fix argscope bug

tensorpack/callbacks/param.py

@@ -49,7 +49,7 @@ class HyperParamSetter(Callback):
         """
         ret = self._get_current_value()
         if ret is not None and ret != self.last_value:
-            logger.info("{} at epoch {} is changed to {}".format(
+            logger.info("{} at epoch {} will change to {}".format(
                 self.op_name, self.epoch_num, ret))
         self.last_value = ret
         return ret

tensorpack/models/_common.py

@@ -5,6 +5,7 @@
 import tensorflow as tf
 from functools import wraps
 import six
+import copy
 
 from ..tfutils import *
 from ..tfutils.modelutils import *
@@ -34,7 +35,7 @@ def layer_register(summary_activation=False, log_shape=True):
             inputs = args[0]
 
             # update from current argument scope
-            actual_args = get_arg_scope()[func.__name__]
+            actual_args = copy.copy(get_arg_scope()[func.__name__])
             actual_args.update(kwargs)
 
             with tf.variable_scope(name) as scope:

tensorpack/models/nonlin.py

@@ -62,12 +62,14 @@ def LeakyReLU(x, alpha, name=None):
         return tf.mul(x, 0.5, name=name)
 
 
-def BNReLU(is_training):
+def BNReLU(is_training, **kwargs):
     """
+    :param is_traning: boolean
+    :param kwargs: args for BatchNorm
     :returns: a activation function that performs BN + ReLU (a too common combination)
     """
     def BNReLU(x, name=None):
-        x = BatchNorm('bn', x, is_training)
+        x = BatchNorm('bn', x, is_training, **kwargs)
         x = tf.nn.relu(x, name=name)
         return x
     return BNReLU

tensorpack/train/base.py

@@ -71,7 +71,7 @@ class Trainer(object):
         # some final operations that might modify the graph
         logger.info("Preparing for training...")
         self._init_summary()
-        get_global_step_var()
+        get_global_step_var()   # ensure there is such var, before finalizing the graph
         callbacks = self.config.callbacks
         callbacks.before_train(self)
         self.config.session_init.init(self.sess)