Merge pull request #2 from abergeron/logsoftmax

theano/tensor/nnet/nnet.py

@@ -736,7 +736,7 @@ class LogSoftmax(gof.Op):
 logsoftmax_op = LogSoftmax()
 
 
-@opt.register_specialize('stabilize')
+@opt.register_specialize('stabilize', 'fast_compile')
 @gof.local_optimizer([tensor.Elemwise])
 def local_logsoftmax(node):
     """
@@ -744,20 +744,17 @@ def local_logsoftmax(node):
 
     Note: only forward pass is affected
     """
-    try:
     if (isinstance(node.op, tensor.Elemwise) and
             isinstance(node.op.scalar_op, scalar.basic.Log) and
             len(node.inputs) == 1 and
-            node.inputs[0].owner and
+            node.inputs[0].owner is not None and
             isinstance(node.inputs[0].owner.op, Softmax)):
         inVars = node.inputs[0].owner.inputs[0]
         new_op = LogSoftmax()
         return [new_op(inVars)]
-    except AttributeError:
-        pass
 
 
-@opt.register_specialize('stabilize')
+@opt.register_specialize('stabilize', 'fast_compile')
 @gof.local_optimizer([SoftmaxGrad])
 def local_logsoftmax_grad(node):
     """
@@ -765,16 +762,19 @@ def local_logsoftmax_grad(node):
 
     Note: only grad is affected
     """
-    try:
     if (isinstance(node.op, SoftmaxGrad) and
         len(node.inputs) == 2 and
+        node.inputs[0].owner is not None and
         isinstance(node.inputs[0].owner.op, tensor.Elemwise) and
+        len(node.inputs[0].owner.inputs) >= 2 and
+        node.inputs[0].owner.inputs[1].owner is not None and
         node.inputs[0].owner.inputs[1].owner.op == softmax_op and
         node.inputs[1] == node.inputs[0].owner.inputs[1] and
         not (
             # skip if it will be optimized by
             # local_advanced_indexing_crossentropy_onehot_grad
             node.inputs[0].owner.op == tensor.true_div and
+            node.inputs[0].owner.inputs[0].owner is not None and
             isinstance(node.inputs[0].owner.inputs[0].owner.op,
                        subtensor.AdvancedIncSubtensor))):
         # get parameters from unoptimized op
@@ -785,8 +785,6 @@ def local_logsoftmax_grad(node):
             grads = tensor.alloc(grads, grads.shape[0], sm.shape[1])
 
         return [grads - tensor.sum(grads, axis=1, keepdims=True) * sm]
-    except AttributeError:
-        pass
 
 
 def softmax_graph(c):