wehub-resource-sync
381 lines
14 KiB
Python
381 lines
14 KiB
Python
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
import unittest
|
|
|
|
import numpy as np
|
|
from op_test import get_device_place
|
|
|
|
import paddle
|
|
from paddle.base import core
|
|
|
|
|
|
class TestPaddleAddZeroSize(unittest.TestCase):
|
|
def setUp(self):
|
|
self.place = get_device_place()
|
|
self.shape = [0, 3]
|
|
self.dtype_pairs = [(paddle.float32, paddle.float32)]
|
|
if core.is_float16_supported(self.place):
|
|
self.dtype_pairs.append((paddle.float32, paddle.float16))
|
|
if core.is_bfloat16_supported(self.place):
|
|
self.dtype_pairs.append((paddle.float32, paddle.bfloat16))
|
|
|
|
def test_0size(self):
|
|
for x_dtype, y_dtype in self.dtype_pairs:
|
|
with self.subTest(msg=f"{x_dtype} + {y_dtype}"):
|
|
x = paddle.randn(self.shape, dtype=x_dtype)
|
|
y = paddle.randn(self.shape, dtype=y_dtype)
|
|
x.stop_gradient = False
|
|
y.stop_gradient = False
|
|
|
|
out = paddle.add(x, y)
|
|
out.backward()
|
|
|
|
self.assertEqual(out.shape, self.shape)
|
|
self.assertEqual(out.dtype, x_dtype)
|
|
self.assertEqual(x.grad.dtype, x_dtype)
|
|
self.assertEqual(y.grad.dtype, y_dtype)
|
|
|
|
|
|
class TestPaddleAddBackward(unittest.TestCase):
|
|
def setUp(self):
|
|
self.place = get_device_place()
|
|
self.x_np_f32 = np.array(
|
|
[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=np.float32
|
|
)
|
|
self.y_np_f32 = np.array(
|
|
[[7.0, 8.0, 9.0], [10.0, 11.0, 12.0]], dtype=np.float32
|
|
)
|
|
self.N = self.x_np_f32.size
|
|
self.expected_grad = np.full(
|
|
self.x_np_f32.shape, 1.0 / self.N, dtype=np.float32
|
|
)
|
|
|
|
def test_backward(self):
|
|
x = paddle.to_tensor(self.x_np_f32, stop_gradient=False)
|
|
y = paddle.to_tensor(self.y_np_f32, stop_gradient=False)
|
|
|
|
out = paddle.add(x, y)
|
|
out.mean().backward()
|
|
|
|
np.testing.assert_allclose(
|
|
x.grad.numpy(), self.expected_grad, rtol=1e-6
|
|
)
|
|
np.testing.assert_allclose(
|
|
y.grad.numpy(), self.expected_grad, rtol=1e-6
|
|
)
|
|
|
|
def test_backward_broadcast(self):
|
|
x_np = self.x_np_f32
|
|
y_np = np.array([[0.1, 0.2, 0.3]], dtype=np.float32)
|
|
x = paddle.to_tensor(x_np, stop_gradient=False)
|
|
y = paddle.to_tensor(y_np, stop_gradient=False)
|
|
|
|
out = paddle.add(x, y)
|
|
loss = out.mean()
|
|
loss.backward()
|
|
|
|
N = out.numel()
|
|
expected_x_grad = np.full(x_np.shape, 1.0 / N, dtype=np.float32)
|
|
expected_y_grad = np.full(y_np.shape, 2.0 / N, dtype=np.float32)
|
|
|
|
np.testing.assert_allclose(x.grad.numpy(), expected_x_grad, rtol=1e-6)
|
|
np.testing.assert_allclose(y.grad.numpy(), expected_y_grad, rtol=1e-6)
|
|
|
|
@unittest.skipUnless(
|
|
core.is_float16_supported(get_device_place()), "Skip float16 test"
|
|
)
|
|
def test_backward_mixed_precision_f16(self):
|
|
# X: float32, Y: float16
|
|
x_np = self.x_np_f32
|
|
y_np = self.y_np_f32.astype(np.float16)
|
|
x = paddle.to_tensor(x_np, stop_gradient=False)
|
|
y = paddle.to_tensor(y_np, stop_gradient=False)
|
|
|
|
out = paddle.add(x, y)
|
|
out.mean().backward()
|
|
|
|
N = out.numel()
|
|
