add test cases to ensure view_as_complex and view_as_real reuses the input's allocation

Author: chenfeiyu

python/paddle/fluid/tests/unittests/test_view_op_reuse_allocation.py

@@ -114,5 +114,164 @@ def view_api_processing(self, var):
  return paddle.flatten(var)
 
 
+# NOTE(chenfeiyu): TestCases for operators with complex view mechanism.
+# a kind of view that shares storage between complex and real tensors
+# though the storage retains, the shape and data type of the two tensors differ
+class TestDygraphViewAsComplexReuseAllocation(unittest.TestCase):
+ def setUp(self):
+ self.init_shape()
+ self.init_dtype()
+
+ def init_shape(self):
+ self.input_shape = [2, 3, 2]
+ self.output_shape = [2, 3]
+
+ def init_dtype(self):
+ self.input_dtype = paddle.float32
+ self.output_dtype = paddle.complex64
+
+ def view_api_processing(self, var):
+ return paddle.view_as_complex(var)
+
+ def test_view_api(self):
+ var = paddle.rand(self.input_shape, dtype=self.input_dtype)
+ view_var = self.view_api_processing(var)
+
+ self.assertEqual(var.shape, self.input_shape)
+ self.assertEqual(view_var.shape, self.output_shape)
+
+ view_var.add_(paddle.to_tensor([1.0]))
+ var_numpy = var.numpy()
+ view_var_numpy = view_var.numpy()
+ self.assertTrue(
+ np.array_equal(
+ np.take(
+ var_numpy, 0, axis=-1), view_var_numpy.real))
+ self.assertTrue(
+ np.array_equal(
+ np.take(
+ var_numpy, 1, axis=-1), view_var_numpy.imag))
+
+ def test_forward_version(self):
+ var = paddle.rand(self.input_shape, dtype=self.input_dtype)
+ self.assertEqual(var.inplace_version, 0)
+ view_var = self.view_api_processing(var)
+ self.assertEqual(view_var.inplace_version, 0)
+
+ view_var.add_(paddle.to_tensor([1.0]))
+ self.assertEqual(var.inplace_version, 1)
+ self.assertEqual(view_var.inplace_version, 1)
+
+ view_var_2 = self.view_api_processing(var)
+ self.assertEqual(view_var_2.inplace_version, 1)
+
+ var.add_(paddle.to_tensor([1.0]))
+ self.assertEqual(view_var.inplace_version, 2)
+ self.assertEqual(view_var_2.inplace_version, 2)
+
+ def test_backward_error(self):
+ # It raises an error because the inplace operator will result
+ # in incorrect gradient computation.
+ with paddle.fluid.dygraph.guard():
+ var_a = paddle.ones(shape=self.input_shape, dtype=self.input_dtype)
+ var_a.stop_gradient = False
+
+ var_b = var_a**2
+
+ # Here, the gradient computation will use the value of var_b
+ var_c = var_b**2
+ view_var_b = self.view_api_processing(var_b)
+ view_var_b.add_(paddle.to_tensor(
+ [1.0])) # var_b is modified inplace
+
+ loss = paddle.nn.functional.relu(var_c)
+ with self.assertRaisesRegexp(
+ RuntimeError,
+ "received tensor_version:{} != wrapper_version_snapshot:{}".
+ format(1, 0)):
+ loss.backward()
+
+
+class TestDygraphViewAsRealReuseAllocation(unittest.TestCase):
+ def setUp(self):
+ self.init_shape()
+ self.init_dtype()
+
+ def init_shape(self):
+ self.input_shape = [2, 3]
+ self.output_shape = [2, 3, 2]
+
+ def init_dtype(self):
+ self.input_dtype = paddle.complex64
+ self.output_dtype = paddle.float32
+
+ def view_api_processing(self, var):
+ return paddle.view_as_real(var)
+
+ def test_view_api(self):
+ var = (
+ paddle.rand(self.input_shape) + 1j * paddle.rand(self.input_shape)
+ ).astype(self.input_dtype)
+ view_var = self.view_api_processing(var)
+
+ self.assertEqual(var.shape, self.input_shape)
+ self.assertEqual(view_var.shape, self.output_shape)
+
+ view_var[0, 0, 0] = 2.0
+ var_numpy = var.numpy()
+ view_var_numpy = view_var.numpy()
+ self.assertTrue(
+ np.array_equal(
+ np.take(
+ view_var_numpy, 0, axis=-1), var_numpy.real))
+ self.assertTrue(
+ np.array_equal(
+ np.take(
+ view_var_numpy, 1, axis=-1), var_numpy.imag))
+
+ def test_forward_version(self):
+ var = (
+ paddle.rand(self.input_shape) + 1j * paddle.rand(self.input_shape)
+ ).astype(self.input_dtype)
+ self.assertEqual(var.inplace_version, 0)
+
+ view_var = self.view_api_processing(var)
+ self.assertEqual(view_var.inplace_version, 0)
+
+ view_var[0, 0, 0] = 2.0
+ self.assertEqual(var.inplace_version, 1)
+ self.assertEqual(view_var.inplace_version, 1)
+
+ view_var_2 = self.view_api_processing(var)
+ self.assertEqual(view_var_2.inplace_version, 1)
+
+ var.add_(paddle.to_tensor([1.0]))
+ self.assertEqual(view_var.inplace_version, 2)
+ self.assertEqual(view_var_2.inplace_version, 2)
+
+ def test_backward_error(self):
+ # It raises an error because the inplace operator will result
+ # in incorrect gradient computation.
+ with paddle.fluid.dygraph.guard():
+ var_a = (paddle.ones(shape=self.input_shape) + 1j * paddle.ones(
+ shape=self.input_shape)).astype(self.input_dtype)
+ var_a.stop_gradient = False
+
+ var_b = paddle.conj(var_a)
+
+ # Here, the gradient computation will use the value of var_b
+ var_c = var_b * var_b
+ view_var_b = self.view_api_processing(var_b)
+ view_var_b.add_(paddle.to_tensor(
+ [1.0])) # var_b is modified inplace
+
+ loss = paddle.abs(var_c).sum()
+ with self.assertRaisesRegexp(
+ RuntimeError,
+ "received tensor_version:{} != wrapper_version_snapshot:{}".
+ format(1, 0)):
+ loss.backward()
+
+
 if __name__ == "__main__":
  unittest.main()