Merge pull request #1669 from abergeron/eq_opt

theano/gof/opt.py

@@ -736,6 +736,14 @@ class LocalOptimizer(object):
             _optimizer_idx[0] += 1
         return self._optimizer_idx
 
+    def tracks(self):
+        """
+        Return the list of op classes that this opt applies to.
+
+        Return None to apply to all nodes.
+        """
+        return None
+
     def transform(self, node):
         """Transform a subgraph whose output is `node`.
 
@@ -772,8 +780,6 @@ class LocalOptimizer(object):
 class FromFunctionLocalOptimizer(LocalOptimizer):
     """WRITEME"""
     def __init__(self, fn, tracks=None):
-        if tracks is None:
-            tracks = []
         self.transform = fn
         self._tracks = tracks
 
@@ -791,9 +797,15 @@ class FromFunctionLocalOptimizer(LocalOptimizer):
                 id(self))
 
 
-def local_optimizer(*tracks):
+def local_optimizer(tracks):
     def decorator(f):
         """WRITEME"""
+        if tracks is not None:
+            if len(tracks) is 0:
+                raise ValueError, ("Use None instead of an empty list to apply to all nodes.", f.__module__, f.__name__)
+            for t in tracks:
+                if not (isinstance(t, op.Op) or issubclass(t, op.PureOp)):
+                    raise ValueError, ("Tracks are op classes or instances", f.__module__, f.__name__)
         rval = FromFunctionLocalOptimizer(f, tracks)
         rval.__name__ = f.__name__
         return rval
@@ -870,7 +882,7 @@ class OpSub(LocalOptimizer):
         return self.op1
 
     def tracks(self):
-        return [[self.op1]]
+        return [self.op1]
 
     def transform(self, node):
         if node.op != self.op1:
@@ -901,7 +913,7 @@ class OpRemove(LocalOptimizer):
         return self.op
 
     def tracks(self):
-        return [[self.op]]
+        return [self.op]
 
     def transform(self, node):
         if node.op != self.op:
@@ -1008,17 +1020,7 @@ class PatternSub(LocalOptimizer):
         return self.op
 
     def tracks(self):
-        def helper(pattern, sofar):
-            if isinstance(pattern, (list, tuple)):
-                sofar = sofar + (pattern[0],)
-                return reduce(tuple.__add__,
-                              tuple(helper(p, sofar) for p in pattern[1:]),
-                              ())
-            elif isinstance(pattern, dict):
-                return helper(pattern['pattern'], sofar)
-            else:
-                return (sofar,)
-        return set(helper(self.in_pattern, ()))
+        return [self.op]
 
     def transform(self, node):
         """
@@ -1500,12 +1502,17 @@ class EquilibriumOptimizer(NavigatorOptimizer):
             None,
             ignore_newtrees=True,
             failure_callback=failure_callback)
-        self.local_optimizers = []
+        self.local_optimizers_map = dict()
+        self.local_optimizers_all = []
         self.global_optimizers = []
 
         for opt in optimizers:
             if isinstance(opt, LocalOptimizer):
-                self.local_optimizers.append(opt)
+                if opt.tracks() is None:
+                    self.local_optimizers_all.append(opt)
+                else:
+                    for c in opt.tracks():
+                        self.local_optimizers_map.setdefault(c, []).append(opt)
             else:
                 self.global_optimizers.append(opt)
         self.max_depth = max_depth
@@ -1513,10 +1520,21 @@ class EquilibriumOptimizer(NavigatorOptimizer):
         assert self.max_use_ratio is not None, (
                 'max_use_ratio has to be a number')
 
+    def get_local_optimizers(self):
+        for opt in self.local_optimizers_all:
+            yield opt
+        # if repeat is not a problem we can drop the set
+        s = set()
+        for lopt in self.local_optimizers_map.values():
+            for opt in lopt:
+                if opt not in s:
+                    yield opt
+                    s.add(opt)
+
     def add_requirements(self, fgraph):
         super(EquilibriumOptimizer, self).add_requirements(fgraph)
         fgraph.attach_feature(ChangeTracker())
-        for opt in self.local_optimizers:
+        for opt in self.get_local_optimizers():
             opt.add_requirements(fgraph)
         for opt in self.global_optimizers:
             opt.add_requirements(fgraph)
@@ -1542,7 +1560,7 @@ class EquilibriumOptimizer(NavigatorOptimizer):
         time_opts = {}
         io_toposort_timing = []
         nb_nodes = []
-        for opt in self.global_optimizers + self.local_optimizers:
+        for opt in self.global_optimizers + list(self.get_local_optimizers()):
             global_process_count.setdefault(opt, 0)
             time_opts.setdefault(opt, 0)
 
@@ -1595,7 +1613,9 @@ class EquilibriumOptimizer(NavigatorOptimizer):
                     node = q.pop()
                     current_node = node
 
-                    for lopt in self.local_optimizers:
+                    for lopt in (self.local_optimizers_all +
+                                 self.local_optimizers_map.get(type(node.op), []) +
+                                 self.local_optimizers_map.get(node.op, [])):
                         t_opt = time.time()
                         lopt_change = self.process_node(fgraph, node, lopt)
                         time_opts[lopt] += time.time() - t_opt
@@ -1634,7 +1654,7 @@ class EquilibriumOptimizer(NavigatorOptimizer):
         print >> stream, "%s%s %s id=%i" % (
                 (' ' * level), self.__class__.__name__, name, id(self))
         if depth != 0:
-            for lopt in self.local_optimizers:
+            for lopt in self.get_local_optimizers():
                 lopt.print_summary(stream, level=(level + 2),
                                    depth=(depth - 1))
 