expected_x_grad = np.full(x_np.shape, 1.0 / N, dtype=np.float32)
|
|
expected_y_grad = np.full(y_np.shape, 1.0 / N, dtype=np.float16)
|
|
|
|
rtol, atol = 1e-3, 1e-3
|
|
actual_x_grad = x.grad.numpy()
|
|
np.testing.assert_allclose(
|
|
actual_x_grad, expected_x_grad, rtol=rtol, atol=atol
|
|
)
|
|
assert actual_x_grad.dtype == expected_x_grad.dtype, (
|
|
f"x.grad dtype mismatch: expected {expected_x_grad.dtype}, got {actual_x_grad.dtype}"
|
|
)
|
|
actual_y_grad = y.grad.numpy()
|
|
np.testing.assert_allclose(
|
|
actual_y_grad, expected_y_grad, rtol=rtol, atol=atol
|
|
)
|
|
assert actual_y_grad.dtype == expected_y_grad.dtype, (
|
|
f"y.grad dtype mismatch: expected {expected_y_grad.dtype}, got {actual_y_grad.dtype}"
|
|
)
|
|
|
|
def test_backward_with_grad(self):
|
|
paddle.disable_static()
|
|
x = paddle.to_tensor(self.x_np_f32, stop_gradient=False)
|
|
y = paddle.to_tensor(self.y_np_f32, stop_gradient=False)
|
|
out = paddle.add(x, y)
|
|
|
|
out_grad_np = np.array(
|
|
[[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32
|
|
)
|
|
out_grad = paddle.to_tensor(out_grad_np)
|
|
out.backward(grad_tensor=out_grad)
|
|
|
|
expected_grad = out_grad_np
|
|
np.testing.assert_allclose(x.grad.numpy(), expected_grad, rtol=1e-6)
|
|
np.testing.assert_allclose(y.grad.numpy(), expected_grad, rtol=1e-6)
|
|
|
|
|
|
class TestPaddleAddNewFeatures(unittest.TestCase):
|
|
def setUp(self):
|
|
self.x_np = np.array([3, 5], dtype='float32')
|
|
self.y_np = np.array([2, 3], dtype='float32')
|
|
self.scalar = 2.0
|
|
self.place = get_device_place()
|
|
|
|
def test_paddle_add_with_alpha(self):
|
|
"""test paddle.add alpha"""
|
|
x = paddle.to_tensor(self.x_np, stop_gradient=False)
|
|
y = paddle.to_tensor(self.y_np, stop_gradient=False)
|
|
out = paddle.add(x, y, alpha=2)
|
|
expected = self.x_np + self.y_np * 2
|
|
np.testing.assert_array_equal(out.numpy(), expected)
|
|
|
|
out.mean().backward()
|
|
expected_x_grad = np.array([0.5, 0.5], dtype='float32')
|
|
expected_y_grad = np.array([1.0, 1.0], dtype='float32') # alpha=2
|
|
np.testing.assert_array_equal(x.grad.numpy(), expected_x_grad)
|
|
np.testing.assert_array_equal(y.grad.numpy(), expected_y_grad)
|
|
|
|
def test_tensor_add_with_alpha(self):
|
|
"""test paddle.Tensor.add alpha"""
|
|
x = paddle.to_tensor(self.x_np, stop_gradient=False)
|
|
y = paddle.to_tensor(self.y_np, stop_gradient=False)
|
|
out = x.add(y, alpha=2)
|
|
expected = self.x_np + self.y_np * 2
|
|
np.testing.assert_array_equal(out.numpy(), expected)
|
|
|
|
out.mean().backward()
|
|
expected_x_grad = np.array([0.5, 0.5], dtype='float32')
|
|
expected_y_grad = np.array([1.0, 1.0], dtype='float32') # alpha=2
|
|
np.testing.assert_array_equal(x.grad.numpy(), expected_x_grad)
|
|
np.testing.assert_array_equal(y.grad.numpy(), expected_y_grad)
|
|
|
|
def test_tensor_add_inplace_with_alpha(self):
|
|
"""test Tensor.add_ alpha"""
|
|
x = paddle.to_tensor(self.x_np)
|
|
y = paddle.to_tensor(self.y_np)
|
|
x.add_(y, alpha=2)
|
|
expected = self.x_np + self.y_np * 2
|
|
np.testing.assert_array_equal(x.numpy(), expected)
|
|
|
|
def test_consistency_between_apis(self):
