3d use new api (#4275)

* PointNet++ and PointRCNN use new API

PaddleCV/Paddle3D/PointNet++/eval_cls.py

@@ -55,7 +55,7 @@ def parse_args():
     parser.add_argument(
         '--num_points',
         type=int,
-        default=4096,
+        default=2048,
         help='number of points in a sample, default: 4096')
     parser.add_argument(
         '--num_classes',
@@ -101,7 +101,7 @@ def eval():
             eval_model.build_model()
             eval_feeds = eval_model.get_feeds()
             eval_outputs = eval_model.get_outputs()
-            eval_pyreader = eval_model.get_pyreader()
+            eval_loader = eval_model.get_loader()
     eval_prog = eval_prog.clone(True)
     eval_keys, eval_values = parse_outputs(eval_outputs)
 
@@ -109,21 +109,20 @@ def eval():
     exe = fluid.Executor(place)
     exe.run(startup)
 
-    assert os.path.exists(args.weights), "weights {} not exists.".format(args.weights)
-    def if_exist(var):
-        return os.path.exists(os.path.join(args.weights, var.name))
-    fluid.io.load_vars(exe, args.weights, eval_prog, predicate=if_exist)
+    assert os.path.exists("{}.pdparams".format(args.weights)), \
+            "Given resume weight {}.pdparams not exist.".format(args.weights)
+    fluid.load(eval_prog, args.weights, exe)
 
     eval_compile_prog = fluid.compiler.CompiledProgram(eval_prog)
     
     # get reader
     modelnet_reader = ModelNet40ClsReader(args.data_dir, mode='test')
     eval_reader = modelnet_reader.get_reader(args.batch_size, args.num_points)
-    eval_pyreader.decorate_sample_list_generator(eval_reader, place)
+    eval_loader.set_sample_list_generator(eval_reader, place)
 
     eval_stat = Stat()
     try:
-        eval_pyreader.start()
+        eval_loader.start()
         eval_iter = 0
         eval_periods = []
         while True:
@@ -141,7 +140,7 @@ def eval():
     except fluid.core.EOFException:
         logger.info("[EVAL] Eval finished, {}average time: {:.2f}".format(eval_stat.get_mean_log(), np.mean(eval_periods[1:])))
     finally:
-        eval_pyreader.reset()
+        eval_loader.reset()
 
 
 if __name__ == "__main__":

PaddleCV/Paddle3D/PointNet++/eval_seg.py

@@ -100,7 +100,7 @@ def eval():
             eval_model.build_model()
             eval_feeds = eval_model.get_feeds()
             eval_outputs = eval_model.get_outputs()
-            eval_pyreader = eval_model.get_pyreader()
+            eval_loader = eval_model.get_loader()
     eval_prog = eval_prog.clone(True)
     eval_keys, eval_values = parse_outputs(eval_outputs)
 
@@ -108,21 +108,20 @@ def eval():
     exe = fluid.Executor(place)
     exe.run(startup)
 
-    assert os.path.exists(args.weights), "weights {} not exists.".format(args.weights)
-    def if_exist(var):
-        return os.path.exists(os.path.join(args.weights, var.name))
-    fluid.io.load_vars(exe, args.weights, eval_prog, predicate=if_exist)
+    assert os.path.exists("{}.pdparams".format(args.weights)), \
+            "Given resume weight {}.pdparams not exist.".format(args.weights)
+    fluid.load(eval_prog, args.weights, exe)
 
     eval_compile_prog = fluid.compiler.CompiledProgram(eval_prog)
     
     # get reader
     indoor_reader = Indoor3DReader(args.data_dir)
     eval_reader = indoor_reader.get_reader(args.batch_size, args.num_points, mode='test')
-    eval_pyreader.decorate_sample_list_generator(eval_reader, place)
+    eval_loader.set_sample_list_generator(eval_reader, place)
 
     eval_stat = Stat()
     try:
-        eval_pyreader.start()
+        eval_loader.start()
         eval_iter = 0
         eval_periods = []
         while True:
@@ -140,7 +139,7 @@ def eval():
     except fluid.core.EOFException:
         logger.info("[EVAL] Eval finished, {}average time: {:.2f}".format(eval_stat.get_mean_log(), np.mean(eval_periods[1:])))
     finally:
-        eval_pyreader.reset()
+        eval_loader.reset()
 
 
 if __name__ == "__main__":