@@ -1654,7 +1674,7 @@ class EquilibriumOptimizer(NavigatorOptimizer):
                 start_nb_nodes, end_nb_nodes, max_nb_nodes)
         print >> stream, blanc, "  time io_toposort %.3fs" % sum(
             io_toposort_timing)
-        s = sum([time_opts[o] for o in opt.local_optimizers])
+        s = sum([time_opts[o] for o in opt.get_local_optimizers()])
         print >> stream, blanc, "  time in local optimizers %.3fs" % s
         s = sum([time_opts[o] for o in opt.global_optimizers])
         print >> stream, blanc, "  time in global optimizers %.3fs" % s
@@ -1679,7 +1699,7 @@ class EquilibriumOptimizer(NavigatorOptimizer):
         not_used = 0
         not_used_time = 0
         process_count = {}
-        for o in opt.global_optimizers + opt.local_optimizers:
+        for o in opt.global_optimizers + list(opt.get_local_optimizers()):
             process_count.setdefault(o, 0)
         for count in loop_process_count:
             for o, v in count.iteritems():
@@ -1707,8 +1727,8 @@ class EquilibriumOptimizer(NavigatorOptimizer):
         #(opt, loop_timing, loop_process_count, max_nb_nodes,
         # global_opt_timing, nb_nodes, time_opts, io_toposort_timing) = prof1
 
-        local_optimizers = set(prof1[0].local_optimizers).union(
-            prof2[0].local_optimizers)
+        local_optimizers = set(prof1[0].get_local_optimizers()).union(
+            prof2[0].get_local_optimizers())
         global_optimizers = set(prof1[0].global_optimizers).union(
             prof2[0].global_optimizers)
         new_opt = EquilibriumOptimizer(

theano/ifelse.py

@@ -384,7 +384,7 @@ def ifelse(condition, then_branch, else_branch, name=None):
         return tuple(rval)
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IfElse])
 def cond_make_inplace(node):
     op = node.op
     if isinstance(op, IfElse) and not op.as_view:
@@ -445,7 +445,7 @@ acceptable_ops = (theano.tensor.basic.Dot,
                   theano.tensor.elemwise.DimShuffle)
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer(acceptable_ops)
 def ifelse_lift_single_if_through_acceptable_ops(main_node):
     """This optimization lifts up certain ifelse instances.
 
@@ -493,7 +493,7 @@ def ifelse_lift_single_if_through_acceptable_ops(main_node):
     return nw_outs
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IfElse])
 def cond_merge_ifs_true(node):
     op = node.op
     if not isinstance(op, IfElse):
@@ -517,7 +517,7 @@ def cond_merge_ifs_true(node):
     return op(*old_ins, **dict(return_list=True))
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IfElse])
 def cond_merge_ifs_false(node):
     op = node.op
     if not isinstance(op, IfElse):
@@ -592,7 +592,7 @@ class CondMerge(gof.Optimizer):
                 fgraph.replace_all_validate(pairs, reason='cond_merge')
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IfElse])
 def cond_remove_identical(node):
     op = node.op
 
@@ -643,7 +643,7 @@ def cond_remove_identical(node):
     return rval
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IfElse])
 def cond_merge_random_op(main_node):
     if isinstance(main_node.op, IfElse):
         return False

theano/sandbox/cuda/GpuConv3D.py

@@ -284,7 +284,7 @@ conv_rows_stack( float* img, float* kern, float* bias, float* out,
 gpu_convd = GpuConv3D()
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([Conv3D])
 def local_gpu_conv3d(node):
     if isinstance(node.op, Conv3D):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu) for i in node.inputs]):

theano/sandbox/cuda/GpuConvGrad3D.py

@@ -341,7 +341,7 @@ convgrad_rows_stack( float* img, float* dCdH, float* dCdW,
 gpu_conv_grad3d = GpuConvGrad3D()
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([ConvGrad3D])
 def local_gpu_conv_gradd(node):
     if isinstance(node.op, ConvGrad3D):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu) for i in node.inputs]):

theano/sandbox/cuda/GpuConvTransp3D.py

@@ -348,7 +348,7 @@ conv_transp_rows_stack( float* H, float* kern, float* bias, float* R,
 gpu_conv_transpd = GpuConvTransp3D()
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([ConvTransp3D])
 def local_gpu_conv_transpd(node):
     if isinstance(node.op, ConvTransp3D):
         if numpy.any([i.owner and isinstance(i.owner.op, HostFromGpu) for i in node.inputs]):

theano/sandbox/cuda/neighbours.py

@@ -405,7 +405,7 @@ def gpu_images2neibs(ten4, neib_shape, neib_step=None, mode='valid'):
     return GpuImages2Neibs(mode)(ten4, neib_shape, neib_step)
 
 
-@local_optimizer()
+@local_optimizer([Images2Neibs])
 def use_gpu_images2neibs(node):
     if (type(node.op) is Images2Neibs and
         node.inputs[0].dtype == 'float32' and

theano/sandbox/cuda/opt.py

@@ -121,7 +121,7 @@ gpu_seqopt.register('InputToGpuOptimizer', InputToGpuOptimizer(),
                     'merge')  # TODO: how to make it mandatory for gpu_seqopt?
 
