before_train option in PeriodicTrigger

tensorpack/callbacks/trigger.py

@@ -15,10 +15,11 @@ class PeriodicTrigger(ProxyCallback):
     with :meth:`trigger()` method. By default the :meth:`trigger()` method will be called every epoch.
     This wrapper can make the callback run at a different frequency.
 
-    All other methods (``before/after_run``, ``trigger_step``, etc) of the given callback are unaffected.
+    All other methods (``before/after_run``, ``trigger_step``, etc) of the given callback
+    are unaffected. They will still be called as-is.
     """
 
-    def __init__(self, triggerable, every_k_steps=None, every_k_epochs=None):
+    def __init__(self, triggerable, every_k_steps=None, every_k_epochs=None, before_train=False):
         """
         Args:
             triggerable (Callback): a Callback instance with a trigger method to be called.
@@ -26,15 +27,23 @@ class PeriodicTrigger(ProxyCallback):
                 None to ignore.
             every_k_epochs (int): trigger when ``epoch_num % k == 0``. Set to
                 None to ignore.
+            before_train (bool): trigger in the :meth:`before_train` method.
 
-        every_k_steps and every_k_epochs can be both set, but cannot be both None.
+        every_k_steps and every_k_epochs can be both set, but cannot be both None unless before_train is True.
         """
         assert isinstance(triggerable, Callback), type(triggerable)
         super(PeriodicTrigger, self).__init__(triggerable)
+        if before_train is False:
             assert (every_k_epochs is not None) or (every_k_steps is not None), \
-            "every_k_steps and every_k_epochs cannot be both None!"
+                "Arguments to PeriodicTrigger have disabled the triggerable!"
         self._step_k = every_k_steps
         self._epoch_k = every_k_epochs
+        self._before_train = before_train
+
+    def _before_train(self):
+        self.cb.before_train()
+        if self._before_train:
+            self.cb.trigger()
 
     def _trigger_step(self):
         self.cb.trigger_step()