Merge pull request #1923 from nouiz/sort_grad

theano/tensor/sort.py

@@ -3,7 +3,7 @@ import numpy as np
 import theano
 from theano.tensor import tensor
 
-from theano.tensor.basic import mul
+from theano.tensor.basic import mul, arange
 
 
 class SortOp(theano.Op):
@@ -27,7 +27,8 @@ class SortOp(theano.Op):
 
     def make_node(self, input, axis=-1):
         input = theano.tensor.as_tensor_variable(input)
-        if axis is None:
+        if (axis is None or
+            (isinstance(axis, theano.Constant) and axis.data is None)):
             axis = theano.Constant(theano.gof.generic, None)
             # axis=None flattens the array before sorting
             out_type = tensor(dtype=input.dtype, broadcastable=[False])
@@ -55,8 +56,35 @@ class SortOp(theano.Op):
         assert inputs_shapes[1] == ()
         return [inputs_shapes[0]]
 
-    #**** It need the argsort, so we can't do it now.
-    #def grad(self, inputs, output_grads):
+    def grad(self, inputs, output_grads):
+        a, axis = inputs
+        inp_grad = theano.gradient.grad_not_implemented(
+            self, 0, axis,
+            "Currently, we only implement the gradient on sort for vector"
+            " and matrix (and axis is None)")
+        if a.ndim == 1:
+            idx = argsort(*inputs, kind=self.kind, order=self.order)
+#            rev_idx = numpy.where(idx[None, :]==numpy.arange(5)[:,None])[1]
+            rev_idx = theano.tensor.eq(idx[None, :],
+                                       arange(a.shape[0])[:, None]).nonzero()[1]
+            inp_grad = output_grads[0][rev_idx]
+        elif a.ndim == 2:
+            if (axis is None or
+                (isinstance(axis, theano.Constant) and axis.data is None)):
+                idx = argsort(*inputs, kind=self.kind, order=self.order)
+                rev_idx = theano.tensor.eq(idx[None, :],
+                                           arange(a.shape[0]*a.shape[1])[:, None]).nonzero()[1]
+                inp_grad = output_grads[0][rev_idx].reshape(a.shape)
+            elif (axis == 0 or
+                  (isinstance(axis, theano.Constant) and axis.data == 0)):
+                idx = argsort(*inputs, kind=self.kind, order=self.order)
+                #not working: numpy.where(idx[None, :]==numpy.arange(2)[:, None, None])
+                pass
+        axis_grad = theano.gradient.grad_undefined(
+            self, 1, axis,
+            "sort is not defined for non-integer axes so"
+            " sort(x, axis+eps) is undefined")
+        return [inp_grad, axis_grad]
     """
     def R_op(self, inputs, eval_points):
         # R_op can receive None as eval_points.
@@ -115,7 +143,8 @@ class ArgSortOp(theano.Op):
 
     def make_node(self, input, axis=-1):
         input = theano.tensor.as_tensor_variable(input)
-        if axis is None:
+        if (axis is None or
+            (isinstance(axis, theano.Constant) and axis.data is None)):
             axis = theano.Constant(theano.gof.generic, None)
             bcast = [False]
         else:

theano/tensor/tests/test_sort.py

@@ -68,6 +68,21 @@ class test_sort(unittest.TestCase):
         gt = np.sort(self.m_val, None)
         assert np.allclose(gv, gt)
 
+    def test_grad_vector(self):
+        a = theano.tensor.vector()
+        data = np.random.rand(10).astype(theano.config.floatX)
+        utt.verify_grad(sort, [data])
+
+    def test_grad_none_axis(self):
+        data = np.random.rand(10).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, None), [data])
+        utt.verify_grad(lambda x: sort(x, 0), [data])
+
+        data = np.random.rand(2, 3).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, None), [data])
+        #utt.verify_grad(lambda x: sort(x, 0), [data])
+        #utt.verify_grad(lambda x: sort(x, 1), [data])
+
 
 class TensorInferShapeTester(utt.InferShapeTester):
     def test_sort(self):