 
-@local_optimizer([])
+@local_optimizer([gpu_from_host, host_from_gpu])
 def local_cut_gpu_host_gpu(node):
     if tensor.opt.opt.check_chain(node, gpu_from_host, host_from_gpu):
         return [node.inputs[0].owner.inputs[0]]
@@ -170,7 +170,7 @@ def dtype_in_elemwise_supported(op):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.Elemwise])
 def local_gpu_elemwise_0(node):
     """elemwise(..., host_from_gpu, ...)
        -> host_from_gpu(elemwise(gpu_from_host, ..., gpu_from_host)
@@ -229,7 +229,7 @@ def local_gpu_elemwise_0(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host])
 def local_gpu_elemwise_1(node):
     """
     gpu_from_host(Elemwise)) -> GpuElemwise(gpu_from_host(...))
@@ -265,7 +265,7 @@ def local_gpu_elemwise_1(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.DimShuffle, gpu_from_host])
 def local_gpu_dimshuffle_0(node):
     """
     dimshuffle(host_from_gpu()) -> host_from_gpu(gpu_dimshuffle)
@@ -290,7 +290,7 @@ def local_gpu_dimshuffle_0(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.SpecifyShape, gpu_from_host])
 def local_gpu_specifyShape_0(node):
     """
     specify_shape(host_from_gpu()) -> host_from_gpu(specify_shape)
@@ -313,7 +313,7 @@ def local_gpu_specifyShape_0(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host]) # XXX: broken: tensor.basic.dot is not an op
 def local_gpu_dot_to_dot22(node):
     """
     gpu_from_host(dot) -> gpudot(gpu_from_host)
@@ -376,7 +376,7 @@ def local_gpu_dot_to_dot22(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([theano.ifelse.IfElse, gpu_from_host])
 def local_gpu_lazy_ifelse(node):
     """
     gpu_from_host(ifelse) -> gpu_ifelse(gpu_from_host)
@@ -434,7 +434,7 @@ def local_gpu_lazy_ifelse(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.blas._dot22])
 def local_gpu_dot22(node):
     """
     gpu_from_host(dot22) -> gpudot(gpu_from_host)
@@ -456,7 +456,7 @@ def local_gpu_dot22(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.blas._dot22scalar])
 def local_gpu_dot22scalar(node):
     """
     gpu_from_host(dot22scalar) -> gpudot(gpu_from_host)
@@ -482,7 +482,7 @@ def local_gpu_dot22scalar(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.blas_c.CGemv, tensor.blas.Gemv])
 def local_gpu_gemv(node):
     """
     gpu_from_host(gemv) -> gpu_gemv(gpu_from_host)
@@ -523,7 +523,8 @@ def local_gpu_gemv(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.blas_c.CGer, tensor.blas.Ger,
+                  tensor.blas_scipy.ScipyGer])
 def local_gpu_ger(node):
     """
     gpu_from_host(ger) -> gpu_ger(gpu_from_host)
@@ -566,7 +567,7 @@ def local_gpu_ger(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.blas.gemm_no_inplace, gpu_from_host])
 def local_gpu_gemm(node):
     """
     gpu_from_host(gemm) -> gpu_gemm(gpu_from_host)
@@ -601,7 +602,13 @@ def local_gpu_gemm(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.elemwise.CAReduce,
+                  tensor.elemwise.All,
+                  tensor.elemwise.Any,
+                  tensor.elemwise.CAReduceDtype,
+                  tensor.elemwise.Sum,
+                  tensor.elemwise.Prod,
+                  tensor.elemwise.ProdWithoutZeros])
 def local_gpu_careduce(node):
     if isinstance(node.op, tensor.elemwise.CAReduce):
         scalar_op = node.op.scalar_op
@@ -671,7 +678,7 @@ def local_gpu_careduce(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.Reshape])
 def local_gpu_reshape(node):
     if node.op == gpu_from_host:
         host_input = node.inputs[0]
@@ -705,7 +712,7 @@ def local_gpu_reshape(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.Flatten])
 def local_gpu_flatten(node):
     if node.op == gpu_from_host:
         host_input = node.inputs[0]
@@ -724,7 +731,7 @@ def local_gpu_flatten(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.Subtensor])
 def local_gpu_subtensor(node):
     if node.op == gpu_from_host:
         host_input = node.inputs[0]
@@ -745,7 +752,7 @@ def local_gpu_subtensor(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.AdvancedSubtensor1])
 def local_gpu_advanced_subtensor1(node):
     if node.op == gpu_from_host:
         host_input = node.inputs[0]
@@ -764,7 +771,7 @@ def local_gpu_advanced_subtensor1(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.AdvancedIncSubtensor1])
 def local_gpu_advanced_incsubtensor1(node):
     if node.op == gpu_from_host:
         host_input = node.inputs[0]
@@ -838,7 +845,7 @@ def local_gpu_advanced_incsubtensor1(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.IncSubtensor])
 def local_gpu_incsubtensor(node):
     if node.op == gpu_from_host:
         host_output = node.inputs[0]
@@ -885,7 +892,7 @@ def local_gpu_incsubtensor(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.Shape])
 def local_gpu_shape(node):
     if isinstance(node.op, tensor.Shape):
         x, = node.inputs
@@ -896,7 +903,7 @@ def local_gpu_shape(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.Rebroadcast])
 def local_gpu_rebroadcast(node):
     '''rebroadcast(host_from_gpu(x)) -> host_from_gpu(rebroadcast(x))'''
     if isinstance(node.op, tensor.Rebroadcast):
@@ -911,7 +918,7 @@ def gpu_print_wrapper(op, cnda):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.printing.Print])
 def local_gpu_print_op(node):
     if isinstance(node.op, tensor.printing.Print):
         x, = node.inputs
@@ -932,7 +939,7 @@ import theano.tensor.nnet
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.nnet.CrossentropySoftmaxArgmax1HotWithBias])
 def local_gpu_crossentorpy_softmax_argmax_1hot_with_bias(node):
     if isinstance(node.op, tensor.nnet.CrossentropySoftmaxArgmax1HotWithBias):
         x, b, y = node.inputs
@@ -962,7 +969,7 @@ def local_gpu_crossentorpy_softmax_argmax_1hot_with_bias(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.nnet.CrossentropySoftmax1HotWithBiasDx])
 def local_gpu_crossentorpy_softmax_1hot_with_bias_dx(node):
     if isinstance(node.op, tensor.nnet.CrossentropySoftmax1HotWithBiasDx):
         dnll, sm, yidx = node.inputs
@@ -977,7 +984,7 @@ def local_gpu_crossentorpy_softmax_1hot_with_bias_dx(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.nnet.Softmax])
 def local_gpu_softmax(node):
     if isinstance(node.op, tensor.nnet.Softmax):
         x, = node.inputs
@@ -989,7 +996,7 @@ def local_gpu_softmax(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.nnet.SoftmaxWithBias])
 def local_gpu_softmax_with_bias(node):
     if isinstance(node.op, tensor.nnet.SoftmaxWithBias):
         x, b = node.inputs
@@ -1005,7 +1012,7 @@ from theano.tensor.nnet import conv
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, conv.ConvOp])
 def local_gpu_conv(node):
     """
     gpu_from_host(conv) -> gpu_conv(gpu_from_host)
@@ -1105,7 +1112,7 @@ import theano.tensor.signal.downsample as downsample
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([downsample.DownsampleFactorMax])
 def local_gpu_downsample_factor_max(node):
     if isinstance(node.op, downsample.DownsampleFactorMax):
         x, = node.inputs
@@ -1115,7 +1122,7 @@ def local_gpu_downsample_factor_max(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([downsample.DownsampleFactorMaxGrad])
 def local_gpu_downsample_factor_max_grad(node):
     if isinstance(node.op, downsample.DownsampleFactorMaxGrad):
         x, z, gz = node.inputs
@@ -1127,11 +1134,11 @@ def local_gpu_downsample_factor_max_grad(node):
                                               gpu_from_host(gz)))]
 