PaddleCV/Paddle3D/PointNet++/models/pointnet2_cls.py

@@ -35,16 +35,22 @@ class PointNet2Cls(object):
         self.num_points = num_points
         self.use_xyz = use_xyz
         self.out_feature = None
-        self.pyreader = None
+        self.loader = None
         self.model_config()
 
     def model_config(self):
         self.SA_confs = []
 
     def build_input(self):
-        self.xyz = fluid.layers.data(name='xyz', shape=[self.num_points, 3], dtype='float32', lod_level=0)
-        self.label = fluid.layers.data(name='label', shape=[1], dtype='int64', lod_level=0)
-        self.pyreader = fluid.io.PyReader(
+        self.xyz = fluid.data(name='xyz',
+                              shape=[None, self.num_points, 3],
+                              dtype='float32',
+                              lod_level=0)
+        self.label = fluid.data(name='label',
+                                shape=[None, 1],
+                                dtype='int64',
+                                lod_level=0)
+        self.loader = fluid.io.DataLoader.from_generator(
                 feed_list=[self.xyz, self.label],
                 capacity=64,
                 use_double_buffer=True,
@@ -65,11 +71,11 @@ class PointNet2Cls(object):
                     **SA_conf)
 
         out = fluid.layers.squeeze(feature, axes=[-1])
-        out = fc_bn(out,out_channels=512, bn=True, bn_momentum=bn_momentum, name="fc_1")
+        out = fc_bn(out, out_channels=512, bn=True, bn_momentum=bn_momentum, name="fc_1")
         out = fluid.layers.dropout(out, 0.5, dropout_implementation="upscale_in_train")
-        out = fc_bn(out,out_channels=256, bn=True, bn_momentum=bn_momentum, name="fc_2")
+        out = fc_bn(out, out_channels=256, bn=True, bn_momentum=bn_momentum, name="fc_2")
         out = fluid.layers.dropout(out, 0.5, dropout_implementation="upscale_in_train")
-        out = fc_bn(out,out_channels=self.num_classes, act=None, name="fc_3")
+        out = fc_bn(out, out_channels=self.num_classes, act=None, name="fc_3")
         pred = fluid.layers.softmax(out)
 
         # calc loss
@@ -87,8 +93,8 @@ class PointNet2Cls(object):
     def get_outputs(self):
         return {"loss": self.loss, "accuracy": self.acc1}
 
-    def get_pyreader(self):
-        return self.pyreader
+    def get_loader(self):
+        return self.loader
 
 
 class PointNet2ClsSSG(PointNet2Cls):

PaddleCV/Paddle3D/PointNet++/models/pointnet2_seg.py

@@ -36,7 +36,7 @@ class PointNet2SemSeg(object):
         self.use_xyz = use_xyz
         self.feed_vars = []
         self.out_feature = None
-        self.pyreader = None
+        self.loader = None
         self.model_config()
 
     def model_config(self):
@@ -44,10 +44,19 @@ class PointNet2SemSeg(object):
         self.FP_confs = []
 
     def build_input(self):
-        self.xyz = fluid.layers.data(name='xyz', shape=[self.num_points, 3], dtype='float32', lod_level=0)
-        self.feature = fluid.layers.data(name='feature', shape=[self.num_points, 6], dtype='float32', lod_level=0)
-        self.label = fluid.layers.data(name='label', shape=[self.num_points, 1], dtype='int64', lod_level=0)
-        self.pyreader = fluid.io.PyReader(
+        self.xyz = fluid.data(name='xyz',
+                              shape=[None, self.num_points, 3],
+                              dtype='float32',
+                              lod_level=0)
+        self.feature = fluid.data(name='feature',
+                                  shape=[None, self.num_points, 6],
+                                  dtype='float32',
+                                  lod_level=0)
+        self.label = fluid.data(name='label',
+                                shape=[None, self.num_points, 1],
+                                dtype='int64',
+                                lod_level=0)
+        self.loader = fluid.io.DataLoader.from_generator(
                 feed_list=[self.xyz, self.feature, self.label],
                 capacity=64,
                 use_double_buffer=True,
@@ -103,8 +112,8 @@ class PointNet2SemSeg(object):
     def get_outputs(self):
         return {"loss": self.loss, "accuracy": self.acc1}
 
-    def get_pyreader(self):
-        return self.pyreader
+    def get_loader(self):
+        return self.loader
 
 
 class PointNet2SemSegSSG(PointNet2SemSeg):

