Better handling of tags and added back ShapeOptimizer to fast_compile

theano/gpuarray/opt.py

@@ -292,10 +292,9 @@ class GraphToGPU(NavigatorOptimizer):
         for i in fgraph.inputs:
             # Do not move *int* scalar to the GPU.
             target = getattr(i.tag, 'target', None)
-            if (target != 'cpu' and
-               isinstance(i.type, tensor.TensorType) and
+            if (isinstance(i.type, tensor.TensorType) and
                (i.ndim > 0 or 'int' not in i.dtype)):
-                mapping[i] = as_gpuarray_variable(i, target)
+                mapping[i] = i.transfer(target)
             else:
                 mapping[i] = i
         for i in fgraph.variables:

theano/tensor/opt.py

@@ -1480,7 +1480,7 @@ class UnShapeOptimizer(Optimizer):
 # As local_shape_to_shape_i is not in fast_compile,
 # dont register this opt in fast_compile.
 theano.compile.mode.optdb.register('ShapeOpt', ShapeOptimizer(),
-                                   0.1, 'fast_run')
+                                   0.1, 'fast_run', 'fast_compile')
 # Not enabled by default for now. Some crossentropy opt use the
 # shape_feature.  They are at step 2.01. uncanonicalize is at step
 # 3. After it goes to 48.5 that move to the gpu. So 10 seem resonable.