Merge pull request #5013 from nouiz/gpu_alloc_empty

theano/gpuarray/basic_ops.py

@@ -7,7 +7,8 @@ import numpy
 from theano import Op, Apply, Type, Variable
 from theano import tensor, config
 from theano.gradient import grad_undefined
-from theano.tensor.basic import Alloc, Join, Split
+from theano.tensor.basic import (
+    Alloc, AllocEmpty, alloc_validate_shape, Join, Split)
 
 from theano.gof import HideC, COp
 from theano.gof.utils import MethodNotDefined
@@ -805,7 +806,7 @@ class GpuAlloc(HideC, Alloc):
 
     def make_node(self, value, *shape):
         value = as_gpuarray_variable(value, context_name=self.context_name)
-        sh, bcast = self.validate_shape(shape)
+        sh, bcast = alloc_validate_shape(shape)
         if value.ndim > len(sh):
             TypeError("The GpuAlloc value to use has more dimensions "
                       "than the specified shape", value.ndim, len(sh))
@@ -941,7 +942,7 @@ def gpu_alloc(ctx, memset_0=False):
 gpu_alloc.cache = {}
 
 
-class GpuAllocEmpty(HideC, Alloc):
+class GpuAllocEmpty(HideC, AllocEmpty):
     """
     Allocate uninitialized memory on the GPU.
 
@@ -958,7 +959,7 @@ class GpuAllocEmpty(HideC, Alloc):
         return get_context(self.context_name)
 
     def make_node(self, *shape):
-        sh, bcast = self.validate_shape(shape)
+        sh, bcast = alloc_validate_shape(shape)
         output = GpuArrayType(dtype=self.dtype, broadcastable=bcast,
                               context_name=self.context_name)()
         output.tag.values_eq_approx = tensor.type.values_eq_approx_always_true

theano/sandbox/cuda/basic_ops.py

@@ -3658,25 +3658,10 @@ class GpuAllocEmpty(GpuOp):
 
     __props__ = ()
 
-    @staticmethod
-    def validate_shape(shape):
-        sh = [tensor.as_tensor_variable(s) for s in shape]
-        bcast = []
-        for s in sh:
-            if s.type.dtype[:3] not in ('int', 'uin'):
-                raise TypeError('Shape arguments must be integers', s)
-            # if s is constant 1, then we're broadcastable in that dim
-            try:
-                const_shp = tensor.get_scalar_constant_value(s)
-            except tensor.NotScalarConstantError:
-                const_shp = None
-            bcast.append(1 == const_shp)
+    def make_node(self, *shape):
+        shape, bcast = tensor.basic.alloc_validate_shape(shape)
         otype = CudaNdarrayType(dtype='float32', broadcastable=bcast)
         output = otype()
-        return sh, output
-
-    def make_node(self, *shape):
-        shape, output = self.validate_shape(shape)
         output.tag.values_eq_approx = tensor.type.values_eq_approx_always_true
         # The outut can contain nan/inf.  output.type is a new
         # instance, so we can do this only for that variable.
@@ -3767,7 +3752,11 @@ class GpuAlloc(GpuAllocEmpty):
         # if there is unneeded transfert generated by the next line
         # the optimizer will remove them.
         v = as_cuda_ndarray_variable(value)
-        shape, output = self.validate_shape(shape)
+
+        shape, bcast = tensor.basic.alloc_validate_shape(shape)
+        otype = CudaNdarrayType(dtype='float32', broadcastable=bcast)
+        output = otype()
+
         return Apply(self, [v] + shape, [output])
 
     # This is required because the superclass (GpuAllocEmpty) also has it.

theano/tensor/basic.py

@@ -2726,27 +2726,7 @@ def identity_like(x):
     return eye(x.shape[0], x.shape[1], k=0, dtype=x.dtype)
 
 
-class Alloc(gof.Op):
-    """Create a Tensor from an initial value and a desired shape.
-
-    alloc(value, shape0, shape1, ..., shapeN)
-
-    Returns an N-dimensional tensor initialized by `value` using something
-    equivalent to
-
-        z = numpy.zeros(shape, value.dtype)
-        z += value
-
-    The result has N dimensions, has the dtype of `value` and is obtained by
-    broadcasting value over the output ndarray.
-
-    This Op is used to replace fill() during optimizations because after shapes
-    are lifted, the first argument to fill can often be pruned from the graph.
-
-    """
-    __props__ = ()
-
-    def validate_shape(self, shape):
+def alloc_validate_shape(shape):
     sh = [as_tensor_variable(s) for s in shape]
     bcast = []
     for i, s in enumerate(sh):
@@ -2775,9 +2755,33 @@ class Alloc(gof.Op):
         bcast.append(1 == const_shp)
     return sh, bcast
 
+
+class Alloc(gof.Op):
+    """Create a Tensor from an initial value and a desired shape.
+
+    alloc(value, shape0, shape1, ..., shapeN)
+
+    Returns an N-dimensional tensor initialized by `value` using something
+    equivalent to
+
+        z = numpy.zeros(shape, value.dtype)
+        z += value
+
+    The result has N dimensions, has the dtype of `value` and is obtained by
+    broadcasting value over the output ndarray.
+
+    This Op is used to replace fill() during optimizations because after shapes
+    are lifted, the first argument to fill can often be pruned from the graph.
+
+    """
+    __props__ = ()
+
+    def validate_shape(self, shape):
+        return alloc_validate_shape(shape)
+
     def make_node(self, value, *shape):
         v = as_tensor_variable(value)
-        sh, bcast = self.validate_shape(shape)
+        sh, bcast = alloc_validate_shape(shape)
         if v.ndim > len(sh):
             raise TypeError("The Alloc value to use has more dimensions"
                             " than the specified dimensions",
@@ -6356,24 +6360,11 @@ class AllocEmpty(gof.Op):
         assert isinstance(dtype, str), dtype
         self.dtype = dtype.lower()
 
-    def validate_shape(self, shape):
-        sh = [as_tensor_variable(s) for s in shape]
-        bcast = []
-        for s in sh:
-            if s.type.dtype[:3] not in ('int', 'uin'):
-                raise TypeError('Shape arguments must be integers', s)
-            # if s is constant 1, then we're broadcastable in that dim
-            try:
-                const_shp = get_scalar_constant_value(s)
-            except NotScalarConstantError:
-                const_shp = None
-            bcast.append(1 == const_shp)
+    def make_node(self, *shape):
+        shape, bcast = alloc_validate_shape(shape)
         otype = TensorType(dtype=self.dtype, broadcastable=bcast)
         output = otype()
-        return sh, output
 
-    def make_node(self, *shape):
-        shape, output = self.validate_shape(shape)
         output.tag.values_eq_approx = values_eq_approx_always_true
         # The outut can contain nan/inf.  output.type is a new
         # instance, so we can do this only for that variable.