m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add kernel for squeeze_op, test=develop (#19656)

Browse Source 
* add kernel for squeeze_op, test=develop

* delete comment, test=develop
expand_as_op_1
zhongpu 5 years ago committed by Jiabin Yang
parent 2a81c3679a
commit 52673956de
5 changed files with 398 additions and 153 deletions
Show Stats Download Patch File Download Diff File

275
paddle/fluid/operators/squeeze_op.cc
View File

File diff suppressed because it is too large Load Diff

44
paddle/fluid/operators/squeeze_op.cu.cc
Unescape View File

@ -0,0 +1,44 @@
/* Copyright (c) 2019 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. */
#include "paddle/fluid/operators/squeeze_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
squeeze, ops::SqueezeKernel<paddle::platform::CUDADeviceContext, float>,
ops::SqueezeKernel<paddle::platform::CUDADeviceContext, double>,
ops::SqueezeKernel<paddle::platform::CUDADeviceContext, int>,
ops::SqueezeKernel<paddle::platform::CUDADeviceContext, int8_t>,
ops::SqueezeKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
squeeze_grad,
ops::SqueezeGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::SqueezeGradKernel<paddle::platform::CUDADeviceContext, double>,
ops::SqueezeGradKernel<paddle::platform::CUDADeviceContext, int>,
ops::SqueezeGradKernel<paddle::platform::CUDADeviceContext, int8_t>,
ops::SqueezeGradKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
squeeze2, ops::Squeeze2Kernel<paddle::platform::CUDADeviceContext, float>,
ops::Squeeze2Kernel<paddle::platform::CUDADeviceContext, double>,
ops::Squeeze2Kernel<paddle::platform::CUDADeviceContext, int>,
ops::Squeeze2Kernel<paddle::platform::CUDADeviceContext, int8_t>,
ops::Squeeze2Kernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
squeeze2_grad,
ops::Squeeze2GradKernel<paddle::platform::CUDADeviceContext, float>,
ops::Squeeze2GradKernel<paddle::platform::CUDADeviceContext, double>,
ops::Squeeze2GradKernel<paddle::platform::CUDADeviceContext, int>,
ops::Squeeze2GradKernel<paddle::platform::CUDADeviceContext, int8_t>,
ops::Squeeze2GradKernel<paddle::platform::CUDADeviceContext, int64_t>);

146
paddle/fluid/operators/squeeze_op.h
Unescape View File

@ -0,0 +1,146 @@
/* Copyright (c) 2019 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. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/pooling.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class SqueezeKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
auto *in = context.Input<framework::LoDTensor>("X");
auto *out = context.Output<framework::LoDTensor>("Out");
auto &axes = context.Attr<std::vector<int>>("axes");
auto x_dims = in->dims();
auto out_dims = GetOutputShape(axes, x_dims);
out->mutable_data(context.GetPlace(), in->type());
framework::TensorCopy(
*in, context.GetPlace(),
context.template device_context<platform::DeviceContext>(), out);
out->Resize(out_dims);
}
static framework::DDim GetOutputShape(const std::vector<int> squeeze_dims,
const framework::DDim &in_dims) {
size_t num_squeeze_dims = squeeze_dims.size();
int cnt_squeezed_dims = 0;
bool should_squeeze[9] = {false};
// Determines number of dimensions of output tensor after squeeze.
// Mark and count the dimensions need to be squeezed
if (num_squeeze_dims == 0) {
for (int idx = 0; idx < in_dims.size(); ++idx) {
if (in_dims[idx] == 1) {
should_squeeze[idx] = true;
++cnt_squeezed_dims;
}
}
} else {
for (size_t idx = 0; idx < num_squeeze_dims; ++idx) {
int current = squeeze_dims[idx] < 0 ? squeeze_dims[idx] + in_dims.size()
: squeeze_dims[idx];
// Check current index, the upper limit has beed checked in line 36.
PADDLE_ENFORCE_GE(current, 0,
"Invalid axis, the negative axis is out of range.");
PADDLE_ENFORCE_EQ(in_dims[current], 1,
"Invalid axis index, the axis that will be squeezed "
"should be equal to 1.");
if (!(should_squeeze[current])) {
++cnt_squeezed_dims;
}
should_squeeze[current] = true;
}
}
// Make output dimensions
std::vector<int64_t> output_shape(in_dims.size() - cnt_squeezed_dims, 0);
for (int in_idx = 0, out_idx = 0; in_idx < in_dims.size(); ++in_idx) {
if (!should_squeeze[in_idx]) {
output_shape[out_idx++] = in_dims[in_idx];
}
}
return framework::make_ddim(output_shape);
}
};
template <typename DeviceContext, typename T>
class SqueezeGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto *d_out =
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
auto *d_x = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
auto in_dims = ctx.Input<framework::LoDTensor>("X")->dims();
d_x->mutable_data(ctx.GetPlace(), d_out->type());
framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
d_x->Resize(in_dims);
}
};
template <typename DeviceContext, typename T>
class Squeeze2Kernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
auto *out = context.Output<framework::LoDTensor>("Out");
auto *in = context.Input<framework::LoDTensor>("X");
auto &axes = context.Attr<std::vector<int>>("axes");
auto x_dims = in->dims();
auto out_dims =
SqueezeKernel<DeviceContext, T>::GetOutputShape(axes, x_dims);
out->mutable_data(context.GetPlace(), in->type());
framework::TensorCopy(
*in, context.GetPlace(),
context.template device_context<platform::DeviceContext>(), out);
out->Resize(out_dims);
}
};
template <typename DeviceContext, typename T>
class Squeeze2GradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto *d_out =
ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
auto *d_x = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
// auto in_dims = d_x->dims();
auto xshape_dims = ctx.Input<framework::LoDTensor>("XShape")->dims();
auto x_dims = framework::slice_ddim(xshape_dims, 1, xshape_dims.size());
d_x->mutable_data(ctx.GetPlace(), d_out->type());
framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
d_x->Resize(x_dims);
}
};
} // namespace operators
} // namespace paddle

75
python/paddle/fluid/tests/unittests/test_squeeze2_op.py
Unescape View File

@ -0,0 +1,75 @@
# Copyright (c) 2019 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.
from __future__ import print_function
import unittest
import numpy as np
from op_test import OpTest
# Correct: General.
class TestSqueezeOp(OpTest):
def setUp(self):
self.op_type = "squeeze2"
self.init_test_case()
self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
self.init_attrs()
self.outputs = {
"Out": self.inputs["X"].reshape(self.new_shape),
"XShape": np.random.random(self.ori_shape).astype("float32")
}
def test_check_output(self):
self.check_output(no_check_set=['XShape'])
def test_check_grad(self):
self.check_grad(["X"], "Out")
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, 2)
self.new_shape = (3, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes}
# Correct: There is mins axis.
class TestSqueezeOp1(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, -2)
self.new_shape = (3, 5)
# Correct: No axes input.
class TestSqueezeOp2(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = ()
self.new_shape = (3, 5)
# Correct: Just part of axes be squeezed.
class TestSqueezeOp3(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 1, 5, 1, 4, 1)
self.axes = (1, -1)
self.new_shape = (3, 5, 1, 4)
if __name__ == "__main__":
unittest.main()

11
python/paddle/fluid/tests/unittests/test_squeeze_op.py
Unescape View File

@ -1,4 +1,4 @@
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2019 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.
@ -23,17 +23,14 @@ from op_test import OpTest
# Correct: General.
class TestSqueezeOp(OpTest):
def setUp(self):
self.op_type = "squeeze2"
self.op_type = "squeeze"
self.init_test_case()
self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
self.init_attrs()
self.outputs = {
"Out": self.inputs["X"].reshape(self.new_shape),
"XShape": np.random.random(self.ori_shape).astype("float32")
}
self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape), }
def test_check_output(self):
self.check_output(no_check_set=['XShape'])
self.check_output()
def test_check_grad(self):
self.check_grad(["X"], "Out")

Write Preview
Loading…
Cancel
Save