pytorch · vishwakftw · Jul 25, 2019 · Jul 27, 2019 · Jul 28, 2019 · Jul 29, 2019
diff --git a/aten/src/ATen/core/Tensor.h b/aten/src/ATen/core/Tensor.h
@@ -731,7 +731,7 @@ class CAFFE2_API Tensor {
   Tensor gather(int64_t dim, const Tensor & index, bool sparse_grad=false) const;
   Tensor addcmul(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
   Tensor addcdiv(const Tensor & tensor1, const Tensor & tensor2, Scalar value=1) const;
-  std::tuple<Tensor,Tensor> gels(const Tensor & A) const;
+  std::tuple<Tensor,Tensor> lstsq(const Tensor & A) const;
   std::tuple<Tensor,Tensor> triangular_solve(const Tensor & A, bool upper=true, bool transpose=false, bool unitriangular=false) const;
   std::tuple<Tensor,Tensor> symeig(bool eigenvectors=false, bool upper=true) const;
   std::tuple<Tensor,Tensor> eig(bool eigenvectors=false) const;

diff --git a/aten/src/ATen/core/TensorMethods.h b/aten/src/ATen/core/TensorMethods.h
@@ -1545,8 +1545,8 @@ inline Tensor Tensor::addcdiv(const Tensor & tensor1, const Tensor & tensor2, Sc
     static auto table = globalATenDispatch().getOpTable("aten::addcdiv(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor");
     return table->getOp<Tensor (const Tensor &, const Tensor &, const Tensor &, Scalar)>(tensorTypeIdToBackend(type_id()), is_variable())(*this, tensor1, tensor2, value);
 }
-inline std::tuple<Tensor,Tensor> Tensor::gels(const Tensor & A) const {
-    static auto table = globalATenDispatch().getOpTable("aten::gels(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)");
+inline std::tuple<Tensor,Tensor> Tensor::lstsq(const Tensor & A) const {
+    static auto table = globalATenDispatch().getOpTable("aten::lstsq(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)");
     return table->getOp<std::tuple<Tensor,Tensor> (const Tensor &, const Tensor &)>(tensorTypeIdToBackend(type_id()), is_variable())(*this, A);
 }
 inline std::tuple<Tensor,Tensor> Tensor::triangular_solve(const Tensor & A, bool upper, bool transpose, bool unitriangular) const {

diff --git a/aten/src/ATen/core/aten_interned_strings.h b/aten/src/ATen/core/aten_interned_strings.h
@@ -334,7 +334,6 @@ _(aten, full) \
 _(aten, full_like) \
 _(aten, gather) \
 _(aten, ge) \
-_(aten, gels) \
 _(aten, geometric) \
 _(aten, geqrf) \
 _(aten, ger) \
@@ -415,6 +414,7 @@ _(aten, logspace) \
 _(aten, logsumexp) \
 _(aten, lstm) \
 _(aten, lstm_cell) \
+_(aten, lstsq) \
 _(aten, lt) \
 _(aten, lu_solve) \
 _(aten, margin_ranking_loss) \

diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
@@ -3747,12 +3747,12 @@
     CPU: legacy::cpu::_th_addcdiv
     CUDA: legacy::cuda::_th_addcdiv
 
-- func: gels(Tensor self, Tensor A, *, Tensor(a!) X, Tensor(b!) qr) -> (Tensor(a!) solution, Tensor(b!) QR)
+- func: lstsq(Tensor self, Tensor A, *, Tensor(a!) X, Tensor(b!) qr) -> (Tensor(a!) solution, Tensor(b!) QR)
   dispatch:
     CPU: legacy::cpu::_th_gels_out
     CUDA: legacy::cuda::_th_gels_out
 
-- func: gels(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)
+- func: lstsq(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)
   variants: method, function
   dispatch:
     CPU: legacy::cpu::_th_gels

diff --git a/docs/source/tensors.rst b/docs/source/tensors.rst
@@ -317,6 +317,7 @@ view of a storage and defines numeric operations on it.
    .. automethod:: log_normal_
    .. automethod:: logsumexp
    .. automethod:: long
+   .. automethod:: lstsq
    .. automethod:: lt
    .. automethod:: lt_
    .. automethod:: lu

diff --git a/docs/source/torch.rst b/docs/source/torch.rst
@@ -334,6 +334,7 @@ BLAS and LAPACK Operations
 .. autofunction:: det
 .. autofunction:: logdet
 .. autofunction:: slogdet
+.. autofunction:: lstsq
 .. autofunction:: lu
 .. autofunction:: lu_solve
 .. autofunction:: lu_unpack

diff --git a/test/test_cuda.py b/test/test_cuda.py
@@ -2757,6 +2757,10 @@ def test_triangular_solve_batched_many_batches(self):
     def test_triangular_solve_batched_dims(self):
         _TestTorchMixin._test_triangular_solve_batched_dims(self, lambda t: t.cuda())
 