PaddleCV/Paddle3D/PointNet++/train_cls.py

@@ -54,7 +54,7 @@ def parse_args():
     parser.add_argument(
         '--num_points',
         type=int,
-        default=4096,
+        default=2048,
         help='number of points in a sample, default: 4096')
     parser.add_argument(
         '--num_classes',
@@ -148,7 +148,7 @@ def train():
                           PointNet2ClsSSG(args.num_classes, args.num_points)
             train_model.build_model(bn_momentum=args.bn_momentum)
             train_feeds = train_model.get_feeds()
-            train_pyreader = train_model.get_pyreader()
+            train_loader = train_model.get_loader()
             train_outputs = train_model.get_outputs()
             train_loss = train_outputs['loss']
             lr = fluid.layers.exponential_decay(
@@ -157,9 +157,13 @@ def train():
                     decay_rate=args.lr_decay,
                     staircase=True)
             lr = fluid.layers.clip(lr, 1e-5, args.lr)
+            params = []
+            for var in train_prog.list_vars():
+                if fluid.io.is_parameter(var):
+                    params.append(var.name)
             optimizer = fluid.optimizer.Adam(learning_rate=lr,
                     regularization=fluid.regularizer.L2Decay(args.weight_decay))
-            optimizer.minimize(train_loss)
+            optimizer.minimize(train_loss, parameter_list=params)
     train_keys, train_values = parse_outputs(train_outputs)
 
     test_prog = fluid.Program()
@@ -171,7 +175,7 @@ def train():
             test_model.build_model()
             test_feeds = test_model.get_feeds()
             test_outputs = test_model.get_outputs()
-            test_pyreader = test_model.get_pyreader()
+            test_loader = test_model.get_loader()
     test_prog = test_prog.clone(True)
     test_keys, test_values = parse_outputs(test_outputs)
 
@@ -180,12 +184,13 @@ def train():
     exe.run(startup)
 
     if args.resume:
-        assert os.path.exists(args.resume), \
-                "Given resume weight dir {} not exist.".format(args.resume)
-        def if_exist(var):
-            return os.path.exists(os.path.join(args.resume, var.name))
-        fluid.io.load_vars(
-            exe, args.resume, predicate=if_exist, main_program=train_prog)
+        assert os.path.exists("{}.pdparams".format(args.resume)), \
+                "Given resume weight {}.pdparams not exist.".format(args.resume)
+        assert os.path.exists("{}.pdopt".format(args.resume)), \
+                "Given resume optimizer state {}.pdopt not exist.".format(args.resume)
+        assert os.path.exists("{}.pdmodel".format(args.resume)), \
+                "Given resume model parameter list {}.pdmodel not exist.".format(args.resume)
+        fluid.load(train_prog, args.resume, exe)
 
     build_strategy = fluid.BuildStrategy()
     build_strategy.memory_optimize = False
@@ -200,7 +205,7 @@ def train():
         if os.path.isdir(path):
             shutil.rmtree(path)
         logger.info("Save model to {}".format(path))
-        fluid.io.save_persistables(exe, path, prog)
+        fluid.save(prog, path)
 
     # get reader
     trans_list = [
@@ -213,10 +218,10 @@ def train():
     ]
     modelnet_reader = ModelNet40ClsReader(args.data_dir, mode='train', transforms=trans_list)
     train_reader = modelnet_reader.get_reader(args.batch_size, args.num_points)
-    train_pyreader.decorate_sample_list_generator(train_reader, place)
+    train_loader.set_sample_list_generator(train_reader, place)
     modelnet_reader = ModelNet40ClsReader(args.data_dir, mode='test', transforms=None)
     test_reader = modelnet_reader.get_reader(args.batch_size, args.num_points)
-    test_pyreader.decorate_sample_list_generator(test_reader, place)
+    test_loader.set_sample_list_generator(test_reader, place)
 
     train_stat = Stat()
     test_stat = Stat()
@@ -226,7 +231,7 @@ def train():
 
     for epoch_id in range(args.epoch):
         try:
-            train_pyreader.start()
+            train_loader.start()
             train_iter = 0
             train_periods = []
             while True:
@@ -251,7 +256,7 @@ def train():
             # evaluation
             if not args.enable_ce:
                 try:
-                    test_pyreader.start()
+                    test_loader.start()
                     test_iter = 0
                     test_periods = []
                     while True:
@@ -269,12 +274,12 @@ def train():
                 except fluid.core.EOFException:
                     logger.info("[TEST] Epoch {} finished, {}average time: {:.2f}".format(epoch_id, test_stat.get_mean_log(), np.mean(test_periods[1:])))
                 finally:
-                    test_pyreader.reset()
+                    test_loader.reset()
                     test_stat.reset()
                     test_periods = []
 
         finally:
-            train_pyreader.reset()
+            train_loader.reset()
             train_stat.reset()
             train_periods = []
 

