext_op/src
     sh make.sh  ${cuda_path} ${cudnn_path} ${nccl_path}
-    '''
+    ```
     其中${cuda_path}、$cudnn_path}和{nccl_path}分别为cuda、cudnn、nccl的安装路径，需通过命令行进行指定
-    成功编译后，`ext_op/src` 目录下将会生成 `rrpn_lib.so` 
-
+    成功编译后，`ext_op/src` 目录下将会生成 `rrpn_lib.so`。
+    需要将`rrpn_lib.so`所在路径以及libpaddle_framework.so路径(即paddle.sysconfig.get_lib()得到路径)设置到环境变量LD_LIBRARY_PATH中:
+    ```
+    # 假如rrpn_lib.so路径是：`rrpn/models/ext_op/src/`，对于Linux环境设置:
+    export LD_LIBRARY_PATH=rrpn/models/ext_op/src/:$( python -c 'import paddle; print(paddle.sysconfig.get_lib())'):$LD_LIBRARY_PATH
+    ```
+
 ## 数据准备
 ### 公开数据集
 在[ICDAR2015数据集](https://rrc.cvc.uab.es/?ch=4&com=downloads)上进行训练，数据集需进入官网进行注册后方可下载。
@@ -58,8 +65,8 @@ dataset/icdar2015/
 │   ├── img_112.jpg
 |   ...
 ├── ch4_test_localization_transcription_gt
-│   ├── img_111.jpg
-│   ├── img_112.jpg
+│   ├── img_111.txt
+│   ├── img_112.txt
 |   ...
 ```
 ### 自定义数据
@@ -88,7 +95,7 @@ x1, y1, x2, y2, x3, y3, x4, y4, class_name
     python train.py \
        --model_save_dir=output/ \
        --pretrained_model=${path_to_pretrain_model} \
-       --data_dir=${path_to_data} \
+       --data_dir=${path_to_icdar2015} \
     ```
 
 
@@ -126,7 +133,7 @@ x1, y1, x2, y2, x3, y3, x4, y4, class_name
 
     ```
     python eval.py \
-        --dataset=icdar2015 \
+        --data_dir=${path_to_icdar2015} \
         --pretrained_model=${path_to_trained_model}
     ```
 
@@ -143,10 +150,6 @@ RRPN
 | [RRPN](https://paddleseg.bj.bcebos.com/deploy/temp/model_final.tar) |8   |    17500       | 0.8048 |
 
 
-
-
-
-
 ## 模型推断及可视化
 
 模型推断可以获取图像中的物体及其对应的类别，`infer.py`是主要执行程序，调用示例如下：

PaddleCV/rrpn/checkpoint.py

@@ -41,6 +41,13 @@ def _load_state(path):
     return state
 
 
+def _strip_postfix(path):
+    path, ext = os.path.splitext(path)
+    assert ext in ['', '.pdparams', '.pdopt', '.pdmodel'], \
+            "Unknown postfix {} from weights".format(ext)
+    return path
+
+
 def load_params(exe, prog, path):
     """
     Load model from the given path.
@@ -50,20 +57,33 @@ def load_params(exe, prog, path):
         path (string): URL string or loca model path.
     """
 
-    if not os.path.exists(path):
+    path = _strip_postfix(path)
+    if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')):
         raise ValueError("Model pretrain path {} does not "
                          "exists.".format(path))
 
     logger.info('Loading parameters from {}...'.format(path))
 
-    def _if_exist(var):
-        param_exist = os.path.exists(os.path.join(path, var.name))
-        do_load = param_exist
-        if do_load:
-            logger.debug('load weight {}'.format(var.name))
-        return do_load
+    ignore_set = set()
+    state = _load_state(path)
 