+    @unittest.skipIf(not TEST_MAGMA, "no MAGMA library detected")
+    def test_lstsq(self):
+        _TestTorchMixin._test_lstsq(self, 'cuda')
+
     @unittest.skipIf(not TEST_MAGMA, "no MAGMA library detected")
     def test_qr(self):
         _TestTorchMixin._test_qr(self, lambda t: t.cuda())

diff --git a/test/test_namedtuple_return_api.py b/test/test_namedtuple_return_api.py
@@ -11,7 +11,7 @@
 aten_native_yaml = os.path.join(path, '../aten/src/ATen/native/native_functions.yaml')
 all_operators_with_namedtuple_return = {
     'max', 'min', 'median', 'mode', 'kthvalue', 'svd', 'symeig', 'eig',
-    'qr', 'geqrf', 'solve', 'slogdet', 'sort', 'topk', 'gels',
+    'qr', 'geqrf', 'solve', 'slogdet', 'sort', 'topk', 'lstsq',
     'triangular_solve'
 }
 
@@ -61,7 +61,7 @@ def test_namedtuple_return(self):
             op(operators=['geqrf'], input=(), names=('a', 'tau'), hasout=True),
             op(operators=['symeig', 'eig'], input=(True,), names=('eigenvalues', 'eigenvectors'), hasout=True),
             op(operators=['triangular_solve'], input=(a,), names=('solution', 'cloned_coefficient'), hasout=True),
-            op(operators=['gels'], input=(a,), names=('solution', 'QR'), hasout=True),
+            op(operators=['lstsq'], input=(a,), names=('solution', 'QR'), hasout=True),
         ]
 
         for op in operators:

diff --git a/test/test_torch.py b/test/test_torch.py
@@ -5771,28 +5771,35 @@ def run_test(A_dims, b_dims, cast, upper, transpose, unitriangular):
     def test_triangular_solve_batched_dims(self):
         self._test_triangular_solve_batched_dims(self, lambda t: t)
 
-    @skipIfNoLapack
-    def test_gels(self):
+    @staticmethod
+    def _test_lstsq(self, device):
+        def cast_fn(tensor):
+            return tensor.to(device=device)
+
         def _test_underdetermined(a, b, expectedNorm):
+            # underdetermined systems are not supported on the GPU
+            if 'cuda' in device:
+                return
+
             m = a.size()[0]
             n = a.size()[1]
             assert(m <= n)
 
             a_copy = a.clone()
             b_copy = b.clone()
-            res1 = torch.gels(b, a)[0]
+            res1 = torch.lstsq(b, a)[0]
             self.assertEqual(a, a_copy, 0)
             self.assertEqual(b, b_copy, 0)
             self.assertEqual((torch.mm(a, res1) - b).norm(), expectedNorm, 1e-8)
 
-            ta = torch.Tensor()
-            tb = torch.Tensor()
-            res2 = torch.gels(b, a, out=(tb, ta))[0]
+            ta = cast_fn(torch.Tensor())
+            tb = cast_fn(torch.Tensor())
+            res2 = torch.lstsq(b, a, out=(tb, ta))[0]
             self.assertEqual(a, a_copy, 0)
             self.assertEqual(b, b_copy, 0)
             self.assertEqual((torch.mm(a, res1) - b).norm(), expectedNorm, 1e-8)
 
-            res3 = torch.gels(b, a, out=(b, a))[0]
+            res3 = torch.lstsq(b, a, out=(b, a))[0]
             self.assertEqual((torch.mm(a_copy, b) - b_copy).norm(), expectedNorm, 1e-8)
             self.assertEqual(res1, tb, 0)
             self.assertEqual(res1, b, 0)
@@ -5806,7 +5813,6 @@ def _test_overdetermined(a, b, expectedNorm):
 
             def check_norm(a, b, expected_norm, gels_result):
                 # Checks |ax - b| and the residual info from the result
-                n = a.size()[1]
 