PaddleCV/Paddle3D/PointNet++/train_seg.py

@@ -147,7 +147,7 @@ def train():
                           PointNet2SemSegSSG(args.num_classes, args.num_points)
             train_model.build_model(bn_momentum=args.bn_momentum)
             train_feeds = train_model.get_feeds()
-            train_pyreader = train_model.get_pyreader()
+            train_loader = train_model.get_loader()
             train_outputs = train_model.get_outputs()
             train_loss = train_outputs['loss']
             lr = fluid.layers.exponential_decay(
@@ -156,9 +156,13 @@ def train():
                     decay_rate=args.lr_decay,
                     staircase=True)
             lr = fluid.layers.clip(lr, 1e-5, args.lr)
+            params = []
+            for var in train_prog.list_vars():
+                if fluid.io.is_parameter(var):
+                    params.append(var.name)
             optimizer = fluid.optimizer.Adam(learning_rate=lr,
                     regularization=fluid.regularizer.L2Decay(args.weight_decay))
-            optimizer.minimize(train_loss)
+            optimizer.minimize(train_loss, parameter_list=params)
     train_keys, train_values = parse_outputs(train_outputs)
 
     test_prog = fluid.Program()
@@ -170,7 +174,7 @@ def train():
             test_model.build_model()
             test_feeds = test_model.get_feeds()
             test_outputs = test_model.get_outputs()
-            test_pyreader = test_model.get_pyreader()
+            test_loader = test_model.get_loader()
     test_prog = test_prog.clone(True)
     test_keys, test_values = parse_outputs(test_outputs)
 
@@ -179,12 +183,13 @@ def train():
     exe.run(startup)
 
     if args.resume:
-        assert os.path.exists(args.resume), \
-                "Given resume weight dir {} not exist.".format(args.resume)
-        def if_exist(var):
-            return os.path.exists(os.path.join(args.resume, var.name))
-        fluid.io.load_vars(
-            exe, args.resume, predicate=if_exist, main_program=train_prog)
+        assert os.path.exists("{}.pdparams".format(args.resume)), \
+                "Given resume weight {}.pdparams not exist.".format(args.resume)
+        assert os.path.exists("{}.pdopt".format(args.resume)), \
+                "Given resume optimizer state {}.pdopt not exist.".format(args.resume)
+        assert os.path.exists("{}.pdmodel".format(args.resume)), \
+                "Given resume model parameter list {}.pdmodel not exist.".format(args.resume)
+        fluid.load(train_prog, args.resume, exe)
 
     build_strategy = fluid.BuildStrategy()
     build_strategy.memory_optimize = False
@@ -199,14 +204,14 @@ def train():
         if os.path.isdir(path):
             shutil.rmtree(path)
         logger.info("Save model to {}".format(path))
-        fluid.io.save_persistables(exe, path, prog)
+        fluid.save(prog, path)
 
     # get reader
     indoor_reader = Indoor3DReader(args.data_dir)
     train_reader = indoor_reader.get_reader(args.batch_size, args.num_points, mode='train')
     test_reader = indoor_reader.get_reader(args.batch_size, args.num_points, mode='test')
-    train_pyreader.decorate_sample_list_generator(train_reader, place)
-    test_pyreader.decorate_sample_list_generator(test_reader, place)
+    train_loader.set_sample_list_generator(train_reader, place)
+    test_loader.set_sample_list_generator(test_reader, place)
 
     train_stat = Stat()
     test_stat = Stat()
@@ -216,7 +221,7 @@ def train():
 
     for epoch_id in range(args.epoch):
         try:
-            train_pyreader.start()
+            train_loader.start()
             train_iter = 0
             train_periods = []
             while True:
@@ -236,12 +241,12 @@ def train():
         except fluid.core.EOFException:
             logger.info("[TRAIN] Epoch {} finished, {}average time: {:.2f}".format(epoch_id, train_stat.get_mean_log(), np.mean(train_periods[1:])))
             ce_time = np.mean(train_periods[1:])
-            save_model(exe, train_prog, os.path.join(args.save_dir, str(epoch_id)))
+            save_model(exe, train_prog, os.path.join(args.save_dir, str(epoch_id), "pointnet2_{}_seg".format(args.model)))
             
             # evaluation
             if not args.enable_ce:
                 try:
-                    test_pyreader.start()
+                    test_loader.start()
                     test_iter = 0
                     test_periods = []
                     while True:
@@ -259,12 +264,12 @@ def train():
                 except fluid.core.EOFException:
                     logger.info("[TEST] Epoch {} finished, {}average time: {:.2f}".format(epoch_id, test_stat.get_mean_log(), np.mean(test_periods[1:])))
                 finally:
-                    test_pyreader.reset()
+                    test_loader.reset()
                     test_stat.reset()
                     test_periods = []
 
         finally:
-            train_pyreader.reset()
+            train_loader.reset()
             train_stat.reset()
             train_periods = []
 

