update new api for rrpn (#4296)

update new api for rrpn

PaddleCV/rrpn/eval.py

@@ -36,7 +36,6 @@ def eval():
 
     place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
     exe = fluid.Executor(place)
-    image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
     class_nums = cfg.class_num
     model = model_builder.RRPN(
         add_conv_body_func=resnet.ResNet(),
@@ -48,19 +47,14 @@ def eval():
     infer_prog = fluid.Program()
     with fluid.program_guard(infer_prog, startup_prog):
         with fluid.unique_name.guard():
-            model.build_model(image_shape)
+            model.build_model()
             pred_boxes = model.eval_bbox_out()
     infer_prog = infer_prog.clone(True)
     exe.run(startup_prog)
-
-    # yapf: disable
-    def if_exist(var):
-        return os.path.exists(os.path.join(cfg.pretrained_model, var.name))
-    if cfg.pretrained_model:
-        checkpoint.load_params(exe, infer_prog, cfg.pretrained_model)
-    # yapf: enable
+    fluid.load(infer_prog, cfg.pretrained_model, exe)
     test_reader = reader.test(1)
-    feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
+    data_loader = model.data_loader
+    data_loader.set_sample_list_generator(test_reader, places=place)
 
     fetch_list = [pred_boxes]
     res_list = []
@@ -68,11 +62,10 @@ def eval():
         'bbox', 'gt_box', 'gt_class', 'is_crowed', 'im_info', 'im_id',
         'is_difficult'
     ]
-    for i, data in enumerate(test_reader()):
-        im_info = [data[0][1]]
+    for i, data in enumerate(data_loader()):
         result = exe.run(infer_prog,
                          fetch_list=[v.name for v in fetch_list],
-                         feed=feeder.feed(data),
+                         feed=data,
                          return_numpy=False)
         pred_boxes_v = result[0]
         nmsed_out = pred_boxes_v

PaddleCV/rrpn/eval_helper.py

@@ -31,11 +31,11 @@ logger = logging.getLogger(__name__)
 
 def get_key_dict(out, data, key):
     res = {}
-    for i in range(len(key)):
-        if i == 0:
-            res[key[i]] = out
+    for name in key:
+        if name == 'bbox':
+            res[name] = np.array(out)
         else:
-            res[key[i]] = data[i]
+            res[name] = np.array(data[name])
     return res
 
 
@@ -167,7 +167,7 @@ def calculate_ap(rec, prec):
 def icdar_map(result, class_name, ovthresh):
     im_ids = []
     for res in result:
-        im_ids.append(res['im_id'])
+        im_ids.append(res['im_id'][0][0])
     recs = {}
 
     for i, im_id in enumerate(im_ids):
@@ -185,11 +185,11 @@ def icdar_map(result, class_name, ovthresh):
     confidence = []
     bbox = []
     for res in result:
-        im_info = res['im_info']
+        im_info = res['im_info'][0]
         pred_boxes = res['bbox']
         for box in pred_boxes:
             if box[0] == class_name:
-                image_ids.append(res['im_id'])
+                image_ids.append(res['im_id'][0][0])
                 confidence.append(box[1])
                 clipd_box = clip_box(box[2:].reshape(-1, 8), im_info)
                 bbox.append(clipd_box[0])
@@ -286,7 +286,7 @@ def icdar_box_eval(result, thresh):
     num_global_care_gt = 0
     num_global_care_det = 0
     for res in result:
-        im_info = res['im_info']
+        im_info = res['im_info'][0]
         h = im_info[1]
         w = im_info[2]
         gt_boxes = res['gt_box']