                 # The first n rows is the least square solution.
                 # Rows n to m-1 contain residual information.
@@ -5819,19 +5825,19 @@ def check_norm(a, b, expected_norm, gels_result):
 
             a_copy = a.clone()
             b_copy = b.clone()
-            res1 = torch.gels(b, a)[0]
+            res1 = torch.lstsq(b, a)[0]
             self.assertEqual(a, a_copy, 0)
             self.assertEqual(b, b_copy, 0)
             check_norm(a, b, expectedNorm, res1)
 
-            ta = torch.Tensor()
-            tb = torch.Tensor()
-            res2 = torch.gels(b, a, out=(tb, ta))[0]
+            ta = cast_fn(torch.Tensor())
+            tb = cast_fn(torch.Tensor())
+            res2 = torch.lstsq(b, a, out=(tb, ta))[0]
             self.assertEqual(a, a_copy, 0)
             self.assertEqual(b, b_copy, 0)
             check_norm(a, b, expectedNorm, res2)
 
-            res3 = torch.gels(b, a, out=(b, a))[0]
+            res3 = torch.lstsq(b, a, out=(b, a))[0]
             check_norm(a_copy, b_copy, expectedNorm, res3)
 
             self.assertEqual(res1, tb, 0)
@@ -5841,51 +5847,55 @@ def check_norm(a, b, expected_norm, gels_result):
 
         # basic test
         expectedNorm = 0
-        a = torch.Tensor(((1.44, -9.96, -7.55, 8.34),
-                          (-7.84, -0.28, 3.24, 8.09),
-                          (-4.39, -3.24, 6.27, 5.28),
-                          (4.53, 3.83, -6.64, 2.06))).t()
-        b = torch.Tensor(((8.58, 8.26, 8.48, -5.28),
-                          (9.35, -4.43, -0.70, -0.26))).t()
+        a = cast_fn(torch.Tensor(((1.44, -9.96, -7.55, 8.34),
+                                  (-7.84, -0.28, 3.24, 8.09),
+                                  (-4.39, -3.24, 6.27, 5.28),
+                                  (4.53, 3.83, -6.64, 2.06)))).t()
+        b = cast_fn(torch.Tensor(((8.58, 8.26, 8.48, -5.28),
+                                  (9.35, -4.43, -0.70, -0.26)))).t()
         _test_underdetermined(a, b, expectedNorm)
 
-        # test overderemined
+        # test overdetermined
         expectedNorm = 17.390200628863
-        a = torch.Tensor(((1.44, -9.96, -7.55, 8.34, 7.08, -5.45),
-                          (-7.84, -0.28, 3.24, 8.09, 2.52, -5.70),
-                          (-4.39, -3.24, 6.27, 5.28, 0.74, -1.19),
-                          (4.53, 3.83, -6.64, 2.06, -2.47, 4.70))).t()
-        b = torch.Tensor(((8.58, 8.26, 8.48, -5.28, 5.72, 8.93),
-                          (9.35, -4.43, -0.70, -0.26, -7.36, -2.52))).t()
+        a = cast_fn(torch.Tensor(((1.44, -9.96, -7.55, 8.34, 7.08, -5.45),
+                                  (-7.84, -0.28, 3.24, 8.09, 2.52, -5.70),
+                                  (-4.39, -3.24, 6.27, 5.28, 0.74, -1.19),
+                                  (4.53, 3.83, -6.64, 2.06, -2.47, 4.70)))).t()
+        b = cast_fn(torch.Tensor(((8.58, 8.26, 8.48, -5.28, 5.72, 8.93),
+                                  (9.35, -4.43, -0.70, -0.26, -7.36, -2.52)))).t()
         _test_overdetermined(a, b, expectedNorm)
 
         # test underdetermined
         expectedNorm = 0
-        a = torch.Tensor(((1.44, -9.96, -7.55),
-                          (-7.84, -0.28, 3.24),
-                          (-4.39, -3.24, 6.27),
-                          (4.53, 3.83, -6.64))).t()
-        b = torch.Tensor(((8.58, 8.26, 8.48),
-                          (9.35, -4.43, -0.70))).t()
+        a = cast_fn(torch.Tensor(((1.44, -9.96, -7.55),
+                                  (-7.84, -0.28, 3.24),
+                                  (-4.39, -3.24, 6.27),
+                                  (4.53, 3.83, -6.64)))).t()
+        b = cast_fn(torch.Tensor(((8.58, 8.26, 8.48),
+                                  (9.35, -4.43, -0.70)))).t()
         _test_underdetermined(a, b, expectedNorm)
 
