Fix recursive import. Do not make dnn import opt. opt stil import dnn

theano/sandbox/cuda/__init__.py

@@ -9,11 +9,28 @@ import sys
 import theano
 from theano.compat import get_unbound_function
 from theano.compile import optdb
+from theano.gof import EquilibriumDB
 from theano.gof.cmodule import get_lib_extension
 from theano.gof.compilelock import get_lock, release_lock
 from theano.configparser import config, AddConfigVar, StrParam, BoolParam
 import nvcc_compiler
 
+# ignore_newtrees is to speed the optimization as this is the pattern
+# we use for optimization. Otherwise, we can iterate 100s of time on
+# the graph and apply only a few optimizations each time.
+gpu_optimizer = EquilibriumDB(ignore_newtrees=False)
+
+
+def register_opt(*tags, **kwargs):
+    def f(local_opt):
+        name = (kwargs and kwargs.pop('name')) or local_opt.__name__
+        gpu_optimizer.register(name, local_opt, 'fast_run', 'fast_compile',
+                               'gpu', *tags)
+        return local_opt
+    return f
+
+
+
 _logger_name = 'theano.sandbox.cuda'
 _logger = logging.getLogger(_logger_name)
 

theano/sandbox/cuda/dnn.py

@@ -13,7 +13,7 @@ from theano.sandbox.cuda.basic_ops import (as_cuda_ndarray_variable,
 from theano.sandbox.cuda.blas import (GpuConv, GpuDownsampleFactorMax,
                                       GpuDownsampleFactorMaxGrad)
 from theano.sandbox.cuda.nnet import GpuSoftmax
-from theano.sandbox.cuda.opt import register_opt
+from theano.sandbox.cuda import register_opt
 
 from theano.sandbox.cuda.nvcc_compiler import NVCC_compiler
 

theano/sandbox/cuda/opt.py

@@ -42,6 +42,7 @@ from theano.sandbox.cuda.elemwise import SupportCodeError
 from theano.scalar.basic_scipy import Erfinv
 from theano.sandbox.cuda.elemwise import erfinv_gpu
 from theano.sandbox.cuda.var import CudaNdarrayConstant
+from theano.sandbox.cuda import gpu_optimizer, register_opt
 from theano.scan_module import scan_utils, scan_op, scan_opt
 from theano.tensor.blas import _is_real_vector, _is_real_matrix
 from theano.tensor import nlinalg
@@ -56,10 +57,6 @@ except ImportError:
 
 #optdb.print_summary()  # shows what is currently registered
 
-#ignore_newtrees is to speed the optimization as this is the pattern
-#we use for optimization. Otherwise, we can iterate 100s of time on
-#the graph and apply only a few optimizations each time.
-gpu_optimizer = EquilibriumDB(ignore_newtrees=False)
 gpu_cut_copies = EquilibriumDB()
 gpu_seqopt = SequenceDB()
 gpu_seqopt.register('gpu_local_optimizations', gpu_optimizer, 1,
@@ -84,14 +81,6 @@ gpu_optimizer.register('gpu_merge', theano.gof.opt.merge_optimizer,
                        'fast_run', 'fast_compile')
 
 
-def register_opt(*tags, **kwargs):
-    def f(local_opt):
-        name = (kwargs and kwargs.pop('name')) or local_opt.__name__
-        gpu_optimizer.register(name, local_opt, 'fast_run', 'fast_compile',
-                               'gpu', *tags)
-        return local_opt
-    return f
-
 #register local_track_shape_i at this level too
 #to make multi-level lift of shape work.
 register_opt()(theano.tensor.opt.local_track_shape_i)