map/filter dataflow

examples/ResNet/svhn_resnet.py

@@ -171,7 +171,7 @@ def get_config():
             InferenceRunner(dataset_test,
                 [ScalarStats('cost'), ClassificationError() ]),
             ScheduledHyperParamSetter('learning_rate',
-                                      [(1, 0.1), (20, 0.01), (33, 0.001), (60, 0.0001)])
+                                      [(1, 0.1), (20, 0.01), (28, 0.001), (50, 0.0001)])
         ]),
         session_config=sess_config,
         model=Model(n=18),

tensorpack/dataflow/common.py

@@ -144,25 +144,29 @@ class FakeData(DataFlow):
             yield [self.rng.random_sample(k) for k in self.shapes]
 
 class MapData(ProxyDataFlow):
-    """ Map a function on the datapoint"""
+    """ Apply map/filter a function on the datapoint"""
     def __init__(self, ds, func):
         """
         :param ds: a :mod:`DataFlow` instance.
-        :param func: a function that takes a original datapoint, returns a new datapoint
+        :param func: a function that takes a original datapoint, returns a new
+            datapoint. return None to skip this data point.
         """
         super(MapData, self).__init__(ds)
         self.func = func
 
     def get_data(self):
         for dp in self.ds.get_data():
-            yield self.func(dp)
+            ret = self.func(dp)
+            if ret is not None:
+                yield ret
 
 class MapDataComponent(ProxyDataFlow):
-    """ Apply a function to the given index in the datapoint"""
+    """ Apply map/filter on the given index in the datapoint"""
     def __init__(self, ds, func, index=0):
         """
         :param ds: a :mod:`DataFlow` instance.
-        :param func: a function that takes a datapoint dp[index], returns a new value of dp[index]
+        :param func: a function that takes a datapoint dp[index], returns a
+        new value of dp[index]. return None to skip this datapoint.
         """
         super(MapDataComponent, self).__init__(ds)
         self.func = func
@@ -170,8 +174,10 @@ class MapDataComponent(ProxyDataFlow):
 
     def get_data(self):
         for dp in self.ds.get_data():
+            repl = self.func(dp[self.index])
+            if repl is not None:
                 dp = copy.deepcopy(dp)  # avoid modifying the original dp
-            dp[self.index] = self.func(dp[self.index])
+                dp[self.index] = repl
                 yield dp
 
 class RandomChooseData(DataFlow):