Merge pull request #2910 from noskill/master

theano/tensor/sort.py

@@ -61,7 +61,7 @@ class SortOp(theano.Op):
         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 or 0)")
+            " matrix (and axis is None or 0) and tensor3")
         if a.ndim == 1:
             idx = argsort(*inputs, kind=self.kind, order=self.order)
 #            rev_idx = numpy.where(idx[None, :]==numpy.arange(5)[:,None])[1]
@@ -80,11 +80,51 @@ class SortOp(theano.Op):
                 idx = argsort(*inputs, kind=self.kind, order=self.order)
                 # not working: numpy.where(idx[None, :]==numpy.arange(2)[:, None, None])
                 pass
+        elif a.ndim == 3:
+            if isinstance(axis, theano.Constant) and axis.data is not None:
+                indices = self.__get_argsort_indices(a, axis)
+                inp_grad = output_grads[0][indices[0], indices[1], indices[2]]
+            elif (axis is None or
+                (isinstance(axis, theano.Constant) and axis.data is None)):
+                rev_idx = self.__get_argsort_indices(a, axis)
+                inp_grad = output_grads[0][rev_idx].reshape(a.shape)
         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 __get_argsort_indices(self, a, axis):
+        """Calculates indices which can be used to reverse
+        sorting operation of "a" tensor along "axis"
+
+        returns:
+          1d array if axis is None
+          list of lenght len(a.shape) otherwise
+        """
+
+        # The goal is to get gradient wrt input from gradient 
+        # wrt sort(input, axis)
+        idx = argsort(a, axis, kind=self.kind, order=self.order)
+        # rev_idx is the reverse of previous argsort operation 
+        rev_idx = argsort(idx, axis, kind=self.kind, order=self.order) 
+        if (axis is None or
+            (isinstance(axis, theano.Constant) and axis.data is None)):
+            return rev_idx
+        indices = []
+        if axis.data >= 0:
+            axis_data = axis.data
+        else:
+            axis_data = a.ndim + axis.data
+        for i in range(a.ndim):
+            if i == axis_data:
+                indices.append(rev_idx)
+            else:
+                index_shape = [1] * a.ndim 
+                index_shape[i] = a.shape[i]
+                # it's a way to emulate numpy.ogrid[0: a.shape[0], 0: a.shape[1], 0: a.shape[2]]
+                indices.append(theano.tensor.arange(a.shape[i]).reshape(index_shape))
+        return indices
     """
     def R_op(self, inputs, eval_points):
         # R_op can receive None as eval_points.

theano/tensor/tests/test_sort.py

@@ -82,6 +82,24 @@ class test_sort(unittest.TestCase):
         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])
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, None), [data])
+
+    def test_grad_negative_axis(self):
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, -1), [data])
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, -2), [data])
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, -3), [data])
+
+    def test_grad_nonnegative_axis(self):
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, 0), [data])
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, 1), [data])
+        data = np.random.rand(2, 3, 4).astype(theano.config.floatX)
+        utt.verify_grad(lambda x: sort(x, 2), [data])
 
 
 class TensorInferShapeTester(utt.InferShapeTester):