|
|
"""test different APIs consistency for add with alpha"""
|
|
x = paddle.to_tensor(self.x_np)
|
|
y = paddle.to_tensor(self.y_np)
|
|
|
|
out1 = paddle.add(x, y, alpha=2)
|
|
out2 = x.add(y, alpha=2)
|
|
x.add_(y, alpha=2)
|
|
|
|
expected = self.x_np + self.y_np * 2
|
|
np.testing.assert_array_equal(out1.numpy(), expected)
|
|
np.testing.assert_array_equal(out2.numpy(), expected)
|
|
np.testing.assert_array_equal(x.numpy(), expected)
|
|
|
|
def test_static_graph_add_with_alpha(self):
|
|
"""test static graph add with alpha and parameter aliases"""
|
|
paddle.enable_static()
|
|
with paddle.static.program_guard(paddle.static.Program()):
|
|
x = paddle.static.data(name='x', shape=[-1, 2], dtype='float32')
|
|
y = paddle.static.data(name='y', shape=[-1, 2], dtype='float32')
|
|
out1 = paddle.add(x, y, alpha=2)
|
|
out2 = paddle.add(input=x, other=y, alpha=2)
|
|
|
|
exe = paddle.static.Executor(self.place)
|
|
res = exe.run(
|
|
feed={
|
|
'x': self.x_np.reshape(1, 2),
|
|
'y': self.y_np.reshape(1, 2),
|
|
},
|
|
fetch_list=[out1, out2],
|
|
)
|
|
|
|
expected = self.x_np + self.y_np * 2
|
|
for result in res:
|
|
np.testing.assert_array_equal(result.flatten(), expected)
|
|
paddle.disable_static()
|
|
|
|
def test_param_alias_input_other(self):
|
|
"""test parameter alias input/other in dynamic graph"""
|
|
x = paddle.to_tensor(self.x_np)
|
|
y = paddle.to_tensor(self.y_np)
|
|
|
|
out1 = paddle.add(input=x, other=y, alpha=2)
|
|
out2 = x.add(other=y, alpha=2)
|
|
x_clone = x.clone()
|
|
x_clone.add_(other=y, alpha=2)
|
|
|
|
expected = self.x_np + self.y_np * 2
|
|
np.testing.assert_array_equal(out1.numpy(), expected)
|
|
np.testing.assert_array_equal(out2.numpy(), expected)
|
|
np.testing.assert_array_equal(x_clone.numpy(), expected)
|
|
|
|
# Note: y does not support scalars separately, but will support them uniformly in the future.
|
|
# def test_scalar_addition(self):
|
|
# """test scalar addition"""
|
|
# x = paddle.to_tensor(self.x_np)
|
|
|
|
# out1 = paddle.add(x, self.scalar)
|
|
# expected1 = self.x_np + self.scalar
|
|
# np.testing.assert_array_equal(out1.numpy(), expected1)
|
|
|
|
# out2 = x.add(self.scalar)
|
|
# np.testing.assert_array_equal(out2.numpy(), expected1)
|
|
|
|
# out3 = paddle.add(x, self.scalar, alpha=2)
|
|
# expected3 = self.x_np + self.scalar * 2
|
|
# np.testing.assert_array_equal(out3.numpy(), expected3)
|
|
|
|
# def test_scalar_addition_inplace(self):
|
|
# """test inplace scalar addition"""
|
|
# x = paddle.to_tensor(self.x_np)
|
|
# x_clone = x.clone()
|
|
|
|
# x_clone.add_(self.scalar)
|
|
# expected = self.x_np + self.scalar
|
|
# np.testing.assert_array_equal(x_clone.numpy(), expected)
|
|
|
|
# x_clone2 = x.clone()
|
|
# x_clone2.add_(self.scalar, alpha=2)
|
|
# expected2 = self.x_np + self.scalar * 2
|
|
# np.testing.assert_array_equal(x_clone2.numpy(), expected2)
|
|
|
|
# def test_different_dtype_scalar(self):
|
|
# """test different dtype scalar addition"""
|
|
# x = paddle.to_tensor(self.x_np)
|
|
|
|
# out1 = x.add(2)
|
|
# expected1 = self.x_np + 2
|
|
