Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Zero-Dim] fix create_scalar to create 0D #51024

Merged
merged 1 commit into from
Mar 3, 2023
Merged

[Zero-Dim] fix create_scalar to create 0D #51024

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions paddle/fluid/pybind/eager_math_op_patch.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -539,8 +539,8 @@ static PyObject* tensor__mul__method(TensorObject* self,
CastPyArg2Scalar(other_obj, "__mul__", 0);
if (PyComplex_Check(other_obj)) {
eager_gil_scoped_release guard;
other_tensor =
full_ad_func({1}, value, DataType::COMPLEX64, self_tensor.place());
other_tensor = full_ad_func(
self_tensor.shape(), value, DataType::COMPLEX64, self_tensor.place());
} else {
eager_gil_scoped_release guard;
other_tensor = full_ad_func(
Expand Down
3 changes: 1 addition & 2 deletions python/paddle/fluid/layers/math_op_patch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,8 +98,7 @@ def create_tensor(block, value, dtype, shape):
return var

def create_scalar(block, value, dtype):
# TODO(zhouwei): will change to [] which is 0-D Tensor
return create_tensor(block, value, dtype, shape=[1])
return create_tensor(block, value, dtype, shape=[])

def create_tensor_with_batchsize(ref_var, value, dtype):
assert isinstance(ref_var, Variable)
Expand Down
178 changes: 93 additions & 85 deletions python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,12 +139,7 @@ def test_static_unary(self):
x = paddle.rand([])
x.stop_gradient = False
out = api(x)
# TODO(zhouwei):
# ScaleLossGradOp / append_backward set grad shape to [1]
# after output 0D, may change it to []
# use out.sum() to avoid this two problem now
loss = out.sum()
paddle.static.append_backward(loss)
paddle.static.append_backward(out)

fetch_list = [x, out]
if block.has_var(x.grad_name):
Expand All @@ -156,12 +151,10 @@ def test_static_unary(self):
self.assertEqual(item.shape, ())

# 2) Test CompiledProgram Program
expect_shape = ()
compile_prog = paddle.static.CompiledProgram(main_prog)

res = exe.run(compile_prog, fetch_list=fetch_list)
for item in res:
self.assertEqual(item.shape, expect_shape)
self.assertEqual(item.shape, ())

paddle.disable_static()

Expand Down Expand Up @@ -229,7 +222,7 @@ def test_static_reduce(self):
x = paddle.rand([])
x.stop_gradient = False
out = api(x, None)
paddle.static.append_backward(out.sum())
paddle.static.append_backward(out)

out_empty_list = api(x, None)
self.assertEqual(out_empty_list.shape, ())
Expand Down Expand Up @@ -437,7 +430,7 @@ def test_static_binary(self):
self.assertEqual(out.shape, out_cls.shape)
else:
out = api(x, y)
paddle.static.append_backward(out.sum())
paddle.static.append_backward(out)

self.assertEqual(x.shape, ())
self.assertEqual(y.shape, ())
Expand All @@ -464,7 +457,7 @@ def test_static_binary(self):
self.assertEqual(out.shape, out_cls.shape)
else:
out = api(x, y)
paddle.static.append_backward(out.sum())
paddle.static.append_backward(out)

self.assertEqual(x.shape, ())
self.assertEqual(y.shape, (2, 3, 4))
Expand All @@ -491,7 +484,7 @@ def test_static_binary(self):
self.assertEqual(out.shape, out_cls.shape)
else:
out = api(x, y)
paddle.static.append_backward(out.sum())
paddle.static.append_backward(out)

self.assertEqual(x.shape, (2, 3, 4))
self.assertEqual(y.shape, ())
Expand All @@ -505,9 +498,6 @@ def test_static_binary(self):
self.assertEqual(y_grad.shape, ())
self.assertEqual(out_grad.shape, (2, 3, 4))

# TODO(zhouwei25):
# will open this UT after fix create_scalar in static graph
'''
# 4) x is 0D , y is scalar
x = paddle.rand([])
x.stop_gradient = False
Expand All @@ -516,17 +506,16 @@ def test_static_binary(self):
out = getattr(paddle.static.Variable, api['cls_method'])(
x, y
)
paddle.static.append_backward(out.sum())
paddle.static.append_backward(out)

self.assertEqual(x.shape, ())
self.assertEqual(out.shape, ())
if block.has_var(x.name):
if block.has_var(x.grad_name):
out_grad = block.var(out.grad_name)
x_grad = block.var(x.grad_name)

self.assertEqual(out_grad.shape, ())
self.assertEqual(x_grad.shape, ())
'''

