Fix CTC gradient disabling optimization

theano/gpuarray/ctc.py

@@ -28,7 +28,7 @@ class GpuConnectionistTemporalClassification(gof.COp):
     compute_grad
         If set to True, enables the computation of gradients of the CTC loss function.
     """
-    __props__ = ('compute_grad',)
+    __props__ = ('compute_grad', 'default_output',)
 
     _cop_num_inputs = 3
     _cop_num_outputs = 2
@@ -39,15 +39,21 @@ class GpuConnectionistTemporalClassification(gof.COp):
     params_type = gpu_context_type
 
     def __init__(self, compute_grad=True):
-        self.compute_grad = compute_grad
-
-        gof.COp.__init__(self, self.func_file, self.func_name)
-
-        if config.ctc.root == "":
+        if not ctc_enabled:
+            raise RuntimeError('Baidu CTC is not enabled and '
+                               'GpuConnectionistTemporalClassification Op '
+                               'can not be constructed.')
+        elif config.ctc.root == "":
             raise ValueError('ctc.root variable is not set, please set it '
                              'to the root directory of the CTC library in '
                              'your system.')
 
+        self.compute_grad = compute_grad
+        # Return only the cost. Gradient will be returned by grad()
+        self.default_output = 0
+
+        gof.COp.__init__(self, self.func_file, self.func_name)
+
     def c_lib_dirs(self):
         dirs = []
         if ctc_enabled:
@@ -86,13 +92,7 @@ class GpuConnectionistTemporalClassification(gof.COp):
         return node.inputs[0].type.context
 
     def make_node(self, activations, labels, input_lengths):
-        if not ctc_enabled:
-            raise RuntimeError('Baidu CTC is not enabled and '
-                               'GpuConnectionistTemporalClassification Op '
-                               'can not be constructed.')
-
         context_name = infer_context_name(activations)
-
         t_activations = as_gpuarray_variable(activations,
                                              context_name=context_name)
         # Ensure activations array is C-contiguous
@@ -111,9 +111,6 @@ class GpuConnectionistTemporalClassification(gof.COp):
         if t_input_lengths.type.dtype != 'int32':
             raise TypeError('Label lengths must use the int32 type!')
 
-        # Return only the cost. Gradient will be returned by grad()
-        self.default_output = 0
-
         costs = GpuArrayType(dtype='float32',
                              broadcastable=(False,),
                              context_name=context_name)()
@@ -133,6 +130,7 @@ class GpuConnectionistTemporalClassification(gof.COp):
             raise RuntimeError('Baidu CTC is not enabled and '
                                'GpuConnectionistTemporalClassification Op '
                                'can not be constructed.')
+        assert len(outputs) == 2
         # Gradients computed by Op
         gradients = outputs[1]
         # Gradients of original function, to compose chain rule
@@ -171,8 +169,7 @@ def gpu_ctc(activations, labels, input_lengths):
     1-D array
         Cost of each example in the minibatch.
     """
-    return GpuConnectionistTemporalClassification()(activations, labels,
-                                                    input_lengths)
+    return GpuConnectionistTemporalClassification()(activations, labels, input_lengths)
 
 
 # Disable gradient computation if not needed
@@ -183,5 +180,5 @@ def local_gpu_ctc_no_grad(node):
     if isinstance(node.op, GpuConnectionistTemporalClassification):
         if len(node.outputs) > 1:
             if len(node.outputs[1].clients) == 0:   # gradient is not used
-                node.op = GpuConnectionistTemporalClassification(compute_grad=False)
-                node.outputs = node.outputs[:1]   # costs only
+                node.op.compute_grad = False
+                return [GpuConnectionistTemporalClassification(compute_grad=False)(*node.inputs)]