-    fluid.io.load_vars(exe, path, prog, predicate=_if_exist)
+    # ignore the parameter which mismatch the shape 
+    # between the model and pretrain weight.
+    all_var_shape = {}
+    for block in prog.blocks:
+        for param in block.all_parameters():
+            all_var_shape[param.name] = param.shape
+    ignore_set.update([
+        name for name, shape in all_var_shape.items()
+        if name in state and shape != state[name].shape
+    ])
+
+    if len(ignore_set) > 0:
+        for k in ignore_set:
+            if k in state:
+                logger.warning('variable {} not used'.format(k))
+                del state[k]
+    fluid.io.set_program_state(prog, state)
 
 
 def save(exe, prog, path):
@@ -83,6 +103,7 @@ def save(exe, prog, path):
 def load_and_fusebn(exe, prog, path):
     """
     Fuse params of batch norm to scale and bias.
+
     Args:
         exe (fluid.Executor): The fluid.Executor object.
         prog (fluid.Program): save weight from which Program object.
@@ -104,19 +125,12 @@ def load_and_fusebn(exe, prog, path):
     #  x is any prefix
     mean_variances = set()
     bn_vars = []
-
-    state = None
-    if os.path.exists(path + '.pdparams'):
-        state = _load_state(path)
+    state = _load_state(path)
 
     def check_mean_and_bias(prefix):
         m = prefix + 'mean'
         v = prefix + 'variance'
-        if state:
-            return v in state and m in state
-        else:
-            return (os.path.exists(os.path.join(path, m)) and
-                    os.path.exists(os.path.join(path, v)))
+        return v in state and m in state
 
     has_mean_bias = True
 
@@ -156,16 +170,14 @@ def check_mean_and_bias(prefix):
                     bn_vars.append(
                         [scale_name, bias_name, mean_name, variance_name])
 
-    if state:
-        fluid.io.set_program_state(prog, state)
-    else:
-        load_params(exe, prog, path)
     if not has_mean_bias:
+        fluid.io.set_program_state(prog, state)
         logger.warning(
             "There is no paramters of batch norm in model {}. "
             "Skip to fuse batch norm. And load paramters done.".format(path))
         return
 
+    fluid.load(prog, path, exe)
     eps = 1e-5
     for names in bn_vars:
         scale_name, bias_name, mean_name, var_name = names

PaddleCV/rrpn/models/ext_op/rrpn_lib.py

@@ -15,7 +15,7 @@
 import paddle.fluid as fluid
 from paddle.fluid.layer_helper import LayerHelper
 from paddle.fluid.framework import Variable
-fluid.load_op_library('models/ext_op/src/rrpn_lib.so')
+fluid.load_op_library('rrpn_lib.so')
 
 
 def rrpn_target_assign(bbox_pred,

PaddleCV/rrpn/models/ext_op/src/make.sh

@@ -27,7 +27,7 @@ git clone https://github.com/NVlabs/cub.git
 
 nvcc rrpn_generate_proposals_op.cu -c -o rrpn_generate_proposals_op.cu.o -ccbin cc -DPADDLE_WITH_MKLDNN -DPADDLE_WITH_CUDA -DEIGEN_USE_GPU -DPADDLE_USE_DSO -Xcompiler -fPIC -std=c++11 -Xcompiler -fPIC -w --expt-relaxed-constexpr -O3 -DNVCC \
     -I ${include_dir} \
-   -I ${include_dir}/third_party \
+    -I ${include_dir}/third_party \
     -I ${CUDA}/include \
     -I ${CUDNN}/include \
     -I ${NCCL}/include \

PaddleCV/rrpn/models/ext_op/src/rrpn_rotated_roi_align_op.cc

@@ -165,8 +165,7 @@ class RRPNRotatedROIAlignGradMaker : public framework::SingleGradOpMaker<T> {
   using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
 
 protected:
-  std::unique_ptr<T> Apply() const override {
-    std::unique_ptr<T> op(new T);
+  void Apply(GradOpPtr<T> op) const override {
     op->SetType("rrpn_rotated_roi_align_grad");
     op->SetInput("X", this->Input("X"));
     op->SetInput("ROIs", this->Input("ROIs"));
@@ -175,12 +174,11 @@ class RRPNRotatedROIAlignGradMaker : public framework::SingleGradOpMaker<T> {
     op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
     op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
     op->SetAttrMap(this->Attrs());
-    return op;
   }
 };
 
-DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(
-    RRPNRotatedRoiAlignGradNoNeedBufVarsInferer, "X");
+DECLARE_NO_NEED_BUFFER_VARS_INFERER(RRPNRotatedRoiAlignGradNoNeedBufVarsInferer,
+                                    "X");
 
 }  // namespace operators
 }  // namespace paddle

PaddleCV/rrpn/reader.py

@@ -96,7 +96,6 @@ def reader():
                     continue
                 batch_out.append(datas)
                 end = time.time()
-                #print('reader time:', end - start)
                 if len(batch_out) == batch_size:
                     yield batch_out
                     count += 1