for api in binary_int_api_list:
main_prog = paddle.static.Program()
Expand Down Expand Up @@ -2154,10 +2143,10 @@ def test_median(self):
@prog_scope()
def test_std(self):
x = paddle.rand([])
x.stop_gradient = False
out1 = paddle.std(x)
out2 = paddle.std(x, [])
paddle.static.append_backward(out1)
paddle.static.append_backward(out2)

prog = paddle.static.default_main_program()
res = self.exe.run(
Expand All @@ -2166,19 +2155,23 @@ def test_std(self):
x,
out1,
out2,
x.grad_name,
out1.grad_name,
],
)
self.assertEqual(res[0].shape, ())
self.assertEqual(res[1].shape, ())
self.assertEqual(res[2].shape, ())
self.assertEqual(res[3].shape, ())
self.assertEqual(res[4].shape, ())

@prog_scope()
def test_var(self):
x = paddle.rand([])
x.stop_gradient = False
out1 = paddle.var(x)
out2 = paddle.var(x, [])
paddle.static.append_backward(out1)
paddle.static.append_backward(out2)

prog = paddle.static.default_main_program()
res = self.exe.run(
Expand All @@ -2187,11 +2180,15 @@ def test_var(self):
x,
out1,
out2,
x.grad_name,
out1.grad_name,
],
)
self.assertEqual(res[0].shape, ())
self.assertEqual(res[1].shape, ())
self.assertEqual(res[2].shape, ())
self.assertEqual(res[3].shape, ())
self.assertEqual(res[4].shape, ())

@prog_scope()
def test_quantile(self):
Expand Down Expand Up @@ -3651,7 +3648,7 @@ class TestUnaryElementwiseAPIWithComplexInput(unittest.TestCase):
def test_dygraph_unary(self):
paddle.disable_static()
for api in unary_apis_with_complex_input:
x = paddle.to_tensor(2.0 + 3.0j).squeeze()
x = paddle.rand([]) + 1j * paddle.rand([])
x.stop_gradient = False
x.retain_grads()
out = api(x)
Expand All @@ -3668,26 +3665,17 @@ def test_dygraph_unary(self):

def test_static_unary(self):
paddle.enable_static()

for api in unary_apis_with_complex_input:
main_prog = paddle.static.Program()
block = main_prog.global_block()
exe = paddle.static.Executor()
with paddle.static.program_guard(
main_prog, paddle.static.Program()
):
# before full support for complex, we cannot create complex tensor with the same code as in dynamic graph
x = paddle.complex(
paddle.to_tensor(2.0), paddle.to_tensor(2.0)
).squeeze()
x = paddle.complex(paddle.rand([]), paddle.rand([]))
x.stop_gradient = False
out = api(x)
# TODO(zhouwei):
# ScaleLossGradOp / append_backward set grad shape to [1]
# after output 0D, may change it to []
# use out.sum() to avoid this two problem now
loss = out.sum()
paddle.static.append_backward(loss)
paddle.static.append_backward(out)

fetch_list = [x, out]
if block.has_var(x.grad_name):
Expand All @@ -3699,79 +3687,99 @@ def test_static_unary(self):
self.assertEqual(item.shape, ())

# 2) Test CompiledProgram Program
expect_shape = ()
compile_prog = paddle.static.CompiledProgram(main_prog)

res = exe.run(compile_prog, fetch_list=fetch_list)
for item in res:
self.assertEqual(item.shape, expect_shape)
self.assertEqual(item.shape, ())

paddle.disable_static()


class TestAsReal(unittest.TestCase):
def test_dygraph(self):
paddle.disable_static()
for api in unary_apis_with_complex_input:
x = paddle.to_tensor(2.0 + 3.0j).squeeze()
x.stop_gradient = False
x.retain_grads()
out = paddle.as_real(x)
out.retain_grads()
out.backward()
x = paddle.rand([]) + 1j * paddle.rand([])
x.stop_gradient = False
x.retain_grads()
out = paddle.as_real(x)
out.retain_grads()
out.backward()

self.assertEqual(x.shape, [])
self.assertEqual(out.shape, [2])
if x.grad is not None:
self.assertEqual(x.grad.shape, [])
self.assertEqual(out.grad.shape, [2])
self.assertEqual(x.shape, [])
self.assertEqual(out.shape, [2])
if x.grad is not None:
self.assertEqual(x.grad.shape, [])
self.assertEqual(out.grad.shape, [2])

paddle.enable_static()

def test_static(self):
paddle.enable_static()