PaddleCV/Paddle3D/PointRCNN/README.md

@@ -188,14 +188,14 @@ RPN训练checkpoints默认保存在`checkpoints/rpn`目录，也可以通过`--s
 python tools/generate_aug_scene.py --class_name 'Car' --split train --aug_times 4
 ```
 
-保存RPN模型对离线增强数据的输出特征和ROI，可以通过参数`--ckpt_dir`来指定RPN训练最终权重保存路径，RPN权重默认保存在`checkpoints/rpn`目录。
+保存RPN模型对离线增强数据的输出特征和ROI，可以通过参数`--weights`来指定RPN训练最终权重保存路径，RPN权重默认保存在`checkpoints/rpn`目录。
 保存输出特征和ROI时须指定`TEST.SPLIT`为`train_aug`，指定`TEST.RPN_POST_NMS_TOP_N`为`300`, `TEST.RPN_NMS_THRESH`为`0.85`。
 通过`--output_dir`指定保存输出特征和ROI的路径，默认保存到`./output`目录。
 
 ```
 python eval.py --cfg=cfgs/default.yml  \
                --eval_mode=rpn \
-               --ckpt_dir=./checkpoints/rpn/199 \
+               --weights=./checkpoints/rpn/199 \
                --save_rpn_feature \
                --output_dir=output \
                --set TEST.SPLIT train_aug TEST.RPN_POST_NMS_TOP_N 300 TEST.RPN_NMS_THRESH 0.85
@@ -260,13 +260,13 @@ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`python -c 'import paddle; print(paddle.
 
