Add optional opt that convert all *AllocEmpty to *Alloc.

theano/sandbox/cuda/opt.py

@@ -2540,6 +2540,19 @@ def local_gpu_allocempty(node):
     return False
 
 
+# Don't register by default.
+@gof.local_optimizer([GpuAllocEmpty])
+def local_gpu_alloc_empty_to_zeros(node):
+    if isinstance(node.op, GpuAllocEmpty):
+        return [gpu_alloc(theano.tensor.constant(0, dtype='float32'),
+                          *node.inputs)]
+optdb.register('local_gpu_alloc_empty_to_zeros',
+               local_gpu_alloc_empty_to_zeros,
+               # After move to gpu and merge2, before inplace.
+               49.3,
+               'alloc_empty_to_zeros',)
+
+
 def typeInfer(node):
     return typeConstructor
 

theano/sandbox/gpuarray/opt.py

@@ -300,6 +300,19 @@ def local_gpualloc_memset_0(node):
             return [new_op(*node.inputs)]
 
 
+# Don't register by default.
+@gof.local_optimizer([GpuAllocEmpty])
+def local_gpua_alloc_empty_to_zeros(node):
+    if isinstance(node.op, GpuAllocEmpty):
+        return [GpuAlloc()(theano.tensor.constant(0, dtype='float32'),
+                           *node.inputs)]
+optdb.register('local_gpua_alloc_empty_to_zeros',
+               local_gpua_alloc_empty_to_zeros,
+               # After move to gpu and merge2, before inplace.
+               49.3,
+               'alloc_empty_to_zeros',)
+
+
 @register_opt()
 @local_optimizer([GpuContiguous])
 def local_gpu_contiguous_gpu_contiguous(node):

theano/tensor/opt.py

@@ -1766,6 +1766,26 @@ def local_useless_alloc(node):
             return [node.inputs[0]]
 
 
+# Don't register by default.
+@gof.local_optimizer([T.AllocEmpty])
+def local_alloc_empty_to_zeros(node):
+    """This convert AllocEmpty to Alloc of 0.
+
+    This help investigate NaN with NanGuardMode.  Not registered by
+    default. To activate it, use the Theano flag
+    optimizer_including=alloc_empty_to_zeros. This also enable
+    the GPU version of this optimizations.
+
+    """
+    if isinstance(node.op, T.AllocEmpty):
+        return [T.zeros(node.inputs, dtype=node.outputs[0].dtype)]
+compile.optdb.register('local_alloc_empty_to_zeros',
+                       local_alloc_empty_to_zeros,
+                       # After move to gpu and merge2, before inplace.
+                       49.3,
+                       'alloc_empty_to_zeros',)
+
+
 @register_specialize
 @register_canonicalize
 @gof.local_optimizer([T.shape])