new default option initial_state;

layers.Recurrence() now accepts a scalar constant as the initial_state

bindings/python/cntk/blocks.py

@@ -34,14 +34,18 @@ _INFERRED = (InferredDimension,)  # as a tuple, makes life easier
 def UntestedBranchError(name):
     raise NotImplementedError("Untested code branch: " + name)
 
+# This record contains the defaults for a number of optional parameters to layers.
+# These can be overwritten temporarily by saying
+#    with default_options(init=..., ...):
+#        # code block within which the changed defaults are active
 _current_default_options = Record(
     init=glorot_uniform(),
-    activation=None,       # Dense() and Convolution() have no activation by default
+    activation=None,                  # Dense() and Convolution() have no activation by default
     pad=False,
     bias=True,
     init_bias=0,
-    enable_self_stabilization=False # Stabilizer() and LSTM()
-    # TODO: add initial_state?
+    enable_self_stabilization=False,  # Stabilizer() and LSTM()
+    initial_state=None                # Recurrence()
 )
 
 _default_sentinel           = Record() # This is a singleton sentinel value we recognize and replace in _initializer_for()
@@ -54,6 +58,7 @@ init_bias_default_or_0                     = _default_sentinel_init_bias
 bias_default_or_True                       = _default_sentinel
 pad_default_or_False                       = _default_sentinel
 enable_self_stabilization_default_or_False = _default_sentinel
+initial_state_default_or_None              = _default_sentinel_init_bias
 
 # check whether a parameter is a default
 # This is meant to be used by implementations of layers that take default values that may default to default-defaults.

bindings/python/cntk/layers.py

@@ -205,9 +205,13 @@ def AveragePooling(filter_shape,  # e.g. (3,3)
     return Pooling(PoolingType_Average, filter_shape, strides=strides, pad=pad)
 
 # Recurrence() -- run a block recurrently over a time sequence
-def Recurrence(over, go_backwards=False, initial_state=None):
+def Recurrence(over, go_backwards=False, initial_state=initial_state_default_or_None):
     # helper to compute previous value
     # can take a single Variable/Function or a tuple
+    initial_state = initial_state if _is_given(initial_state) else _current_default_options.initial_state
+    # if initial state is given and a numeric constant, then turn it into a Constant() object
+    if np.isscalar(initial_state):
+        initial_state = Constant(initial_state, shape=(1)) # TODO: This should be automatically done inside the API.
     if go_backwards:
         UntestedBranchError("Recurrence, go_backwards option") # TODO: test in SLUHandsOn.py bidirectional
     def previous_hook(state):

bindings/python/examples/SLUHandsOn/SLUHandsOn.py

@@ -48,15 +48,13 @@ def create_reader(path):
 ########################
 
 def create_model():
-  #with default_options(enable_self_stabilization=False):
+  # the following injects an option to mimic the BS version identically
+  with default_options(enable_self_stabilization=False, initial_state=0.1):
     return Sequential([
         #Stabilizer(),
         Embedding(emb_dim),
         #BatchNormalization(),
-        Recurrence(LSTM(hidden_dim), go_backwards=False,
-                   #),
-                   initial_state=Constant(0.1, shape=(1))),   # (this last option mimics a default in BS to recreate identical results)
-                   # BUGBUG: initial_state=0.1 should work
+        Recurrence(LSTM(hidden_dim), go_backwards=False),
         #Stabilizer(),
         #BatchNormalization(),
         Dense(label_dim)