 
-from theano.sandbox.cuda.basic_ops import gpu_join
+from theano.sandbox.cuda.basic_ops import gpu_join, GpuJoin
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([tensor.Join])
 def local_gpu_join(node):
     """
     Inspired by the opt for convop.
@@ -1188,6 +1195,14 @@ def local_gpu_join(node):
 
             return [replacement_node]
 
+# This is a copy of the same opt in tensor to make the tests happy,
+# but I'm not convinced it is actually needed.
+@register_opt()
+@local_optimizer([GpuJoin])
+def local_gpujoin_1(node):
+    tensors = node.inputs[1:]
+    if len(tensors) == 1:
+        return [tensors[0]]
 
 # Commented out because it can result in
 #   shared =  dimshuffle(gemm_inplace(dimshuffle(shared)))
@@ -1205,7 +1220,7 @@ def local_inplace_gemv(node):
         return [gpu_gemv_inplace(*node.inputs)]
 
 
-@local_optimizer([gpu_gemm_no_inplace])
+@local_optimizer([gpu_ger_no_inplace])
 def local_inplace_ger(node):
     if node.op == gpu_ger_no_inplace:
         return [gpu_ger_inplace(*node.inputs)]
@@ -1336,7 +1351,7 @@ optdb.register('gpu_inplace_elemwise_opt', gpu_inplace_elemwise_optimizer, 75,
 
 
 @register_opt()
-@local_optimizer([tensor.Alloc])
+@local_optimizer([tensor.alloc])
 def local_gpualloc(node):
     replace = False
     if node.op == tensor.alloc:
@@ -1383,7 +1398,7 @@ def local_gpualloc(node):
 
 
 @register_opt()
-@local_optimizer([tensor.Alloc])
+@local_optimizer([GpuAlloc])
 def local_gpualloc_memset_0(node):
     if isinstance(node.op, GpuAlloc) and not node.op.memset_0:
         inp = node.inputs[0]
@@ -1395,7 +1410,7 @@ def local_gpualloc_memset_0(node):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer([gpu_from_host, tensor.Eye])
 def local_gpu_eye(node):
     """
     gpu_from_host(eye) -> gpueye(gpu_from_host)
@@ -1479,7 +1494,7 @@ def tensor_to_cuda(x):
 
 
 @register_opt()
-@local_optimizer([])
+@local_optimizer(None) # XXX: linalg is in sandbox, so don't import it globally
 def local_gpu_extract_diagonal(node):
     """
     extract_diagonal(host_from_gpu()) -> host_from_gpu(extract_diagonal)
@@ -1505,7 +1520,7 @@ def local_gpu_extract_diagonal(node):
 
 
 @register_opt('scan')
-@local_optimizer([])
+@local_optimizer([gpu_from_host, scan_op.Scan])
 def gpuScanOptimization(node):
     """
     scan(host_from_gpu) -> host_from_gpu(GPUscan)

theano/sandbox/cuda/rng_curand.py

@@ -346,7 +346,7 @@ class CURAND_RandomStreams(object):
         return  rval
 
 
-@local_optimizer([None])
+@local_optimizer([CURAND_Base])
 def local_destructive(node):
     op = node.op
     if isinstance(op, CURAND_Base) and not op.destructive:

theano/sandbox/gpuarray/opt.py

@@ -112,7 +112,7 @@ gpu_seqopt.register('InputToGpuArrayOptimizer', InputToGpuOptimizer(),
                     0, 'fast_run', 'fast_compile', 'merge')
 
