Fix the loading of json stats (fix #904 introduced by 4a917d6e)

tensorpack/callbacks/inference.py

@@ -78,6 +78,9 @@ class ScalarStats(Inferencer):
     """
     Statistics of some scalar tensor.
     The value will be averaged over all given datapoints.
+
+    Note that the average of accuracy over all batches is not necessarily the
+    accuracy of the whole dataset. See :class:`ClassificationError` for details.
     """
 
     def __init__(self, names, prefix='validation'):

tensorpack/callbacks/monitor.py

@@ -306,7 +306,13 @@ class JSONWriter(TrainingMonitor):
         except Exception:
             return None
 
+    # initialize the stats here, because before_train from other callbacks may use it
     def _setup_graph(self):
+        self._stats = []
+        self._stat_now = {}
+        self._last_gs = -1
+
+    def _before_train(self):
         stats = JSONWriter.load_existing_json()
         self._fname = os.path.join(logger.get_logger_dir(), JSONWriter.FILENAME)
         if stats is not None:
@@ -315,33 +321,27 @@ class JSONWriter(TrainingMonitor):
             except Exception:
                 epoch = None
 
+            # check against the current training settings
+            # therefore this logic needs to be in before_train stage
             starting_epoch = self.trainer.loop.starting_epoch
             if epoch is None or epoch == starting_epoch:
                 logger.info("Found existing JSON inside {}, will append to it.".format(logger.get_logger_dir()))
                 self._stats = stats
             else:
                 logger.warn(
-                    "History epoch value {} from JSON is not the predecessor of the starting_epoch value {}".format(
+                    "History epoch={} from JSON is not the predecessor of the current starting_epoch={}".format(
                         epoch - 1, starting_epoch))
                 logger.warn("If you want to resume old training, either use `AutoResumeTrainConfig` "
-                            "or correctly set the starting_epoch yourself to avoid inconsistency. "
-                            "Epoch number will not be automatically loaded by JSONWriter.")
+                            "or correctly set the new starting_epoch yourself to avoid inconsistency. ")
 
                 backup_fname = JSONWriter.FILENAME + '.' + datetime.now().strftime('%m%d-%H%M%S')
                 backup_fname = os.path.join(logger.get_logger_dir(), backup_fname)
 
-                logger.warn("Now, we will start training at epoch {} and backup old json to {}".format(
+                logger.warn("Now, we will train with starting_epoch={} and backup old json to {}".format(
                     self.trainer.loop.starting_epoch, backup_fname))
                 shutil.move(self._fname, backup_fname)
-                self._stats = []
-        else:
-            self._stats = []
-        self._stat_now = {}
-
-        self._last_gs = -1
 
-    # in case we have something to log here.
-    def _before_train(self):
+        # in case we have something to log here.
         self._trigger()
 
     def _trigger_step(self):

tensorpack/dataflow/dataset/cifar.py

@@ -85,7 +85,7 @@ def _parse_meta(filename, cifar_classnum):
 
 
 class CifarBase(RNGDataFlow):
-    def __init__(self, train_or_test, shuffle=True, dir=None, cifar_classnum=10):
+    def __init__(self, train_or_test, shuffle=None, dir=None, cifar_classnum=10):
         assert train_or_test in ['train', 'test']
         assert cifar_classnum == 10 or cifar_classnum == 100
         self.cifar_classnum = cifar_classnum
@@ -104,6 +104,9 @@ class CifarBase(RNGDataFlow):
         self.train_or_test = train_or_test
         self.data = read_cifar(self.fs, cifar_classnum)
         self.dir = dir
+
+        if shuffle is None:
+            shuffle = train_or_test == 'train'
         self.shuffle = shuffle
 
     def __len__(self):
@@ -149,18 +152,18 @@ class Cifar10(CifarBase):
     image is 32x32x3 in the range [0,255].
     label is an int.
     """
-    def __init__(self, train_or_test, shuffle=True, dir=None):
+    def __init__(self, train_or_test, shuffle=None, dir=None):
         """
         Args:
             train_or_test (str): either 'train' or 'test'.
-            shuffle (bool): shuffle the dataset.
+            shuffle (bool): shuffle the dataset, default to shuffle in training
         """
         super(Cifar10, self).__init__(train_or_test, shuffle, dir, 10)
 
 
 class Cifar100(CifarBase):
     """ Similar to Cifar10"""
-    def __init__(self, train_or_test, shuffle=True, dir=None):
+    def __init__(self, train_or_test, shuffle=None, dir=None):
         super(Cifar100, self).__init__(train_or_test, shuffle, dir, 100)
 
 

tensorpack/train/config.py

@@ -222,7 +222,7 @@ class AutoResumeTrainConfig(TrainConfig):
             if last_epoch is not None:
                 now_epoch = last_epoch + 1
                 logger.info("Found history statistics from JSON. "
-                            "Overwrite the starting epoch to epoch #{}.".format(now_epoch))
+                            "Setting starting_epoch to {}.".format(now_epoch))
                 kwargs['starting_epoch'] = now_epoch
 
         super(AutoResumeTrainConfig, self).__init__(**kwargs)