# np.testing.assert_array_equal(out1.numpy(), expected1)
|
|
|
|
# out2 = x.add(2.5)
|
|
# expected2 = self.x_np + 2.5
|
|
# np.testing.assert_array_equal(out2.numpy(), expected2)
|
|
|
|
# def test_scalar_addition_static_graph(self):
|
|
# """test static graph scalar addition"""
|
|
# paddle.enable_static()
|
|
# with paddle.static.program_guard(paddle.static.Program()):
|
|
# x = paddle.static.data(name='x', shape=[-1, 2], dtype='float32')
|
|
# out1 = paddle.add(x, self.scalar)
|
|
# out2 = paddle.add(x, self.scalar, alpha=2)
|
|
|
|
# exe = paddle.static.Executor(self.place)
|
|
# res = exe.run(
|
|
# feed={'x': self.x_np.reshape(1, 2)},
|
|
# fetch_list=[out1, out2],
|
|
# )
|
|
|
|
# expected1 = self.x_np + self.scalar
|
|
# expected2 = self.x_np + self.scalar * 2
|
|
# np.testing.assert_array_equal(res[0].flatten(), expected1)
|
|
# np.testing.assert_array_equal(res[1].flatten(), expected2)
|
|
# paddle.disable_static()
|
|
|
|
|
|
class TestAddOut(unittest.TestCase):
|
|
def setUp(self):
|
|
paddle.disable_static()
|
|
self.place = get_device_place()
|
|
|
|
def test_add_with_alpha_out(self):
|
|
def run_add_with_alpha(test_type):
|
|
x = paddle.to_tensor([1.0, 2.0, 3.0], stop_gradient=False)
|
|
y = paddle.to_tensor([4.0, 5.0, 6.0], stop_gradient=False)
|
|
out = paddle.zeros_like(x)
|
|
out.stop_gradient = False
|
|
alpha = 2.0
|
|
|
|
if test_type == "return":
|
|
out = paddle.add(x, y, alpha=alpha)
|
|
elif test_type == "input_out":
|
|
paddle.add(x, y, alpha=alpha, out=out)
|
|
elif test_type == "both_return":
|
|
out = paddle.add(x, y, alpha=alpha, out=out)
|
|
elif test_type == "both_input_out":
|
|
tmp = paddle.add(x, y, alpha=alpha, out=out)
|
|
|
|
expected = x + y * alpha
|
|
np.testing.assert_allclose(
|
|
out.numpy(),
|
|
expected.numpy(),
|
|
rtol=1e-20,
|
|
atol=1e-20,
|
|
)
|
|
|
|
loss = out.sum()
|
|
loss.backward()
|
|
|
|
return out, x.grad, y.grad, out.grad
|
|
|
|
out1, x1, y1, o1 = run_add_with_alpha("return")
|
|
out2, x2, y2, o2 = run_add_with_alpha("input_out")
|
|
out3, x3, y3, o3 = run_add_with_alpha("both_return")
|
|
out4, x4, y4, o4 = run_add_with_alpha("both_input_out")
|
|
|
|
np.testing.assert_allclose(
|
|
out1.numpy(), out2.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
out1.numpy(), out3.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
out1.numpy(), out4.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
|
|
np.testing.assert_allclose(
|
|
x1.numpy(), x2.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
x1.numpy(), x3.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
x1.numpy(), x4.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
|
|
np.testing.assert_allclose(
|
|
y1.numpy(), y2.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
y1.numpy(), y3.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
np.testing.assert_allclose(
|
|
y1.numpy(), y4.numpy(), rtol=1e-20, atol=1e-20
|
|
)
|
|
|
|
np.testing.assert_equal(o1, None)
|
|
np.testing.assert_equal(o2, None)
|
|
np.testing.assert_equal(o3, None)
|
|
np.testing.assert_equal(o4, None)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|