 
-@local_optimizer([])
+@local_optimizer([gpu_from_host, host_from_gpu])
 def local_cut_gpu_host_gpu(node):
     if tensor.opt.opt.check_chain(node, gpu_from_host, host_from_gpu):
         return [node.inputs[0].owner.inputs[0]]

theano/sandbox/linalg/ops.py

@@ -72,7 +72,7 @@ def hints(variable):
 
 
 @register_canonicalize
-@local_optimizer([])
+@local_optimizer([Hint])
 def remove_hint_nodes(node):
     if is_hint_node(node):
         # transfer hints from graph to Feature
@@ -224,7 +224,7 @@ def is_positive(v):
 
 
 @register_stabilize
-@local_optimizer([])
+@local_optimizer([Dot, Dot22])
 def inv_as_solve(node):
     if not imported_scipy:
         return False
@@ -242,7 +242,7 @@ def inv_as_solve(node):
 @register_canonicalize
 @register_stabilize
 @register_specialize
-@local_optimizer([])
+@local_optimizer([DimShuffle])
 def no_transpose_symmetric(node):
     if isinstance(node.op, DimShuffle):
         x = node.inputs[0]
@@ -253,7 +253,7 @@ def no_transpose_symmetric(node):
 
 
 @register_stabilize
-@local_optimizer([])
+@local_optimizer(None) # XXX: solve is defined later and can't be used here
 def psd_solve_with_chol(node):
     if node.op == solve:
         A, b = node.inputs  # result is solution Ax=b
@@ -269,7 +269,7 @@ def psd_solve_with_chol(node):
 
 @register_stabilize
 @register_specialize
-@local_optimizer([])
+@local_optimizer(None) # XXX: det is defined later and can't be used here
 def local_det_chol(node):
     """
     If we have det(X) and there is already an L=cholesky(X)
@@ -287,7 +287,7 @@ def local_det_chol(node):
 @register_canonicalize
 @register_stabilize
 @register_specialize
-@local_optimizer([])
+@local_optimizer([tensor.log])
 def local_log_prod_sqr(node):
     if node.op == tensor.log:
         x, = node.inputs
@@ -307,7 +307,7 @@ def local_log_prod_sqr(node):
 @register_canonicalize
 @register_stabilize
 @register_specialize
-@local_optimizer([])
+@local_optimizer([tensor.log])
 def local_log_pow(node):
     if node.op == tensor.log:
         x, = node.inputs

theano/sandbox/multinomial.py

@@ -337,7 +337,7 @@ class GpuMultinomialFromUniform(MultinomialFromUniform, GpuOp):
         """ % locals()
 
 
-@local_optimizer()
+@local_optimizer([MultinomialFromUniform])
 def local_gpu_multinomial(node):
     if type(node.op) is MultinomialFromUniform:
         p, u = node.inputs

theano/sandbox/rng_mrg.py

@@ -941,7 +941,7 @@ class MRG_RandomStreams(object):
         return final_samples
 
 
-@local_optimizer([None])
+@local_optimizer([mrg_uniform])
 def mrg_random_make_inplace(node):
     op = node.op
     if isinstance(op, mrg_uniform) and not op.inplace:

theano/scan_module/scan_opt.py

@@ -49,7 +49,7 @@ def info(*msg):
     _logger.info('INFO theano.scan: ' + ' '.join(msg))
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([scan_op.Scan])
 def remove_constants_and_unused_inputs_scan(node):
     '''
     Move constants into the inner graph, and remove unused inputs.
@@ -1337,7 +1337,7 @@ def make_equiv(lo, li):
     return left, right
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([scan_op.Scan])
 def scan_merge_inouts(node):
     if not isinstance(node.op, scan_op.Scan):
         return False

theano/sparse/opt.py

@@ -32,7 +32,7 @@ sparse.register_specialize(local_csm_properties_csm)
 
 
 # This is tested in tests/test_basic.py:test_remove0
-@gof.local_optimizer([None])
+@gof.local_optimizer([sparse.Remove0])
 def local_inplace_remove0(node):
     """
     Optimization to insert inplace versions of Remove0.
@@ -49,7 +49,7 @@ theano.compile.optdb.register('local_inplace_remove0',
                               gof.TopoOptimizer(local_inplace_remove0,
     failure_callback=gof.TopoOptimizer.warn_inplace),
                               60, 'fast_run', 'inplace')
-@gof.local_optimizer([None])
+@gof.local_optimizer([sparse.AddSD])
 def local_inplace_addsd(node):
     """
     Optimization to insert inplace versions of AddSD.

theano/tensor/blas.py

@@ -1645,7 +1645,7 @@ class Dot22(GemmRelated):
 _dot22 = Dot22()
 
 
-@local_optimizer([T._dot])
+@local_optimizer([T.Dot])
 def local_dot_to_dot22(node):
     # This works for tensor.outer too because basic.outer is a macro that
     # produces a dot(dimshuffle,dimshuffle) of form 4 below
@@ -2025,7 +2025,7 @@ blas_optdb.register('local_dot22_to_dot22scalar',
 
 #from opt import register_specialize, register_canonicalize
 #@register_specialize
-@local_optimizer([])
+@local_optimizer([T.sub, T.add])
 def local_print_as_we_go_along(node):
     if node.op in (T.sub, T.add):
         debugprint(node)

theano/tensor/nnet/conv3d2d.py

@@ -266,7 +266,7 @@ def make_gpu_optimizer(op, to_gpu):
     :param to_gpu: a list of op inputs that are moved to the GPU.
 