PaddleCV/rrpn/infer.py

@@ -32,7 +32,6 @@ from utility import print_arguments, parse_args, check_gpu
 def infer():
     place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
     exe = fluid.Executor(place)
-    image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
     class_nums = cfg.class_num
     model = model_builder.RRPN(
         add_conv_body_func=resnet.ResNet(),
@@ -43,31 +42,25 @@ def infer():
     infer_prog = fluid.Program()
     with fluid.program_guard(infer_prog, startup_prog):
         with fluid.unique_name.guard():
-            model.build_model(image_shape)
+            model.build_model()
             pred_boxes = model.eval_bbox_out()
     infer_prog = infer_prog.clone(True)
     exe.run(startup_prog)
-
-    # yapf: disable
-    def if_exist(var):
-        return os.path.exists(os.path.join(cfg.pretrained_model, var.name))
-    if cfg.pretrained_model:
-        checkpoint.load_params(exe, infer_prog, cfg.pretrained_model)
-    # yapf: enable
+    fluid.load(infer_prog, cfg.pretrained_model, exe)
     infer_reader = reader.infer(cfg.image_path)
-    feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
-
+    data_loader = model.data_loader
+    data_loader.set_sample_list_generator(infer_reader, places=place)
     fetch_list = [pred_boxes]
     imgs = os.listdir(cfg.image_path)
     imgs.sort()
 
-    for i, data in enumerate(infer_reader()):
+    for i, data in enumerate(data_loader()):
         result = exe.run(infer_prog,
                          fetch_list=[v.name for v in fetch_list],
-                         feed=feeder.feed(data),
+                         feed=data,
                          return_numpy=False)
         nmsed_out = result[0]
-        im_info = data[0][1]
+        im_info = np.array(data[0]['im_info'])[0]
         im_scale = im_info[2]
         outs = np.array(nmsed_out)
         draw_bounding_box_on_image(cfg.image_path, imgs[i], outs, im_scale,

PaddleCV/rrpn/models/model_builder.py

@@ -35,8 +35,8 @@ class RRPN(object):
         self.use_pyreader = use_pyreader
         self.use_random = use_random
 
-    def build_model(self, image_shape):
-        self.build_input(image_shape)
+    def build_model(self):
+        self.build_input()
         body_conv = self.add_conv_body_func(self.image)
         # RPN
         self.rpn_heads(body_conv)
@@ -61,56 +61,42 @@ class RRPN(object):
     def eval_bbox_out(self):
         return self.pred_result
 
-    def build_input(self, image_shape):
-        if self.use_pyreader:
-            in_shapes = [[-1] + image_shape, [-1, 5], [-1, 1], [-1, 1],
-                         [-1, 3], [-1, 1]]
-            lod_levels = [0, 1, 1, 1, 0, 0]
-            dtypes = [
-                'float32', 'float32', 'int32', 'int32', 'float32', 'int64'
+    def build_input(self):
+        self.image = fluid.data(
+            name='image', shape=[None, 3, None, None], dtype='float32')
+        if self.mode == 'train':
+            self.gt_box = fluid.data(
+                name='gt_box', shape=[None, 5], dtype='float32', lod_level=1)
+        else:
+            self.gt_box = fluid.data(
+                name='gt_box', shape=[None, 8], dtype='float32', lod_level=1)
+        self.gt_label = fluid.data(
+            name='gt_class', shape=[None, 1], dtype='int32', lod_level=1)
+        self.is_crowd = fluid.data(
+            name='is_crowed', shape=[None, 1], dtype='int32', lod_level=1)
+        self.im_info = fluid.data(
+            name='im_info', shape=[None, 3], dtype='float32')
+        self.im_id = fluid.data(name='im_id', shape=[None, 1], dtype='int64')
+        self.difficult = fluid.data(
+            name='is_difficult', shape=[None, -1], dtype='float32', lod_level=1)
+        if self.mode == 'train':
+            feed_data = [
+                self.image, self.gt_box, self.gt_label, self.is_crowd,
+                self.im_info, self.im_id
             ]
-            self.py_reader = fluid.layers.py_reader(
-                capacity=64,
-                shapes=in_shapes,
-                lod_levels=lod_levels,
-                dtypes=dtypes,
-                use_double_buffer=True)
-            ins = fluid.layers.read_file(self.py_reader)
-            self.image = ins[0]
-            self.gt_box = ins[1]
-            self.gt_label = ins[2]
-            self.is_crowd = ins[3]
-            self.im_info = ins[4]
-            self.im_id = ins[5]
+        elif self.mode == 'infer':
+            feed_data = [self.image, self.im_info]
         else:
-            self.image = fluid.layers.data(
-                name='image', shape=image_shape, dtype='float32')
-            self.gt_box = fluid.layers.data(
-                name='gt_box', shape=[4], dtype='float32', lod_level=1)
-            self.gt_label = fluid.layers.data(
-                name='gt_label', shape=[1], dtype='int32', lod_level=1)
-            self.is_crowd = fluid.layers.data(
-                name='is_crowd', shape=[1], dtype='int32', lod_level=1)
-            self.im_info = fluid.layers.data(
-                name='im_info', shape=[3], dtype='float32')
-            self.im_id = fluid.layers.data(
-                name='im_id', shape=[1], dtype='int64')
-
-            self.difficult = fluid.layers.data(
-                name='difficult', shape=[1], dtype='float32', lod_level=1)
-
-    def feeds(self):
-        if self.mode == 'infer':
-            return [self.image, self.im_info]
-        if self.mode == 'val':
-            return [
+            feed_data = [
                 self.image, self.gt_box, self.gt_label, self.is_crowd,
                 self.im_info, self.im_id, self.difficult
             ]
-        return [
-            self.image, self.gt_box, self.gt_label, self.is_crowd, self.im_info,
-            self.im_id
-        ]
+        if self.mode == 'train':
+            self.data_loader = fluid.io.DataLoader.from_generator(
+                feed_list=feed_data, capacity=64, iterable=False)
+        else:
+            self.data_loader = fluid.io.DataLoader.from_generator(
+                feed_list=feed_data, capacity=64, iterable=True)
 