for api in unary_apis_with_complex_input:
main_prog = paddle.static.Program()
block = main_prog.global_block()
exe = paddle.static.Executor()
with paddle.static.program_guard(
main_prog, paddle.static.Program()
):
# before full support for complex, we cannot create complex tensor with the same code as in dynamic graph
x = paddle.complex(
paddle.to_tensor(2.0), paddle.to_tensor(2.0)
).squeeze()
x.stop_gradient = False
out = paddle.as_real(x)
self.assertEqual(x.shape, ())
self.assertEqual(out.shape, (2,))
# TODO(zhouwei):
# ScaleLossGradOp / append_backward set grad shape to [1]
# after output 0D, may change it to []
# use out.sum() to avoid this two problem now
loss = out.abs().sum()
paddle.static.append_backward(loss)
main_prog = paddle.static.Program()
block = main_prog.global_block()
exe = paddle.static.Executor()
with paddle.static.program_guard(main_prog, paddle.static.Program()):
x = paddle.complex(paddle.rand([]), paddle.rand([]))
x.stop_gradient = False
out = paddle.as_real(x)
self.assertEqual(x.shape, ())
self.assertEqual(out.shape, (2,))
paddle.static.append_backward(out.sum())

fetch_list = [x, out]
if block.has_var(x.grad_name):
fetch_list.extend([x.grad_name, out.grad_name])
fetch_list = [x, out]
if block.has_var(x.grad_name):
fetch_list.extend([x.grad_name, out.grad_name])

# 1) Test Program
res = exe.run(main_prog, fetch_list=fetch_list)
self.assertEqual(res[0].shape, ())
self.assertEqual(res[1].shape, (2,))
self.assertEqual(res[2].shape, ())
self.assertEqual(res[3].shape, (2,))
res = exe.run(main_prog, fetch_list=fetch_list)
self.assertEqual(res[0].shape, ())
self.assertEqual(res[1].shape, (2,))
self.assertEqual(res[2].shape, ())
self.assertEqual(res[3].shape, (2,))

# 2) Test CompiledProgram Program
expect_shapes = (), (2,), (), (2,)
compile_prog = paddle.static.CompiledProgram(main_prog)
paddle.disable_static()

res = exe.run(compile_prog, fetch_list=fetch_list)
print(res)
for actual, expect in zip(res, expect_shapes):
self.assertEqual(actual.shape, expect)

class TestAsComplex(unittest.TestCase):
def test_dygraph(self):
paddle.disable_static()
x = paddle.rand([2])
x.stop_gradient = False
x.retain_grads()
out = paddle.as_complex(x)
out.retain_grads()
out.backward()

self.assertEqual(x.shape, [2])
self.assertEqual(out.shape, [])
if x.grad is not None:
self.assertEqual(x.grad.shape, [2])
self.assertEqual(out.grad.shape, [])

paddle.enable_static()

def test_static(self):
paddle.enable_static()
main_prog = paddle.static.Program()
block = main_prog.global_block()
exe = paddle.static.Executor()
with paddle.static.program_guard(main_prog, paddle.static.Program()):
x = paddle.rand([2])
x.stop_gradient = False
out = paddle.as_complex(x)
self.assertEqual(x.shape, (2,))
self.assertEqual(out.shape, ())
paddle.static.append_backward(out.sum())

fetch_list = [x, out]
if block.has_var(x.grad_name):
fetch_list.extend([x.grad_name, out.grad_name])

res = exe.run(main_prog, fetch_list=fetch_list)
self.assertEqual(res[0].shape, (2,))
self.assertEqual(res[1].shape, ())
self.assertEqual(res[2].shape, (2,))
self.assertEqual(res[3].shape, ())

paddle.disable_static()

Expand Down
8 changes: 2 additions & 6 deletions python/paddle/tensor/math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4449,19 +4449,15 @@ def gcd(x, y, name=None):
y = paddle.broadcast_to(y, shape)
x = paddle.abs(x)
y = paddle.abs(y)
# TODO(zhouwei25): Support 0D for not_equal tensor with scalar
zero = paddle.full([], 0)

def _gcd_cond_fn(x, y):
# return paddle.any(y != 0)
return paddle.any(y != zero)
return paddle.any(y != 0)

def _gcd_body_fn(x, y):
# paddle.mod will raise an error when any element of y is 0. To avoid
# that, we change those zeros to ones. Their values don't matter because
# they won't be used.
# y_not_equal_0 = y != 0
y_not_equal_0 = y != zero
y_not_equal_0 = y != 0
y_safe = paddle.where(y_not_equal_0, y, paddle.ones(y.shape, y.dtype))
x, y = (
paddle.where(y_not_equal_0, y, x),
Expand Down