     """
-    @theano.gof.local_optimizer([])
+    @theano.gof.local_optimizer([op, cuda.gpu_from_host])
     def local_to_gpu(node):
         """
         op(host_from_gpu()) -> host_from_gpu(op)
@@ -302,7 +302,7 @@ if cuda.cuda_available:
     make_gpu_optimizer(IncDiagonalSubtensor, [0, 3])
 
 
-@theano.gof.local_optimizer([None])
+@theano.gof.local_optimizer([DiagonalSubtensor, IncDiagonalSubtensor])
 def local_inplace_DiagonalSubtensor(node):
     """ also work for IncDiagonalSubtensor """
     if (isinstance(node.op, (DiagonalSubtensor, IncDiagonalSubtensor)) and

theano/tensor/nnet/nnet.py

@@ -589,7 +589,7 @@ opt.local_mul_canonizer.add_simplifier(softmax_simplifier,
 
 if 0:
     @opt.register_specialize
-    @gof.local_optimizer([])
+    @gof.local_optimizer([tensor.add])
     def local_softmax_grad(node):
         '''dy*sm - DimShuffle{0,'x'}(sum{1}(dy*sm))*sm -> softmax_grad(dy,sm)'''
         #TODO what if the signs are changed?
@@ -1417,7 +1417,7 @@ def _is_const(z, val, approx=False):
 
 
 @opt.register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([subtensor.AdvancedSubtensor, tensor.log])
 def local_advanced_indexing_crossentropy_onehot(node):
     log = None
     sm = None

theano/tensor/opt.py

@@ -347,7 +347,7 @@ compile.optdb['canonicalize'].register(
 
 @register_canonicalize
 @register_stabilize
-@gof.local_optimizer([None])
+@gof.local_optimizer([T.Dot])
 def local_0_dot_x(node):
     if not isinstance(node.op, T.Dot):
         return False
@@ -390,7 +390,7 @@ def local_0_dot_x(node):
 ######################
 
 
-@gof.local_optimizer([None, None])
+@gof.local_optimizer([DimShuffle])
 def local_dimshuffle_lift(node):
     """
     "Lifts" DimShuffle through Elemwise operations and merges
@@ -431,7 +431,7 @@ def local_dimshuffle_lift(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.DimShuffle])
 def local_lift_transpose_through_dot(node):
     """
     dot(x,y).T -> dot(y.T, x.T)
@@ -456,7 +456,7 @@ def local_lift_transpose_through_dot(node):
         return [T.dot(y.T, x.T)]
 
 
-@gof.local_optimizer([])
+@gof.local_optimizer([DimShuffle])
 def dimshuffle_as_view(node):
     op = node.op
     if not isinstance(op, DimShuffle) or op.inplace:
@@ -476,7 +476,7 @@ register_specialize(local_dimshuffle_lift)
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.DimShuffle])
 def local_dimshuffle_no_inplace_at_canonicalize(node):
     if isinstance(node.op, T.DimShuffle) and node.op.inplace:
         return [T.DimShuffle(node.op.input_broadcastable,
@@ -1211,9 +1211,9 @@ def local_useless_alloc(node):
 
 @register_specialize
 @register_canonicalize
-@gof.local_optimizer([T._shape])
+@gof.local_optimizer([T.shape])
 def local_shape_to_shape_i(node):
-    if node.op == T._shape:
+    if node.op == T.shape:
         # This optimization needs ShapeOpt and fgraph.shape_feature
         if not hasattr(node.fgraph, 'shape_feature'):
             return
@@ -1221,9 +1221,10 @@ def local_shape_to_shape_i(node):
         return [shape_feature.make_vector_shape(node.inputs[0])]
 
 
+# TODO: Not sure what type of node we are expecting here
 @register_specialize
 @register_canonicalize
-@gof.local_optimizer([T._shape])
+@gof.local_optimizer(None)
 def local_track_shape_i(node):
     try:
         shape_feature = node.fgraph.shape_feature
@@ -1423,7 +1424,7 @@ def local_remove_useless_assert(node):
             return [assert_(node.inputs[0], *cond)]
 
 
-@gof.local_optimizer([T.Alloc])
+@gof.local_optimizer([T.Elemwise])
 def local_alloc_elemwise(node):
     """
     elemwise(alloc(x, shp), ..., y.TensorType(BROADCAST CONDITION))
@@ -1542,7 +1543,7 @@ else:
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Elemwise])
 def local_upcast_elemwise_constant_inputs(node):
     """This explicitly upcasts constant inputs to elemwise Ops, when
     those Ops do implicit upcasting anyway.
@@ -1690,7 +1691,7 @@ def local_useless_subtensor(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([Subtensor])
 def local_subtensor_lift(node):
     """
     unary(x)[idx] -> unary(x[idx])#any broadcast pattern.
@@ -1900,7 +1901,7 @@ def merge_two_slices(slice1, len1, slice2, len2):
 
 @register_canonicalize
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([Subtensor])
 def local_subtensor_merge(node):
     """
     Refactored optimization to deal with all cases of tensor merging.
@@ -1962,7 +1963,7 @@ def local_subtensor_merge(node):
 
 @register_canonicalize
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([Subtensor])
 def local_subtensor_of_alloc(node):
     """alloc[x:y] -> alloc"""
     if not isinstance(node.op, Subtensor):
@@ -2015,7 +2016,7 @@ def local_subtensor_of_alloc(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([None])
+@gof.local_optimizer([T.add])
 def local_IncSubtensor_serialize(node):
     """
     When using Subtensor, gradient graphs can be ugly.
@@ -2087,7 +2088,7 @@ compile.optdb.register('pre_local_IncSubtensor_serialize',
 #after priority 50 Destructive inplace operations
 #gemm is the first one now, at priority 70
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([IncSubtensor]) # XXX: GPU
 def local_inplace_setsubtensor(node):
     """
     Also work for GpuIncSubtensor
@@ -2106,7 +2107,7 @@ compile.optdb.register('local_inplace_setsubtensor',
                        'fast_run', 'inplace')  # DEBUG
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([AdvancedIncSubtensor1]) # XXX: GPU
 def local_inplace_incsubtensor1(node):
     """ also work for GpuAdvancedIncSubtensor1 """
     if isinstance(node.op, AdvancedIncSubtensor1) and not node.op.inplace:
@@ -2124,7 +2125,7 @@ compile.optdb.register('local_inplace_incsubtensor1',
 