     def eval_bbox(self):
         self.im_scale = fluid.layers.slice(
@@ -151,23 +137,37 @@ class RRPN(object):
             dimension = fluid.layers.fill_constant(
                 shape=[1, 1], value=2, dtype='int32')
             cond = fluid.layers.less_than(dimension, res_dimension)
-            res = fluid.layers.create_global_var(
-                shape=[1, 10], value=0.0, dtype='float32', persistable=False)
-            with fluid.layers.control_flow.Switch() as switch:
-                with switch.case(cond):
-                    coordinate = fluid.layers.fill_constant(
-                        shape=[9], value=0.0, dtype='float32')
-                    pred_class = fluid.layers.fill_constant(
-                        shape=[1], value=i + 1, dtype='float32')
-                    add_class = fluid.layers.concat(
-                        [pred_class, coordinate], axis=0)
-                    normal_result = fluid.layers.elementwise_add(pred_result,
-                                                                 add_class)
-                    fluid.layers.assign(normal_result, res)
-                with switch.default():
-                    normal_result = fluid.layers.fill_constant(
-                        shape=[1, 10], value=-1.0, dtype='float32')
-                    fluid.layers.assign(normal_result, res)
+
+            def case1():
+                res = fluid.layers.create_global_var(
+                    shape=[1, 10],
+                    value=0.0,
+                    dtype='float32',
+                    persistable=False)
+                coordinate = fluid.layers.fill_constant(
+                    shape=[9], value=0.0, dtype='float32')
+                pred_class = fluid.layers.fill_constant(
+                    shape=[1], value=i + 1, dtype='float32')
+                add_class = fluid.layers.concat(
+                    [pred_class, coordinate], axis=0)
+                normal_result = fluid.layers.elementwise_add(pred_result,
+                                                             add_class)
+                fluid.layers.assign(normal_result, res)
+                return res
+
+            def case2():
+                res = fluid.layers.create_global_var(
+                    shape=[1, 10],
+                    value=0.0,
+                    dtype='float32',
+                    persistable=False)
+                normal_result = fluid.layers.fill_constant(
+                    shape=[1, 10], value=-1.0, dtype='float32')
+                fluid.layers.assign(normal_result, res)
+                return res
+
+            res = fluid.layers.case(
+                pred_fn_pairs=[(cond, case1)], default=case2)
             results.append(res)
         if len(results) == 1:
             self.pred_result = results[0]