 2. 保存RPN模型对评估数据的输出特征和ROI
 
-保存RPN模型对评估数据的输出特征和ROI命令如下，可以通过参数`--ckpt_dir`来指定RPN训练最终权重保存路径，RPN权重默认保存在`checkpoints/rpn`目录。
+保存RPN模型对评估数据的输出特征和ROI命令如下，可以通过参数`--weights`来指定RPN训练最终权重保存路径，RPN权重默认保存在`checkpoints/rpn`目录。
 通过`--output_dir`指定保存输出特征和ROI的路径，默认保存到`./output`目录。
 
 ```
 python eval.py --cfg=cfgs/default.yml \
                --eval_mode=rpn \
-               --ckpt_dir=./checkpoints/rpn/199 \
+               --weights=./checkpoints/rpn/199 \
                --save_rpn_feature \
                --output_dir=output/val
 ```
@@ -280,7 +280,7 @@ python eval.py --cfg=cfgs/default.yml \
 ```
 python eval.py --cfg=cfgs/default.yml \
                --eval_mode=rcnn_offline \
-               --ckpt_dir=./checkpoints/rcnn_offline/29 \
+               --weights=./checkpoints/rcnn_offline/29 \
                --rcnn_eval_roi_dir=output/val/detections/data \
                --rcnn_eval_feature_dir=output/val/features \
                --save_result

PaddleCV/Paddle3D/PointRCNN/eval.py

@@ -59,10 +59,10 @@ def parse_args():
         default=1,
         help='evaluation batch size, default 1')
     parser.add_argument(
-        '--ckpt_dir',
+        '--weights',
         type=str,
         default='checkpoints/199',
-        help='specify a ckpt directory to be evaluated if needed')
+        help='specify weights to be evaluated if needed')
     parser.add_argument(
         '--data_dir',
         type=str,
@@ -146,7 +146,7 @@ def eval():
         with fluid.unique_name.guard():
             eval_model = PointRCNN(cfg, args.batch_size, True, 'TEST')
             eval_model.build()
-            eval_pyreader = eval_model.get_pyreader()
+            eval_loader = eval_model.get_loader()
             eval_feeds = eval_model.get_feeds()
             eval_outputs = eval_model.get_outputs()
     eval_prog = eval_prog.clone(True)
@@ -164,13 +164,10 @@ def eval():
 
     exe.run(startup)
 
-    # load checkpoint
-    assert os.path.isdir(
-        args.ckpt_dir), "ckpt_dir {} not a directory".format(args.ckpt_dir)
-
-    def if_exist(var):
-        return os.path.exists(os.path.join(args.ckpt_dir, var.name))
-    fluid.io.load_vars(exe, args.ckpt_dir, eval_prog, predicate=if_exist)
+    # load weights 
+    assert os.path.exists("{}.pdparams".format(args.weights)), \
+            "Given resume weight {}.pdparams not exist.".format(args.weights)
+    fluid.load(eval_prog, args.weights)
 
     kitti_feature_dir = os.path.join(args.output_dir, 'features')
     kitti_output_dir = os.path.join(args.output_dir, 'detections', 'data')
@@ -207,7 +204,7 @@ def eval():
                                         rcnn_eval_roi_dir=args.rcnn_eval_roi_dir,
                                         rcnn_eval_feature_dir=args.rcnn_eval_feature_dir)
     eval_reader = kitti_rcnn_reader.get_multiprocess_reader(args.batch_size, eval_feeds)
-    eval_pyreader.decorate_sample_list_generator(eval_reader, place)
+    eval_loader.set_sample_list_generator(eval_reader, place)
 
     thresh_list = [0.1, 0.3, 0.5, 0.7, 0.9]
     queue = multiprocessing.Queue(128)
@@ -249,7 +246,7 @@ def eval():
             p_list[-1].start()
 
     try:
-        eval_pyreader.start()
+        eval_loader.start()
         eval_iter = 0
         start_time = time.time()
         
@@ -336,7 +333,7 @@ def eval():
                             "run 'python3 tools/kitti_eval.py' to evaluate KITTI mAP.")
 
     finally:
-        eval_pyreader.reset()
+        eval_loader.reset()
 
 
 if __name__ == "__main__":

PaddleCV/Paddle3D/PointRCNN/models/point_rcnn.py

@@ -39,47 +39,47 @@ class PointRCNN(object):
         self.num_points = self.cfg.RPN.NUM_POINTS
         self.prog = prog
         self.inputs = None
-        self.pyreader = None
+        self.loader = None
 
     def build_inputs(self):
         self.inputs = OrderedDict()
 
         if self.cfg.RPN.ENABLED:
-            self.inputs['sample_id'] = fluid.layers.data(name='sample_id', shape=[1], dtype='int32')
-            self.inputs['pts_input'] = fluid.layers.data(name='pts_input', shape=[self.num_points, 3], dtype='float32')
-            self.inputs['pts_rect'] = fluid.layers.data(name='pts_rect', shape=[self.num_points, 3], dtype='float32')
-            self.inputs['pts_features'] = fluid.layers.data(name='pts_features', shape=[self.num_points, 1], dtype='float32')
-            self.inputs['rpn_cls_label'] = fluid.layers.data(name='rpn_cls_label', shape=[self.num_points], dtype='int32')
-            self.inputs['rpn_reg_label'] = fluid.layers.data(name='rpn_reg_label', shape=[self.num_points, 7], dtype='float32')
-            self.inputs['gt_boxes3d'] = fluid.layers.data(name='gt_boxes3d', shape=[7], lod_level=1, dtype='float32')
+            self.inputs['sample_id'] = fluid.data(name='sample_id', shape=[None, 1], dtype='int32')
+            self.inputs['pts_input'] = fluid.data(name='pts_input', shape=[None, self.num_points, 3], dtype='float32')
+            self.inputs['pts_rect'] = fluid.data(name='pts_rect', shape=[None, self.num_points, 3], dtype='float32')
+            self.inputs['pts_features'] = fluid.data(name='pts_features', shape=[None, self.num_points, 1], dtype='float32')
+            self.inputs['rpn_cls_label'] = fluid.data(name='rpn_cls_label', shape=[None, self.num_points], dtype='int32')
