use statmonitor similar to PVANet

tensorpack/callbacks/inference.py

@@ -188,7 +188,7 @@ class ClassificationError(Inferencer):
     testing (because the size of test set might not be a multiple of batch size).
     Therefore the result is different from averaging the error rate of each batch.
     """
-    def __init__(self, wrong_var_name='wrong:0', summary_name='validation_error'):
+    def __init__(self, wrong_var_name='wrong:0', summary_name='val_error'):
         """
         :param wrong_var_name: name of the `wrong` variable
         :param summary_name: the name for logging

tensorpack/callbacks/param.py

@@ -199,18 +199,16 @@ class ScheduledHyperParamSetter(HyperParamSetter):
 
 class StatMonitorParamSetter(HyperParamSetter):
     """
-    Set hyperparameter by a func, if a specific stat wasn't
-    monotonically decreasing/increasing $a$ times out of the last $b$ epochs
+    Set hyperparameter by a func, when a specific stat wasn't
+    decreasing/increasing enough in the last $k$ epochs
     """
-    def __init__(self, param, stat_name, value_func,
-            last_k=5,
-            min_non_decreasing=2,
-            reverse=False
+    def __init__(self, param, stat_name, value_func, threshold,
+            last_k, reverse=False
             ):
         """
         Change param by `new_value = value_func(old_value)`,
-        if `stat_name` wasn't decreasing >=2 times in the lastest 5 times of
-            statistics update.
+        if `stat_name` wasn't decreasing > threshold  times in the lastest
+            last_k times of statistics update.
 
         For example, if error wasn't decreasing, anneal the learning rate:
             StatMonitorParamSetter('learning_rate', 'val-error', lambda x: x * 0.2)
@@ -221,13 +219,10 @@ class StatMonitorParamSetter(HyperParamSetter):
         self.stat_name = stat_name
         self.value_func = value_func
         self.last_k = last_k
-        self.min_non_decreasing = min_non_decreasing
-        self.last_changed_epoch = 0
+        self.threshold = threshold
+        self.reverse = reverse
 
-        if not reverse:
-            self.less_than = lambda x, y: x <= y
-        else:
-            self.less_than = lambda x, y: x >= y
+        self.last_changed_epoch = 0
 
     def _get_value_to_set(self):
         holder = self.trainer.stat_holder
@@ -236,13 +231,16 @@ class StatMonitorParamSetter(HyperParamSetter):
                 self.epoch_num - self.last_changed_epoch < self.last_k:
             return None
         hist = hist[-self.last_k-1:]    # len==last_k+1
-        cnt = 0
-        for k in range(self.last_k):
-            if self.less_than(hist[k], hist[k+1]):
-                cnt += 1
-        if cnt >= self.min_non_decreasing \
-                and self.less_than(hist[0], hist[-1]):
+
+        hist_first = hist[0]
+        if not self.reverse:
+            hist_min = min(hist)
+            if hist_min <= hist_first - self.threshold: # small enough
+                return None
+        else:
+            hist_max = max(hist)
+            if hist_max >= hist_first + self.threshold: # large enough
+                return None
         self.last_changed_epoch = self.epoch_num
         return self.value_func(self.get_current_value())
-        return None