Merge pull request #4932 from nouiz/no_complex

theano/gpuarray/basic_ops.py

@@ -626,6 +626,8 @@ class GpuFromHost(Op):
     def make_node(self, x):
         if not isinstance(x.type, tensor.TensorType):
             raise TypeError(x)
+        if "complex" in x.dtype:
+            raise TypeError("complex not supported in the new gpuarray back-end.", x)
         return Apply(self, [x], [GpuArrayType(broadcastable=x.broadcastable,
                                               context_name=self.context_name,
                                               dtype=x.dtype)()])

theano/gpuarray/dnn.py

@@ -1270,7 +1270,7 @@ def dnn_pool(img, ws, stride=(1, 1), mode='max', pad=(0, 0)):
         Subsampling window size.
     stride : tuple
         Subsampling stride (default: (1, 1)).
-    mode : {'max', 'average_inc_pad', 'average_exc_pad'}
+    mode : {'max', 'average_inc_pad', 'average_exc_pad', 'sum'}
     pad : tuple
         (padX, padY) or (padX, padY, padZ)
         default: (0, 0)
@@ -1285,6 +1285,11 @@ def dnn_pool(img, ws, stride=(1, 1), mode='max', pad=(0, 0)):
 
     """
     img = gpu_contiguous(img)
+    if mode == "sum":
+        ret = GpuDnnPool(mode="average_inc_pad")(img, ws, stride, pad)
+        context_name = ret.type.context_name
+        window_elem = theano.tensor.prod(ws).astype(ret.dtype)
+        return as_gpuarray_variable(ret * window_elem, context_name)
     return GpuDnnPool(mode=mode)(img, ws, stride, pad)
 
 

theano/gpuarray/opt.py

@@ -191,8 +191,9 @@ def op_lifter(OP, cuda_only=False):
 
                 # Check if we should replace
                 if (not replace or
-                    (cuda_only and
-                     get_context(context_name).kind != b'cuda')):
+                        (cuda_only and
+                         get_context(context_name).kind != b'cuda') or
+                        any(["complex" in i.dtype for i in node.inputs])):
                     return False
 
                 # tag the inputs with the context in case
@@ -298,7 +299,8 @@ class GraphToGPU(Optimizer):
         for i in fgraph.inputs:
             # Do not move *int* scalar to the GPU.
             if (isinstance(i.type, tensor.TensorType) and
-               (i.ndim > 0 or 'int' not in i.dtype)):
+                    (i.ndim > 0 or 'int' not in i.dtype) and
+                    "complex" not in i.dtype):
                 mapping[i] = i.transfer(getattr(i.tag, 'target', target))
             else:
                 mapping[i] = i
@@ -344,6 +346,10 @@ class GraphToGPU(Optimizer):
                          self.local_optimizers_map.get(type(c.op), []))):
                         move_to_GPU = True
             new_ops = None
+            if move_to_GPU and any(["complex" in getattr(i, 'dtype', "")
+                                    for i in node.inputs]):
+                move_to_GPU = False
+
             # Apply the lifter
             if move_to_GPU:
                 for lopt in (self.local_optimizers_map.get(node.op, []) +

theano/gpuarray/tests/test_dnn.py

@@ -356,6 +356,18 @@ def test_pooling_opt():
 
     f(numpy.zeros((10, 10), dtype='float32'))
 
+    # Test sum pooling
+    f = theano.function(
+        [x],
+        pool_2d(x, ds=(2, 3), mode='sum',
+                ignore_border=True),
+        mode=mode_with_gpu)
+
+    assert any([isinstance(n.op, dnn.GpuDnnPool)
+                for n in f.maker.fgraph.toposort()])
+    data = numpy.random.rand(10, 10).astype('float32')
+    f(data)
+
 
 def test_dnn_tag():
     """

theano/gpuarray/tests/test_opt.py

@@ -472,3 +472,12 @@ def test_local_assert_no_cpu_op():
         theano.function([], out, mode=mode_local_assert)
     finally:
         theano.config.assert_no_cpu_op = old
+
+
+def test_no_complex():
+    width_var = tensor.cscalar()
+    freq_var = tensor.fscalar()
+    signal_var = tensor.fscalar()
+    stft_out = tensor.exp(width_var * freq_var) * signal_var
+    theano.function([width_var, freq_var, signal_var], stft_out,
+                    mode=mode_with_gpu)

theano/sandbox/cuda/dnn.py

@@ -1991,7 +1991,7 @@ def dnn_pool(img, ws, stride=(1, 1), mode='max', pad=(0, 0)):
         Subsampling window size.
     stride
         Subsampling stride (default: (1, 1)).
-    mode : {'max', 'average_inc_pad', 'average_exc_pad}
+    mode : {'max', 'average_inc_pad', 'average_exc_pad, 'sum'}
     pad :
         (pad_h, pad_w) padding information.
         pad_h is the number of zero-valued pixels added to each of the top and
@@ -2009,6 +2009,11 @@ def dnn_pool(img, ws, stride=(1, 1), mode='max', pad=(0, 0)):
 
     """
     img = gpu_contiguous(img)
+    if mode == "sum":
+        ret = GpuDnnPool(mode="average_inc_pad")(img, ws, stride, pad)
+        window_elem = theano.tensor.prod(ws).astype(ret.dtype)
+        return as_cuda_ndarray_variable(ret * window_elem)
+
     return GpuDnnPool(mode=mode)(img, ws, stride, pad)
 
 

theano/sandbox/cuda/tests/test_dnn.py

@@ -545,6 +545,17 @@ def test_pooling_opt():
                 for n in f.maker.fgraph.toposort()])
 
     f(numpy.zeros((10, 10), dtype='float32'))
+    # Test sum pooling
+    f = theano.function(
+        [x],
+        pool_2d(x, ds=(2, 3), mode='sum',
+                ignore_border=True),
+        mode=mode_with_gpu)
+
+    assert any([isinstance(n.op, dnn.GpuDnnPool)
+                for n in f.maker.fgraph.toposort()])
+    data = numpy.random.rand(10, 10).astype('float32')
+    f(data)
 
 
 class test_DnnSoftMax(test_nnet.test_SoftMax):