+            self.inputs['rpn_reg_label'] = fluid.data(name='rpn_reg_label', shape=[None, self.num_points, 7], dtype='float32')
+            self.inputs['gt_boxes3d'] = fluid.data(name='gt_boxes3d', shape=[None, 7], lod_level=1, dtype='float32')
 
         if self.cfg.RCNN.ENABLED:
             if self.cfg.RCNN.ROI_SAMPLE_JIT:
-                self.inputs['sample_id'] = fluid.layers.data(name='sample_id', shape=[1], dtype='int32', append_batch_size=False)
-                self.inputs['rpn_xyz'] = fluid.layers.data(name='rpn_xyz', shape=[self.num_points, 3], dtype='float32', append_batch_size=False)
-                self.inputs['rpn_features'] = fluid.layers.data(name='rpn_features', shape=[self.num_points,128], dtype='float32', append_batch_size=False)
-                self.inputs['rpn_intensity'] = fluid.layers.data(name='rpn_intensity', shape=[self.num_points], dtype='float32', append_batch_size=False)
-                self.inputs['seg_mask'] = fluid.layers.data(name='seg_mask', shape=[self.num_points], dtype='float32', append_batch_size=False)
-                self.inputs['roi_boxes3d'] = fluid.layers.data(name='roi_boxes3d', shape=[-1, -1, 7], dtype='float32', append_batch_size=False, lod_level=0)
-                self.inputs['pts_depth'] = fluid.layers.data(name='pts_depth', shape=[self.num_points], dtype='float32', append_batch_size=False)
-                self.inputs['gt_boxes3d'] = fluid.layers.data(name='gt_boxes3d', shape=[-1, -1, 7], dtype='float32', append_batch_size=False, lod_level=0)
+                self.inputs['sample_id'] = fluid.data(name='sample_id', shape=[1], dtype='int32')
+                self.inputs['rpn_xyz'] = fluid.data(name='rpn_xyz', shape=[self.num_points, 3], dtype='float32')
+                self.inputs['rpn_features'] = fluid.data(name='rpn_features', shape=[self.num_points, 128], dtype='float32')
+                self.inputs['rpn_intensity'] = fluid.data(name='rpn_intensity', shape=[self.num_points], dtype='float32')
+                self.inputs['seg_mask'] = fluid.data(name='seg_mask', shape=[self.num_points], dtype='float32')
+                self.inputs['roi_boxes3d'] = fluid.data(name='roi_boxes3d', shape=[None, None, 7], dtype='float32', lod_level=0)
+                self.inputs['pts_depth'] = fluid.data(name='pts_depth', shape=[self.num_points], dtype='float32')
+                self.inputs['gt_boxes3d'] = fluid.data(name='gt_boxes3d', shape=[None, None, 7], dtype='float32', lod_level=0)
             else:
-                self.inputs['sample_id'] = fluid.layers.data(name='sample_id', shape=[-1], dtype='int32', append_batch_size=False)
-                self.inputs['pts_input'] = fluid.layers.data(name='pts_input', shape=[-1,512,133], dtype='float32', append_batch_size=False)
-                self.inputs['pts_feature'] = fluid.layers.data(name='pts_feature', shape=[-1,512,128], dtype='float32', append_batch_size=False)
-                self.inputs['roi_boxes3d'] = fluid.layers.data(name='roi_boxes3d', shape=[-1,7], dtype='float32', append_batch_size=False)
+                self.inputs['sample_id'] = fluid.data(name='sample_id', shape=[None], dtype='int32')
+                self.inputs['pts_input'] = fluid.data(name='pts_input', shape=[None, 512, 133], dtype='float32')
+                self.inputs['pts_feature'] = fluid.data(name='pts_feature', shape=[None, 512, 128], dtype='float32')
+                self.inputs['roi_boxes3d'] = fluid.data(name='roi_boxes3d', shape=[None,7], dtype='float32')
                 if self.is_train:
-                    self.inputs['cls_label'] = fluid.layers.data(name='cls_label', shape=[-1], dtype='float32', append_batch_size=False)
-                    self.inputs['reg_valid_mask'] = fluid.layers.data(name='reg_valid_mask', shape=[-1], dtype='float32', append_batch_size=False)
-                    self.inputs['gt_boxes3d_ct'] = fluid.layers.data(name='gt_boxes3d_ct', shape=[-1,7], dtype='float32', append_batch_size=False)
-                    self.inputs['gt_of_rois'] = fluid.layers.data(name='gt_of_rois', shape=[-1,7], dtype='float32', append_batch_size=False)
+                    self.inputs['cls_label'] = fluid.data(name='cls_label', shape=[None], dtype='float32')
+                    self.inputs['reg_valid_mask'] = fluid.data(name='reg_valid_mask', shape=[None], dtype='float32')
+                    self.inputs['gt_boxes3d_ct'] = fluid.data(name='gt_boxes3d_ct', shape=[None, 7], dtype='float32')
+                    self.inputs['gt_of_rois'] = fluid.data(name='gt_of_rois', shape=[None, 7], dtype='float32')
                 else:
-                    self.inputs['roi_scores'] = fluid.layers.data(name='roi_scores', shape=[-1,], dtype='float32', append_batch_size=False)
-                    self.inputs['gt_iou'] = fluid.layers.data(name='gt_iou', shape=[-1], dtype='float32', append_batch_size=False)
-                    self.inputs['gt_boxes3d'] = fluid.layers.data(name='gt_boxes3d', shape=[-1,-1,7], dtype='float32', append_batch_size=False, lod_level=0)
+                    self.inputs['roi_scores'] = fluid.data(name='roi_scores', shape=[None], dtype='float32')
+                    self.inputs['gt_iou'] = fluid.data(name='gt_iou', shape=[None], dtype='float32')