 @register_canonicalize
 @register_stabilize
-@gof.local_optimizer([None])
+@gof.local_optimizer([IncSubtensor])
 def local_incsubtensor_of_allocs(node):
     """
     IncSubtensor(x, zeros, idx) -> x
@@ -2147,7 +2148,7 @@ def local_incsubtensor_of_allocs(node):
 
 @register_canonicalize
 @register_stabilize
-@gof.local_optimizer([None])
+@gof.local_optimizer([IncSubtensor])
 def local_setsubtensor_of_allocs(node):
     """
     SetSubtensor(x, x[idx], idx) -> x
@@ -2294,7 +2295,7 @@ def local_join_1(node):
 ###############
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Elemwise])
 def local_remove_switch_const_cond(node):
     """
     This optimization makes the following changes in the graph:
@@ -2377,7 +2378,7 @@ def local_mul_switch_sink(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([T.true_div])
+@gof.local_optimizer([T.true_div, T.int_div, T.floor_div])
 def local_div_switch_sink(node):
     """
     This optimization makes the folowing changes in the graph:
@@ -2421,7 +2422,7 @@ def local_div_switch_sink(node):
 ################
 @register_canonicalize
 @register_stabilize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Flatten])
 def local_flatten_lift(node):
     """
     Flatten(UnaryElemwise(x)) -> UnaryElemwise(Flatten(x))
@@ -2442,7 +2443,7 @@ def local_flatten_lift(node):
 ##################
 
 
-@gof.local_optimizer([None, None])
+@gof.local_optimizer([T.Reshape])
 def local_reshape_chain(node):
     """
     Reshape(Reshape(shape1),shape2) -> Reshape(shape2)
@@ -2470,7 +2471,7 @@ register_canonicalize(local_reshape_chain)
 
 @register_canonicalize
 @register_stabilize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Reshape])
 def local_reshape_lift(node):
     """
     Reshape(UnaryElemwise(x)) -> UnaryElemwise(Reshape(x))
@@ -2490,7 +2491,7 @@ def local_reshape_lift(node):
 if 0:
     # TODO: Test that this optimziation works.
     @register_canonicalize
-    @gof.local_optimizer([])
+    @gof.local_optimizer([T.Reshape])
     def local_scalar_reshape(node):
         """Eliminate reshape Ops whose inputs and outputs are scalars """
         if isinstance(node.op, T.Reshape):
@@ -2506,7 +2507,7 @@ if 0:
     # TODO: Remember to take into account the new sum dtype argument if this
     #       optimization is enabled.
     @register_canonicalize
-    @gof.local_optimizer([])
+    @gof.local_optimizer([T.Sum])
     def local_sum_over_empty(node):
         if isinstance(node.op, T.Sum):
             # This optimization needs ShapeOpt and fgraph.shape_feature
@@ -2528,7 +2529,7 @@ if 0:
 ##################
 
 
-@gof.local_optimizer([None, T.fill])
+@gof.local_optimizer([T.Elemwise])
 def local_fill_cut(node):
     """
     f(fill(a,b), c) -> f(b, c)
@@ -2582,7 +2583,7 @@ register_canonicalize(local_fill_cut)
 register_canonicalize(gof.OpRemove(T.tensor_copy), name='remove_tensor_copy')
 
 
-@gof.local_optimizer([None, T.fill])
+@gof.local_optimizer([T.Elemwise])
 def local_fill_sink(node):
     """
     f(fill(a, b), fill(c, d), e) -> fill(a, fill(c, f(b, d, e)))
@@ -2670,8 +2671,7 @@ class Canonizer(gof.LocalOptimizer):
         self.external_simplifiers.append((reason, simplifier))
 
     def tracks(self):
-        return [[self.main, None], [self.inverse, None],
-                [self.reciprocal, None]]
+        return [self.main, self.inverse, self.reciprocal]
 
     def get_num_denum(self, input):
         """
@@ -3059,7 +3059,7 @@ register_canonicalize(local_neg_to_mul)
 
 
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Sum])
 def local_sum_mul_by_scalar(node):
     """sum(scalar * smth) -> scalar * sum(smth)
        sum(-smth) -> -sum(smth)
@@ -3096,7 +3096,7 @@ def local_sum_mul_by_scalar(node):
 
 
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Elemwise])
 def local_elemwise_sub_zeros(node):
     """
     Elemwise{sub}(X,X) -> zeros_like(X)
@@ -3110,7 +3110,7 @@ def local_elemwise_sub_zeros(node):
 
 @register_canonicalize
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Sum])
 def local_sum_div_dimshuffle(node):
     '''sum(a / dimshuffle{...}(b), axis=l) -> sum(a, axis={...}) / b,
     if dimension l of the DimShuffle is 'x'.'''
@@ -3199,7 +3199,7 @@ def local_sum_div_dimshuffle(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Sum])
 def local_sum_all_to_none(node):
     """Sum{0,1,...N} -> Sum{}"""
     if isinstance(node.op, T.Sum):
@@ -3212,7 +3212,7 @@ def local_sum_all_to_none(node):
 