PaddleCV/rrpn/train.py

@@ -56,7 +56,7 @@ def get_device_num():
 
 def train():
     learning_rate = cfg.learning_rate
-    image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size]
+    #image_shape = [-1, 3, cfg.TRAIN.max_size, cfg.TRAIN.max_size]
 
     devices_num = get_device_num()
     total_batch_size = devices_num * cfg.TRAIN.im_per_batch
@@ -71,7 +71,7 @@ def train():
                 add_roi_box_head_func=resnet.ResNetC5(),
                 use_pyreader=cfg.use_pyreader,
                 use_random=use_random)
-            model.build_model(image_shape)
+            model.build_model()
             losses, keys, rpn_rois = model.loss()
             loss = losses[0]
             fetch_list = losses
@@ -132,16 +132,16 @@ def train():
             if num_trainers > 1:
                 train_reader = fluid.contrib.reader.distributed_batch_reader(
                     train_reader)
-        py_reader = model.py_reader
-        py_reader.decorate_paddle_reader(train_reader)
+        data_loader = model.data_loader
+        data_loader.set_sample_list_generator(train_reader, places=place)
     else:
         if num_trainers > 1: shuffle = False
         train_reader = reader.train(
             batch_size=total_batch_size, shuffle=shuffle)
         feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
 
-    def train_loop_pyreader():
-        py_reader.start()
+    def train_loop():
+        data_loader.start()
         train_stats = TrainingStats(cfg.log_window, keys)
         try:
             start_time = time.time()
@@ -173,48 +173,9 @@ def train():
             total_time = end_time - start_time
             last_loss = np.array(outs[0]).mean()
         except (StopIteration, fluid.core.EOFException):
-            py_reader.reset()
-
-    def train_loop():
-        start_time = time.time()
-        prev_start_time = start_time
-        start = start_time
-        train_stats = TrainingStats(cfg.log_window, keys)
-        for iter_id, data in enumerate(train_reader()):
-            prev_start_time = start_time
-            start_time = time.time()
-            if data[0][1].shape[0] == 0:
-                continue
-
-            outs = exe.run(compiled_train_prog,
-                           fetch_list=[v.name for v in fetch_list],
-                           feed=feeder.feed(data))
-            stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
-            train_stats.update(stats)
-            logs = train_stats.log()
-            if iter_id % 10 == 0:
-                strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
-                    now_time(), iter_id,
-                    np.mean(outs[-1]), logs, start_time - prev_start_time)
-                print(strs)
-            sys.stdout.flush()
-            if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0 and iter_id != 0:
-                save_name = "{}".format(iter_id + 1)
-                checkpoint.save(exe, train_prog,
-                                os.path.join(cfg.model_save_dir, save_name))
-            if (iter_id + 1) == cfg.max_iter:
-                checkpoint.save(exe, train_prog,
-                                os.path.join(cfg.model_save_dir, "model_final"))
-                break
-
-        end_time = time.time()
-        total_time = end_time - start_time
-        last_loss = np.array(outs[0]).mean()
+            data_loader.reset()
 
-    if cfg.use_pyreader:
-        train_loop_pyreader()
-    else:
-        train_loop()
+    train_loop()
 
 
 if __name__ == '__main__':

PaddleCV/rrpn/utility.py

@@ -133,7 +133,6 @@ def parse_args():
     add_arg('dataset',          str,   'icdar2015',  "icdar2015, icdar2017.")
     add_arg('class_num',        int,   2,          "Class number.")
     add_arg('data_dir',         str,   'dataset/icdar2015',        "The data root path.")
-    add_arg('use_pyreader',     bool,   False,           "Use pyreader.")
     add_arg('use_profile',         bool,   False,       "Whether use profiler.")
     add_arg('padding_minibatch',bool,   False,
         "If False, only resize image and not pad, image shape is different between"