PaddleCV/rrpn/roidbs.py

@@ -101,7 +101,6 @@ def get_roidb(self):
                     elif edge2 >= edge1:
                         width = edge2
                         height = edge1
-                        # print pt2[0], pt3[0]
                         if pt2[0] - pt3[0] != 0:
                             angle = -np.arctan(
                                 float(pt2[1] - pt3[1]) /
@@ -160,7 +159,6 @@ def get_roidb(self):
                     else:
                         hard_boxes.append([x_ctr, y_ctr, width, height, angle])
 
-            #print(easy_boxes)
             if self.mode == 'train':
                 boxes.extend(easy_boxes)
                 # hard box only get 1/3 for train
@@ -173,8 +171,6 @@ def get_roidb(self):
                 is_difficult = [0] * len(easy_boxes)
                 is_difficult.extend([1] * int(len(hard_boxes)))
             len_of_bboxes = len(boxes)
-            #is_difficult = [0] * len(easy_boxes)
-            #is_difficult.extend([1] * int(len(hard_boxes)))
             is_difficult = np.array(is_difficult).reshape(
                 1, len_of_bboxes).astype(np.int32)
             if self.mode == 'train':
@@ -221,11 +217,9 @@ class ICDAR2017Dataset(object):
     def __init__(self, mode):
         print('Creating: {}'.format(cfg.dataset))
         self.name = cfg.data_dir
-        #print('**************', self.name)
         self.mode = mode
         data_path = DatasetPath(mode, self.name)
         data_dir = data_path.get_data_dir()
-        #print("&**************", data_dir)
         file_list = data_path.get_file_list()
         self.image_dir = data_dir
         self.gt_dir = file_list
@@ -245,15 +239,12 @@ def get_roidb(self):
             labels_map = get_labels_maps()
         for image in image_list:
             prefix = image[:-4]
-            #print(image)
 
             if image.split('.')[-1] not in post_fix:
                 continue
             img_name = os.path.join(self.image_dir, image)
             gt_name = os.path.join(self.gt_dir, 'gt_' + prefix + '.txt')
             gt_classes = []
-            #boxes = []
-            #hard_boxes = []
             boxes = []
             gt_obj = open(gt_name, 'r', encoding='UTF-8-sig')
             gt_txt = gt_obj.read()
@@ -293,7 +284,6 @@ def get_roidb(self):
                     elif edge2 >= edge1:
                         width = edge2
                         height = edge1
-                        # print pt2[0], pt3[0]
                         if pt2[0] - pt3[0] != 0:
                             angle = -np.arctan(
                                 float(pt2[1] - pt3[1]) /
@@ -312,7 +302,6 @@ def get_roidb(self):
                     else:
                         boxes.append([x_ctr, y_ctr, width, height, angle])
             len_of_bboxes = len(boxes)
-            #print(len_of_bboxes)
             is_difficult = np.zeros((len_of_bboxes, 1), dtype=np.int32)
             if self.mode == 'train':
                 gt_boxes = np.zeros((len_of_bboxes, 5), dtype=np.int32)
@@ -332,7 +321,6 @@ def get_roidb(self):
                         boxes[idx][3], boxes[idx][4], boxes[idx][5],
                         boxes[idx][6], boxes[idx][7]
                     ]
-                #gt_classes[idx] = 1
             if gt_boxes.shape[0] <= 0:
                 continue
             gt_boxes = gt_boxes.astype(np.float64)

PaddleCV/rrpn/utility.py

@@ -154,7 +154,7 @@ def parse_args():
     add_arg('pixel_means',     float,   [0.485, 0.456, 0.406], "pixel mean")
     add_arg('nms_thresh',    float, 0.3,    "NMS threshold.")
     add_arg('score_thresh',    float, 0.01,    "score threshold for NMS.")
-    add_arg('snapshot_stride',  int,    1000,    "save model every snapshot stride.")
+    add_arg('snapshot_iter',  int,    1000,    "save model every snapshot iter.")
     # SINGLE EVAL AND DRAW
     add_arg('draw_threshold',  float, 0.8,    "Confidence threshold to draw bbox.")
     add_arg('image_path',       str,   'ICDAR2015/tmp/',  "The image path used to inference and visualize.")