fixed reader bug and trainer bug

dygraph/tsn/compose.py

@@ -25,11 +25,11 @@
 class TSN_UCF101_Dataset(Dataset):
     def __init__(self, cfg, mode):
         self.mode = mode
-        self.format = cfg.MODEL.format  #'videos' or 'frames'
+        self.format = cfg.MODEL.format  # 'videos' or 'frames'
         self.seg_num = cfg.MODEL.seg_num
         self.seglen = cfg.MODEL.seglen
-        self.short_size = cfg.TRAIN.short_size
-        self.target_size = cfg.TRAIN.target_size
+        self.short_size = cfg[mode.upper()]['short_size']
+        self.target_size = cfg[mode.upper()]['target_size']
         self.img_mean = np.array(cfg.MODEL.image_mean).reshape(
             [3, 1, 1]).astype(np.float32)
         self.img_std = np.array(cfg.MODEL.image_std).reshape(
@@ -52,7 +52,7 @@ def _construct_loader(self):
                 if self.format == "videos":
                     path, label = path_label.split()
                     self._path_to_videos.append(path + '.avi')
-                    self._num_frames.append(0)  # unused 
+                    self._num_frames.append(0)  # unused
                     self._labels.append(int(label))
                 elif self.format == "frames":
                     path, num_frames, label = path_label.split()
@@ -92,11 +92,11 @@ def __getitem__(self, idx):
                         format(path, ir))
                     return None, None
             label = self._labels[idx]
-            return frames, np.array([label])  #, np.array([idx])
+            return frames, np.array([label])  # , np.array([idx])
 
     def pipline(self, filepath, num_frames, format, seg_num, seglen, short_size,
                 target_size, img_mean, img_std, mode):
-        #Loader
+        # Loader
         if format == 'videos':
             Loader_ops = [
                 VideoDecoder(filepath), VideoSampler(seg_num, seglen, mode)
@@ -106,7 +106,7 @@ def pipline(self, filepath, num_frames, format, seg_num, seglen, short_size,
                 FrameLoader(filepath, num_frames, seg_num, seglen, mode)
             ]
 
-        #Augmentation
+        # Augmentation
         if mode == 'train':
             Aug_ops = [
                 Scale(short_size), RandomCrop(target_size), RandomFlip(),

dygraph/tsn/eval.py

@@ -19,12 +19,13 @@
 import ast
 import logging
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph.base import to_variable
-
+from paddle.io import DataLoader, DistributedBatchSampler
+from compose import TSN_UCF101_Dataset
 from model import TSN_ResNet
 from utils.config_utils import *
-from reader.ucf101_reader import UCF101Reader
 
 logging.root.handlers = []
 FORMAT = '[%(levelname)s: %(filename)s: %(lineno)4d]: %(message)s'
@@ -68,21 +69,28 @@ def test(args):
         model_dict, _ = fluid.load_dygraph(args.weights)
         video_model.set_dict(model_dict)
 
-        test_reader = UCF101Reader(name="TSN", mode="test", cfg=test_config)
-        test_reader = test_reader.create_reader()
+        test_dataset = TSN_UCF101_Dataset(test_config, 'test')
+        test_sampler = DistributedBatchSampler(
+            test_dataset,
+            batch_size=test_config.VALID.batch_size,
+            shuffle=False,
+            drop_last=False)
+        test_loader = DataLoader(
+            test_dataset,
+            batch_sampler=test_sampler,
+            places=place,
+            num_workers=4,
+            return_list=True)
 
         video_model.eval()
         total_loss = 0.0
         total_acc1 = 0.0
         total_acc5 = 0.0
         total_sample = 0
 
-        for batch_id, data in enumerate(test_reader()):
-            x_data = np.array([item[0] for item in data])
-            y_data = np.array([item[1] for item in data]).reshape([-1, 1])
-
-            imgs = to_variable(x_data)
-            labels = to_variable(y_data)
+        for batch_id, data in enumerate(test_loader):
+            imgs = paddle.to_tensor(data[0])
+            labels = paddle.to_tensor(data[1])
             labels.stop_gradient = True
             outputs = video_model(imgs)
             loss = fluid.layers.cross_entropy(