+                    self.inputs['gt_boxes3d'] = fluid.data(name='gt_boxes3d', shape=[None, None, 7], dtype='float32', lod_level=0)
                 
 
-        self.pyreader = fluid.io.PyReader(
+        self.loader = fluid.io.DataLoader.from_generator(
                 feed_list=list(self.inputs.values()),
                 capacity=64,
                 use_double_buffer=True,
@@ -120,6 +120,6 @@ class PointRCNN(object):
         rcnn_loss, _, _ = self.rcnn.get_loss()
         return rpn_loss + rcnn_loss
 
-    def get_pyreader(self):
-        return self.pyreader
+    def get_loader(self):
+        return self.loader
         

PaddleCV/Paddle3D/PointRCNN/train.py

@@ -22,7 +22,6 @@ import numpy as np
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid.layers import control_flow
-from paddle.fluid.contrib.extend_optimizer import extend_with_decoupled_weight_decay
 import paddle.fluid.layers.learning_rate_scheduler as lr_scheduler
 
 from models.point_rcnn import PointRCNN
@@ -169,7 +168,7 @@ def train():
         with fluid.unique_name.guard():
             train_model = PointRCNN(cfg, args.batch_size, True, 'TRAIN')
             train_model.build()
-            train_pyreader = train_model.get_pyreader()
+            train_loader = train_model.get_loader()
             train_feeds = train_model.get_feeds()
             train_outputs = train_model.get_outputs()
             train_loss = train_outputs['loss']
@@ -179,7 +178,7 @@ def train():
                           decay_factor=1e-5,
                           total_step=steps_per_epoch * args.epoch,
                           warmup_pct=cfg.TRAIN.PCT_START,
-                          train_program=train_prog,
+                          train_prog=train_prog,
                           startup_prog=startup,
                           weight_decay=cfg.TRAIN.WEIGHT_DECAY,
                           clip_norm=cfg.TRAIN.GRAD_NORM_CLIP)
@@ -188,13 +187,13 @@ def train():
     exe.run(startup)
 
     if args.resume:
-        assert os.path.exists(args.resume), \
-                "Given resume weight dir {} not exist.".format(args.resume)
-        def if_exist(var):
-            logger.debug("{}: {}".format(var.name, os.path.exists(os.path.join(args.resume, var.name))))
-            return os.path.exists(os.path.join(args.resume, var.name))
-        fluid.io.load_vars(
-            exe, args.resume, predicate=if_exist, main_program=train_prog)
+        assert os.path.exists("{}.pdparams".format(args.resume)), \
+                "Given resume weight {}.pdparams not exist.".format(args.resume)
+        assert os.path.exists("{}.pdopt".format(args.resume)), \
+                "Given resume optimizer state {}.pdopt not exist.".format(args.resume)
+        assert os.path.exists("{}.pdmodel".format(args.resume)), \
+                "Given resume model parameter list {}.pdmodel not exist.".format(args.resume)
+        fluid.load(train_prog, args.resume, exe)
 
     build_strategy = fluid.BuildStrategy()
     build_strategy.memory_optimize = False
@@ -208,19 +207,19 @@ def train():
         if os.path.isdir(path):
             shutil.rmtree(path)
         logger.info("Save model to {}".format(path))
-        fluid.io.save_persistables(exe, path, prog)
+        fluid.save(prog, path)
 
     # get reader
     train_reader = kitti_rcnn_reader.get_multiprocess_reader(args.batch_size,
                                                              train_feeds,
                                                              proc_num=args.worker_num,
                                                              drop_last=True)
-    train_pyreader.decorate_sample_list_generator(train_reader, place)
+    train_loader.set_sample_list_generator(train_reader, place)
 
     train_stat = Stat()
     for epoch_id in range(args.resume_epoch, args.epoch):
         try:
-            train_pyreader.start()
+            train_loader.start()
             train_iter = 0
             train_periods = []
             while True:
@@ -241,7 +240,7 @@ def train():
             train_stat.reset()
             train_periods = []
         finally:
-            train_pyreader.reset()
+            train_loader.reset()
 
 
 if __name__ == "__main__":

PaddleCV/Paddle3D/PointRCNN/utils/optimizer.py

@@ -79,7 +79,7 @@ def optimize(loss,
              decay_factor,
              total_step,
              warmup_pct,
-             train_program,
+             train_prog,
              startup_prog,
              weight_decay,
              clip_norm,
@@ -105,11 +105,15 @@ def optimize(loss,
     param_list = dict()
 
     if weight_decay > 0:
-        for param in train_program.global_block().all_parameters():
+        for param in train_prog.all_parameters():
             param_list[param.name] = param * 1.0
             param_list[param.name].stop_gradient = True
 
-    _, param_grads = optimizer.minimize(loss)
+    opt_param_list = []
+    for var in train_prog.list_vars():
+        if fluid.io.is_parameter(var):
+            opt_param_list.append(var.name)
+    _, param_grads = optimizer.minimize(loss, parameter_list=opt_param_list)
 
     if weight_decay > 0:
         for param, grad in param_grads: