654 lines
21 KiB
Python
654 lines
21 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, is_custom_device
|
|
|
|
import paddle
|
|
from paddle import core
|
|
|
|
|
|
class TestNonBroadcastableMismatchedShapeCase(unittest.TestCase):
|
|
"""Unittest from PyTorch comparison and handcrafted backward result
|
|
Note that this unit test might fail, if you modify the implementation
|
|
of scatter and gather kernel, especially the ordering of atomic writes
|
|
|
|
So make sure you know what you are doing, otherwise
|
|
you may need to update this unittest.
|
|
"""
|
|
|
|
def setUp(self):
|
|
self.input = paddle.to_tensor(
|
|
[
|
|
[
|
|
[
|
|
[1.9693925, 2.2913685],
|
|
[-0.19461553, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.29458013, 0.51647896],
|
|
[0.79423386, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
stop_gradient=False,
|
|
)
|
|
self.index = paddle.to_tensor(
|
|
[[[[0], [1]]], [[[1], [0]]]], dtype='int64', stop_gradient=True
|
|
)
|
|
self.src = paddle.to_tensor(
|
|
[
|
|
[
|
|
[[-2.1342657], [-0.6801669], [-0.741744]],
|
|
[[-0.15918107], [1.5543042], [-0.35116914]],
|
|
],
|
|
[
|
|
[[0.39571938], [0.5322498], [-0.35833976]],
|
|
[[1.3826214], [0.6314196], [0.891596]],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
stop_gradient=False,
|
|
)
|
|
self.no_grad = False
|
|
self.dim = 2
|
|
self.include_self = True
|
|
|
|
def test_no_grad_add(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='add',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[-0.16487312, 2.2913685],
|
|
[-0.87478244, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.8268299, 0.51647896],
|
|
[1.1899532, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
def test_with_grad_assign(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='assign',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[-2.1342657, 2.2913685],
|
|
[-0.6801669, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.5322498, 0.51647896],
|
|
[0.39571938, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
result.backward()
|
|
gt_input_grad = np.array(
|
|
[
|
|
[
|
|
[[0.0, 1.0], [0.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
[
|
|
[[0.0, 1.0], [0.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
gt_src_grad = np.array(
|
|
[[[[1.0], [1.0]]], [[[1.0], [1.0]]]], dtype='float32'
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.input.grad.numpy(), gt_input_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.src.grad.numpy(), gt_src_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
def test_no_grad_mul(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='mul',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[-4.203207, 2.2913685],
|
|
[0.13237104, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.15679021, 0.51647896],
|
|
[0.31429374, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
def test_with_grad_amin(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='amin',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[-2.1342657, 2.2913685],
|
|
[-0.6801669, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.29458013, 0.51647896],
|
|
[0.39571938, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
result.backward()
|
|
gt_input_grad = np.array(
|
|
[
|
|
[
|
|
[[0.0, 1.0], [0.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
[
|
|
[[1.0, 1.0], [0.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
gt_src_grad = np.array(
|
|
[[[[1.0], [1.0]]], [[[0.0], [0.0]]]], dtype='float32'
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.input.grad.numpy(), gt_input_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.src.grad.numpy(), gt_src_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
def test_with_grad_amax(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='amax',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[1.9693925, 2.2913685],
|
|
[-0.19461553, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.5322498, 0.51647896],
|
|
[0.79423386, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
result.backward()
|
|
gt_input_grad = np.array(
|
|
[
|
|
[
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
[
|
|
[[0.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
[[1.0, 1.0], [1.0, 1.0], [1.0, 1.0]],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
gt_src_grad = np.array(
|
|
[[[[0.0], [0.0]]], [[[0.0], [0.0]]]], dtype='float32'
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.input.grad.numpy(), gt_input_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.src.grad.numpy(), gt_src_grad, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
def test_no_grad_mean(self):
|
|
self.input.clear_grad()
|
|
self.src.clear_grad()
|
|
result = paddle.put_along_axis(
|
|
self.input,
|
|
indices=self.index,
|
|
values=self.src,
|
|
axis=self.dim,
|
|
reduce='mean',
|
|
include_self=self.include_self,
|
|
broadcast=False,
|
|
)
|
|
gt_result = np.array(
|
|
[
|
|
[
|
|
[
|
|
[-0.08243656, 2.2913685],
|
|
[-0.43739122, 0.298859],
|
|
[-0.86006254, 0.28243607],
|
|
],
|
|
[
|
|
[-0.09577879, -0.10506158],
|
|
[-0.12375893, 1.4438118],
|
|
[-0.66273206, 1.0404967],
|
|
],
|
|
],
|
|
[
|
|
[
|
|
[0.41341496, 0.51647896],
|
|
[0.5949766, -1.5084593],
|
|
[0.405428, -0.8155419],
|
|
],
|
|
[
|
|
[0.27907062, 0.70933336],
|
|
[-1.2590513, 0.7363407],
|
|
[1.078117, -0.03632839],
|
|
],
|
|
],
|
|
],
|
|
dtype='float32',
|
|
)
|
|
np.testing.assert_allclose(
|
|
result.numpy(), gt_result, rtol=1e-6, atol=1e-6
|
|
)
|
|
|
|
|
|
class TestPutAlongAxisNonIncludeSelf2ndGrad(unittest.TestCase):
|
|
"""Test case from issue 72803"""
|
|
|
|
def setUp(self):
|
|
self.x = np.array(
|
|
[
|
|
[1.6947253, 1.7280283, -1.1000537, -1.7621638, -0.46924523],
|
|
[-0.17813402, 0.9851728, 0.8784995, -0.35652128, 0.63679916],
|
|
[-0.2506482, 0.46839848, 1.6940045, 1.2753638, -1.5601108],
|
|
[-1.4223574, -0.30286825, -0.6940945, 0.4153872, -1.598482],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.indices = np.array(
|
|
[
|
|
[3, 2, 2, 2, 0],
|
|
[1, 1, 3, 1, 3],
|
|
[0, 0, 3, 2, 3],
|
|
[0, 1, 2, 0, 3],
|
|
],
|
|
dtype="int64",
|
|
)
|
|
self.values = np.array(
|
|
[
|
|
[-0.3371469, -2.3898945, -0.6047427, -0.18021728, 1.0270963],
|
|
[-0.4792783, -0.06155855, -1.1657414, -0.22004248, -1.2116293],
|
|
[-1.2325171, -1.2428453, -0.53471214, 0.64549965, 0.3991431],
|
|
[-0.45945236, -0.2563897, -1.2712464, 1.7996459, -0.08381622],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.dout = np.array(
|
|
[
|
|
[-0.19797462, -0.98365456, 1.936407, -0.0050864, -1.0364918],
|
|
[1.0826564, -2.1047552, 0.9298107, 0.6769417, 0.9323797],
|
|
[-0.68968654, -0.5532966, 0.24068666, 0.5625817, 1.8991498],
|
|
[0.84938127, -0.5345554, -0.6814333, -1.0064939, 2.419181],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.ddx = np.array(
|
|
[
|
|
[0.3573612, -0.6587053, -1.0527273, 0.7391721, -0.16440763],
|
|
[-1.67882, -0.46170056, -0.81231886, 0.6644795, 1.0688623],
|
|
[-1.3970909, 0.17792162, 0.35944283, -0.00945398, -1.8379706],
|
|
[0.99883825, 0.47824964, -1.4997529, 0.80206966, -0.24591826],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.ddv = np.array(
|
|
[
|
|
[0.31652406, -0.41458955, -0.46466753, -0.23473991, 0.25190634],
|
|
[-1.3948212, -0.84799731, 0.5940094, 0.46881115, 0.4054867],
|
|
[-2.0037501, 0.087257907, 1.0091733, -0.002437128, 0.67401189],
|
|
[-0.10354018, 0.51002628, -2.5794835, -1.7636456, -0.59410858],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.gt_result = np.array(
|
|
[
|
|
[-1.6919695, -1.2428453, -1.1000537, 1.7996459, 1.0270963],
|
|
[-0.4792783, -0.31794825, 0.8784995, -0.22004248, 0.63679916],
|
|
[-0.2506482, -2.3898945, -1.8759892, 0.46528238, -1.5601108],
|
|
[-0.3371469, -0.30286825, -1.7004535, 0.4153872, -0.8963024],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.gt_dx = np.array(
|
|
[
|
|
[0.0, 0.0, 1.936407, 0.0, 0.0],
|
|
[0.0, 0.0, 0.9298107, 0.0, 0.9323797],
|
|
[-0.68968654, 0.0, 0.0, 0.0, 1.8991498],
|
|
[0.0, -0.5345554, 0.0, -1.0064939, 0.0],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.gt_dv = np.array(
|
|
[
|
|
[0.84938127, -0.5532966, 0.24068666, 0.5625817, -1.0364918],
|
|
[1.0826564, -2.1047552, -0.6814333, 0.6769417, 2.419181],
|
|
[-0.19797462, -0.98365456, -0.6814333, 0.5625817, 2.419181],
|
|
[-0.19797462, -2.1047552, 0.24068666, -0.0050864, 2.419181],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
self.gt_ddout = np.array(
|
|
[
|
|
[-2.1072903, 0.08725791, -1.0527273, -1.7636456, 0.25190634],
|
|
[-1.3948212, -0.33797103, -0.81231886, 0.46881115, 1.0688623],
|
|
[-1.3970909, -0.41458955, -3.044151, -0.23717704, -1.8379706],
|
|
[0.31652406, 0.47824964, 1.6031827, 0.80206966, 0.48538995],
|
|
],
|
|
dtype="float32",
|
|
)
|
|
|
|
def test_2nd_grad(self):
|
|
x = paddle.to_tensor(self.x)
|
|
x.stop_gradient = False
|
|
include_self = False
|
|
axis = 0
|
|
|
|
indices = paddle.to_tensor(self.indices)
|
|
|
|
values = paddle.to_tensor(self.values)
|
|
values.stop_gradient = False
|
|
|
|
out = paddle.put_along_axis(
|
|
x,
|
|
indices,
|
|
values,
|
|
axis,
|
|
'add',
|
|
include_self=include_self,
|
|
)
|
|
|
|
dout = paddle.to_tensor(self.dout)
|
|
dout.stop_gradient = False
|
|
|
|
dx, dv = paddle.grad(
|
|
out,
|
|
[x, values],
|
|
dout,
|
|
create_graph=True,
|
|
)
|
|
|
|
ddx = paddle.to_tensor(self.ddx)
|
|
ddx.stop_gradient = False
|
|
ddv = paddle.to_tensor(self.ddv)
|
|
ddv.stop_gradient = False
|
|
|
|
ddout = paddle.grad(
|
|
[dx, dv],
|
|
dout,
|
|
[ddx, ddv],
|
|
)[0]
|
|
|
|
np.testing.assert_allclose(out.numpy(), self.gt_result, 1e-6, 1e-6)
|
|
np.testing.assert_allclose(dx.numpy(), self.gt_dx, 1e-6, 1e-6)
|
|
np.testing.assert_allclose(dv.numpy(), self.gt_dv, 1e-6, 1e-6)
|
|
np.testing.assert_allclose(ddout.numpy(), self.gt_ddout, 1e-6, 1e-6)
|
|
|
|
|
|
@unittest.skipIf(
|
|
not (core.is_compiled_with_cuda() or is_custom_device()),
|
|
"CPU FP16 is not supported",
|
|
)
|
|
class TestPutAlongAxisFP16MulDuplicatedIndices(unittest.TestCase):
|
|
def setUp(self):
|
|
self.input = paddle.ones(16, dtype=paddle.float16)
|
|
self.src = paddle.arange(
|
|
0.9, 0.9 + 0.02 * 16, 0.02, dtype=paddle.float16
|
|
)
|
|
self.index = paddle.zeros(16, dtype=paddle.int64)
|
|
|
|
def test_fp16_mul_reduce(self):
|
|
res = paddle.put_along_axis(
|
|
self.input, self.index, self.src, axis=0, reduce='mul'
|
|
)
|
|
gt = np.ones(16, dtype=np.float64)
|
|
gt[0] = np.arange(0.9, 0.9 + 16 * 0.02, 0.02).prod()
|
|
np.testing.assert_allclose(
|
|
res.numpy().astype(np.float64), gt, rtol=1e-2, atol=1e-2
|
|
)
|
|
|
|
|
|
class TestPutAlongAxisIntegerMean(unittest.TestCase):
|
|
def setUp(self):
|
|
self.gt_include_self = np.array(
|
|
[
|
|
[[-8, -7, -7, -7], [-12, -11, -10, -9]],
|
|
[[-5, -5, -4, -4], [-4, -3, -2, -1]],
|
|
[[-2, -2, -2, -1], [4, 5, 6, 7]],
|
|
[[0, 1, 1, 1], [12, 13, 14, 15]],
|
|
],
|
|
dtype='int32',
|
|
)
|
|
self.gt_exclude_self = np.array(
|
|
[
|
|
[[-3, -3, -3, -3], [-12, -11, -10, -9]],
|
|
[[-3, -3, -3, -3], [-4, -3, -2, -1]],
|
|
[[-3, -3, -3, -3], [4, 5, 6, 7]],
|
|
[[-3, -3, -3, -3], [12, 13, 14, 15]],
|
|
],
|
|
dtype='int32',
|
|
)
|
|
|
|
def _make_static_mean_int(self, gt, include_self, place):
|
|
paddle.enable_static()
|
|
with paddle.static.program_guard(paddle.static.Program()):
|
|
input_ = paddle.arange(-16, 16, 1, dtype=paddle.int32).reshape(
|
|
[4, 2, 4]
|
|
)
|
|
src = paddle.full([4, 2, 4], -3, dtype=paddle.int32)
|
|
index = paddle.zeros([4, 2, 4], dtype=paddle.int64)
|
|
result = paddle.put_along_axis(
|
|
input_,
|
|
indices=index,
|
|
values=src,
|
|
axis=1,
|
|
reduce='mean',
|
|
include_self=include_self,
|
|
)
|
|
|
|
exe = paddle.static.Executor(place)
|
|
result_np = exe.run(fetch_list=[result])
|
|
np.testing.assert_array_equal(result_np[0], gt)
|
|
paddle.disable_static()
|
|
|
|
def test_mean_int(self):
|
|
# try testing with both CPU and GPU places
|
|
if paddle.is_compiled_with_cuda() or is_custom_device():
|
|
self._make_static_mean_int(
|
|
self.gt_include_self, True, get_device_place()
|
|
)
|
|
self._make_static_mean_int(
|
|
self.gt_exclude_self, False, get_device_place()
|
|
)
|
|
self._make_static_mean_int(
|
|
self.gt_include_self, True, paddle.CPUPlace()
|
|
)
|
|
self._make_static_mean_int(
|
|
self.gt_exclude_self, False, paddle.CPUPlace()
|
|
)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|