allow a trainer without get_predict_func

tensorpack/tfutils/summary.py

@@ -101,6 +101,7 @@ def add_moving_summary(v, *args):
         assert x.get_shape().ndims == 0
         tf.add_to_collection(MOVING_SUMMARY_VARS_KEY, x)
 
+@memoized
 def summary_moving_average(tensors=None):
     """
     Create a MovingAverage op and summary for tensors

tensorpack/train/base.py

@@ -59,10 +59,9 @@ class Trainer(object):
         """ run an iteration"""
         pass
 
-    @abstractmethod
     def get_predict_func(self, input_names, output_names):
         """ return a online predictor"""
-        pass
+        raise NotImplementedError()
 
     def get_predict_funcs(self, input_names, output_names, n):
         """ return n predictor functions.