         # test reuse
         expectedNorm = 0
-        a = torch.Tensor(((1.44, -9.96, -7.55, 8.34),
-                          (-7.84, -0.28, 3.24, 8.09),
-                          (-4.39, -3.24, 6.27, 5.28),
-                          (4.53, 3.83, -6.64, 2.06))).t()
-        b = torch.Tensor(((8.58, 8.26, 8.48, -5.28),
-                          (9.35, -4.43, -0.70, -0.26))).t()
-        ta = torch.Tensor()
-        tb = torch.Tensor()
-        torch.gels(b, a, out=(tb, ta))
+        a = cast_fn(torch.Tensor(((1.44, -9.96, -7.55, 8.34),
+                                  (-7.84, -0.28, 3.24, 8.09),
+                                  (-4.39, -3.24, 6.27, 5.28),
+                                  (4.53, 3.83, -6.64, 2.06)))).t()
+        b = cast_fn(torch.Tensor(((8.58, 8.26, 8.48, -5.28),
+                                  (9.35, -4.43, -0.70, -0.26)))).t()
+        ta = cast_fn(torch.Tensor())
+        tb = cast_fn(torch.Tensor())
+        torch.lstsq(b, a, out=(tb, ta))
         self.assertEqual((torch.mm(a, tb) - b).norm(), expectedNorm, 1e-8)
-        torch.gels(b, a, out=(tb, ta))
+        torch.lstsq(b, a, out=(tb, ta))
         self.assertEqual((torch.mm(a, tb) - b).norm(), expectedNorm, 1e-8)
-        torch.gels(b, a, out=(tb, ta))
+        torch.lstsq(b, a, out=(tb, ta))
         self.assertEqual((torch.mm(a, tb) - b).norm(), expectedNorm, 1e-8)
 
+    @skipIfNoLapack
+    def test_lstsq(self):
+        self._test_lstsq(self, 'cpu')
+
     @skipIfNoLapack
     def test_eig(self):
         a = torch.Tensor(((1.96, 0.00, 0.00, 0.00, 0.00),
@@ -8906,9 +8916,9 @@ def fn(torchfn, *args):
             q, r = fn(torch.qr, (3, 0), False)
             self.assertEqual([(3, 3), (3, 0)], [q.shape, r.shape])
 
-            # gels
-            self.assertRaises(RuntimeError, lambda: torch.gels(torch.randn(0, 0), torch.randn(0, 0)))
-            self.assertRaises(RuntimeError, lambda: torch.gels(torch.randn(0,), torch.randn(0, 0)))
+            # lstsq
+            self.assertRaises(RuntimeError, lambda: torch.lstsq(torch.randn(0, 0), torch.randn(0, 0)))
+            self.assertRaises(RuntimeError, lambda: torch.lstsq(torch.randn(0,), torch.randn(0, 0)))
 
     def test_expand(self):
         tensor = torch.rand(1, 8, 1)

diff --git a/tools/autograd/derivatives.yaml b/tools/autograd/derivatives.yaml
@@ -345,10 +345,6 @@
   self: zeros_like(self)
   other: zeros_like(other)
 
-- name: gels(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)
-  self: not_implemented("gels")
-  A: not_implemented("gels")
-
 - name: geometric_(Tensor(a!) self, float p, *, Generator? generator=None) -> Tensor(a!)
   self: zeros_like(grad)
 
@@ -467,6 +463,10 @@
 - name: logsumexp(Tensor self, int[1] dim, bool keepdim=False) -> Tensor
   self: logsumexp_backward(grad, self, result, dim, keepdim)
 
+- name: lstsq(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)
+  self: not_implemented("lstsq")
+  A: not_implemented("lstsq")
+
 - name: lt_(Tensor(a!) self, Scalar other) -> Tensor(a!)
   self: zeros_like(self)
 

diff --git a/torch/_tensor_docs.py b/torch/_tensor_docs.py
@@ -1040,13 +1040,6 @@ def add_docstr_all(method, docstr):
 In-place version of :meth:`~Tensor.ge`
 """)
 
-add_docstr_all('gels',
-               r"""
-gels(A) -> Tensor
-
-See :func:`torch.gels`
-""")
-
 add_docstr_all('geometric_',
                r"""
 geometric_(p, *, generator=None) -> Tensor
@@ -1466,6 +1459,13 @@ def add_docstr_all(method, docstr):
 See :func:`torch.logsumexp`
 """)
 
+add_docstr_all('lstsq',
+               r"""
+lstsq(A) -> (Tensor, Tensor)
+
+See :func:`torch.lstsq`
+""")
+
 add_docstr_all('lt',
                r"""
 lt(other) -> Tensor