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Sum])
 def local_sum_sum(node):
     """
     Sum(Sum()) -> Sum
@@ -3278,9 +3278,12 @@ def local_sum_sum(node):
                 combined_sum = T.Sum(newaxis, dtype=out_dtype)
                 return [combined_sum(summed.owner.inputs[0])]
 
+ALL_REDUCE = [T.elemwise.CAReduce, T.elemwise.All, T.elemwise.Any,
+              T.elemwise.Sum, T.elemwise.Prod,
+              T.elemwise.ProdWithoutZeros]
 
 @register_canonicalize
-@gof.local_optimizer([])
+@gof.local_optimizer(ALL_REDUCE)
 def local_cut_useless_reduce(node):
     """Sum(a, axis=[]) -> a  """
     if isinstance(node.op, T.CAReduce):
@@ -3296,7 +3299,7 @@ def local_cut_useless_reduce(node):
 #
 #@register_canonicalize
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer(ALL_REDUCE)
 def local_reduce_broadcastable(node):
     """Remove reduction over broadcastable dimensions"""
     if isinstance(node.op, T.CAReduce):
@@ -3335,7 +3338,7 @@ def local_reduce_broadcastable(node):
 
 
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.Sum])
 def local_sum_alloc(node):
     """ sum(alloc(constant,shapes...)) => constant*prod(shapes)"""
     if isinstance(node.op, T.Sum):
@@ -3742,7 +3745,7 @@ def local_abs_lift(node):
 
 
 @register_specialize
-@gof.local_optimizer([])
+@gof.local_optimizer([T.mul, T.true_div])
 def local_abs_merge(node):
     """
     merge abs generated by local_abs_lift when the canonizer don't
@@ -3917,8 +3920,7 @@ def attempt_distribution(factor, num, denum):
                                    neg_pairs))), num, denum
 
 
-@gof.local_optimizer([T.mul, T.add, T.mul], [T.mul, T.sub, T.mul],
-                     [T.mul, T.add, T.true_div], [T.mul, T.sub, T.true_div])
+@gof.local_optimizer([T.mul])
 def local_greedy_distributor(node):
     """
     This optimization tries to apply distributivity of multiplication
@@ -3984,7 +3986,7 @@ register_canonicalize(local_greedy_distributor)
 register_stabilize(local_greedy_distributor)
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer(None)
 def constant_folding(node):
     for input in node.inputs:
         if not isinstance(input, Constant):

theano/tensor/opt_uncanonicalize.py

@@ -55,7 +55,7 @@ def local_max_and_argmax(node):
             return [new, None]
 
 @register_uncanonicalize
-@gof.local_optimizer([T._shape])
+@gof.local_optimizer([T.neg])
 def local_max_to_min(node):
     """
     change -(max(-x)) to min

theano/tensor/raw_random.py

@@ -816,7 +816,7 @@ def multinomial(random_state, size=None, n=1, pvals=[0.5, 0.5],
     return op(random_state, size, n, pvals)
 
 
-@gof.local_optimizer([None])
+@gof.local_optimizer([RandomFunction])
 def random_make_inplace(node):
     op = node.op
     if isinstance(op, RandomFunction) and not op.inplace:

theano/tensor/tests/test_basic.py

@@ -3361,10 +3361,8 @@ class T_Join_and_Split(unittest.TestCase):
         utt.verify_grad((lambda a, b: join(0, a, b)), [a_val, b_val], rng=rng)
 
     def test_broadcastable_single_input_broadcastable_dimension(self):
-        """
-        Test that all broadcastable flags are preserved by a
-        single-input join.
-        """
+        # Test that all broadcastable flags are preserved by a
+        # single-input join.
         rng = numpy.random.RandomState(seed=utt.fetch_seed())
         a_val = rng.rand(1, 4, 1).astype(self.floatX)
         a = self.shared(a_val, broadcastable=(True, False, True))
@@ -3387,10 +3385,8 @@ class T_Join_and_Split(unittest.TestCase):
         #self.assertRaises(TypeError, f, bad_a_val)
 
     def test_broadcastable_flags_many_dims_and_inputs(self):
-        """
-        Test that the right broadcastable flags get set for a join
-        with many inputs and many input dimensions.
-        """
+        # Test that the right broadcastable flags get set for a join
+        # with many inputs and many input dimensions.
         a = TensorType(dtype=self.floatX, broadcastable=[1, 0, 1, 0, 0, 0])()
         b = TensorType(dtype=self.floatX, broadcastable=[1, 1, 1, 0, 0, 0])()
         c = TensorType(dtype=self.floatX, broadcastable=[1, 0, 0, 0, 0, 0])()
@@ -3479,20 +3475,16 @@ class T_Join_and_Split(unittest.TestCase):
             f(get_mat(3, 4), get_mat(3, 4), get_mat(2, 5))
 
     def test_rebroadcast(self):
-        """
-        Regression test for a crash that used to happen when rebroadcasting.
-        """
+        # Regression test for a crash that used to happen when rebroadcasting.
         x = tensor.TensorType(self.floatX, [False, False, True])()
         u = tensor.TensorType(self.floatX, [False, False, True])()
         # This line used to crash.
         z = tensor.concatenate([x, -u], axis=2)
 
     def test_concatenate_same(self):
-        """
-        Test that we can concatenate the same tensor multiple time.
+        # Test that we can concatenate the same tensor multiple time.
 
-        In the past it was broken on the GPU.
-        """
+        # In the past it was broken on the GPU.
         rng = numpy.random.RandomState(seed=utt.fetch_seed())
         T_shared = self.shared(rng.rand(3, 4).astype(self.floatX))
         Tout = tensor.concatenate([T_shared, T_shared])