avoid consuming data for summary, in SimpleTrainer

tensorpack/models/model_desc.py

@@ -161,6 +161,8 @@ class ModelFromMetaGraph(ModelDesc):
     Only useful for inference.
     """
 
+    # TODO can this be really used for inference?
+
     def __init__(self, filename):
         """
         Args:

tensorpack/train/feedfree.py

@@ -25,8 +25,8 @@ class FeedfreeTrainerBase(Trainer):
     """
 
     def _trigger_epoch(self):
-        # need to run summary_op every epoch
-        # note that summary_op will take a data from the queue
+        # run summary_op every epoch
+        # TODO summary_op will take a data! This is not good for TensorInput.
         if self.summary_op is not None:
             summary_str = self.summary_op.eval()
             self.add_summary(summary_str)

tensorpack/train/input_data.py

@@ -45,8 +45,12 @@ class FeedInput(InputData):
     def next_feed(self):
         data = next(self.data_producer)
         feed = dict(zip(self.input_vars, data))
+        self._last_feed = feed
         return feed
 
+    def last_feed(self):
+        return self._last_feed
+
 
 class FeedfreeInput(InputData):
     """ Abstract base for input without feed,

tensorpack/train/trainer.py

@@ -66,7 +66,7 @@ class SimpleTrainer(Trainer):
         self._predictor_factory = PredictorFactory(self.sess, self.model, [0])
         if config.dataflow is None:
             self._input_method = config.data
-            assert isinstance(self._input_method, FeedInput)
+            assert isinstance(self._input_method, FeedInput), type(self._input_method)
         else:
             self._input_method = FeedInput(config.dataflow)
 
@@ -93,7 +93,7 @@ class SimpleTrainer(Trainer):
 
     def _trigger_epoch(self):
         if self.summary_op is not None:
-            feed = self._input_method.next_feed()
+            feed = self._input_method.last_feed()
             summary_str = self.summary_op.eval(feed_dict=feed)
             self.add_summary(summary_str)