old interface compatible

theano/compile/debugmode.py

@@ -1879,7 +1879,7 @@ class _Linker(gof.link.LocalLinker):
                 thunk.outputs = [storage_map[v] for v in node.outputs]
                 thunk_other = thunk
             else:
-                new_node = node.op.prepare_node(node)
+                new_node = node.op.prepare_node(node, storage_map, compute_map)
                 if new_node is not None:
                     node = new_node
 

theano/gof/op.py

@@ -836,7 +836,7 @@ class Op(utils.object2, PureOp, CLinkerOp):
         else:
             return NotImplemented
 
-    def prepare_node(self, node):
+    def prepare_node(self, node, storage_map, compute_map):
         """
         Make any special modifications that the Op needs before doing
         make_thunk().
@@ -955,7 +955,8 @@ class Op(utils.object2, PureOp, CLinkerOp):
         """
         logger = logging.getLogger('theano.gof.op.Op')
 
-        new_node = self.prepare_node(node)
+        new_node = self.prepare_node(node, storage_map=storage_map,
+                                     compute_map=compute_map)
         if new_node is not None:
             node = new_node
 
@@ -1214,7 +1215,8 @@ int main( int argc, const char* argv[] )
                 self.openmp = False
                 theano.config.openmp = False
 
-    def prepare_node(self, node):
+    def prepare_node(self, node, storage_map,
+                     compute_map):
         self.update_self_openmp()
 
 

theano/sandbox/cuda/dnn.py

@@ -1379,6 +1379,24 @@ class GpuDnnPool(DnnBase):
         assert mode in ('max', 'average_inc_pad', 'average_exc_pad')
         self.mode = mode
 
+    def prepare_node(self, node, storage_map, compute_map):
+        if len(node.inputs) == 2:
+            warnings.warn("Theano GPUDnnPoolGrad internal changed.", stacklevel=3)
+            # Old interface
+            self.mode = node.inputs[1].owner.op.mode
+            ws = theano.tensor.constant(node.inputs[1].owner.op.ws)
+            st = theano.tensor.constant(node.inputs[1].owner.op.stride)
+            pad = theano.tensor.constant(node.inputs[1].owner.op.pad)
+            node.inputs[1] = ws
+            node.inputs.append(st)
+            node.inputs.append(pad)
+            storage_map[ws] = [None]
+            storage_map[st] = [None]
+            storage_map[pad] = [None]
+            compute_map[ws] = [False]
+            compute_map[st] = [False]
+            compute_map[pad] = [False]
+
     def make_node(self, img, ws, stride, pad):
         img = as_cuda_ndarray_variable(img)
         assert (img.ndim in [4, 5])
@@ -1442,13 +1460,6 @@ if (pool%(name)s != NULL) { cudnnDestroyPoolingDescriptor(pool%(name)s); }
 """ % dict(name=name)
 
     def c_code(self, node, name, inputs, outputs, sub):
-        if len(inputs) == 2:
-            warnings.warn("Theano GPUDnnPoolGrad internal changed.", stacklevel=3)
-            desc = inputs[1]
-            ws = desc.ws
-            stride = desc.stride
-            pad = desc.pad
-        else:
         ws = inputs[1]
         stride = inputs[2]
         pad = inputs[3]
@@ -1593,6 +1604,24 @@ class GpuDnnPoolGrad(DnnBase):
         assert mode in ('max', 'average_inc_pad', 'average_exc_pad')
         self.mode = mode
 
+    def prepare_node(self, node, storage_map, compute_map):
+        if len(node.inputs) == 4:
+            warnings.warn("Theano GPUDnnPoolGrad internal changed.", stacklevel=3)
+            # Old interface
+            self.mode = node.inputs[3].owner.op.mode
+            ws = theano.tensor.constant(node.inputs[3].owner.op.ws)
+            st = theano.tensor.constant(node.inputs[3].owner.op.stride)
+            pad = theano.tensor.constant(node.inputs[3].owner.op.pad)
+            node.inputs[3] = ws
+            node.inputs.append(st)
+            node.inputs.append(pad)
+            storage_map[ws] = [None]
+            storage_map[st] = [None]
+            storage_map[pad] = [None]
+            compute_map[ws] = [False]
+            compute_map[st] = [False]
+            compute_map[pad] = [False]
+
     def make_node(self, inp, out, inp_grad, ws, stride, pad):
         inp = as_cuda_ndarray_variable(inp)
         assert (inp.ndim in [4, 5])
@@ -1675,13 +1704,6 @@ if (pool%(name)s != NULL) { cudnnDestroyPoolingDescriptor(pool%(name)s); }
         # Here the name out and inp are based on the cudnn definition.
         # Not the definition of this class.
         # This make it complicated.
-        if len(inputs) == 4:
-            warnings.warn("Theano GPUDnnPoolGrad internal changed.", stacklevel=3)
-            out, inp, inp_grad, desc = inputs
-            ws = desc.ws
-            stride = desc.stride
-            pad = desc.pad
-        else:
         out, inp, inp_grad, ws, stride, pad = inputs
 
         out_grad, = outputs

theano/sandbox/cuda/tests/test_dnn.py

@@ -347,6 +347,8 @@ def test_pooling():
 
 
 def test_pooling_with_tensor_vars():
+    if not cuda.dnn.dnn_available():
+        raise SkipTest(cuda.dnn.dnn_available.msg)
     x = T.ftensor4()
     ws = theano.shared(numpy.array([2, 2], dtype='int32'))
     st = theano.shared(numpy.array([1, 1], dtype='int32'))
@@ -393,6 +395,13 @@ def test_pooling_with_tensor_vars():
         utt.assert_allclose(a, b)
 
 
+def test_old_pool_interface():
+    if not cuda.dnn.dnn_available():
+        raise SkipTest(cuda.dnn.dnn_available.msg)
+    import pickle
+    pickle.load(open('old_pool_interface.pkl', 'rb'))
+
+
 def test_pooling3d():
     # CuDNN 3d pooling requires CuDNN v3. Don't test if the CuDNN version is
     # too old.
@@ -654,7 +663,8 @@ class test_DnnSoftMax(test_nnet.test_SoftMax):
             input_val = numpy.random.normal(0, 1, inp_shape).astype("float32")
 
             out = f(input_val)
-            expected_out = numpy.log(numpy.exp(input_val) /
+            expected_out = numpy.log(
+                numpy.exp(input_val) /
                 numpy.exp(input_val).sum(1)[:, None, :, :])
 
             utt.assert_allclose(out, expected_out)

theano/tensor/elemwise.py

@@ -792,7 +792,7 @@ class Elemwise(OpenMPOp):
 
         return ret
 
-    def prepare_node(self, node):
+    def prepare_node(self, node, storage_map, compute_map):
         # Postpone the ufunc building to the last minutes
         # NumPy ufunc support only up to 31 inputs.
         # But our c code support more.