Zhang Ting
5 years ago
committed by
GitHub
parent
1c08a2136e
commit
4773e3f541
@ -0,0 +1,119 @@
|
|||||||
|
// Copyright (c) 2020 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/dist_op.h"
|
||||||
|
#include <string>
|
||||||
|
#include <vector>
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
class DistOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||||
|
OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Dist");
|
||||||
|
OP_INOUT_CHECK(ctx->HasInput("Y"), "Input", "Y", "Dist");
|
||||||
|
OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "Dist");
|
||||||
|
ctx->SetOutputDim("Out", {1});
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class DistOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||||
|
public:
|
||||||
|
void Make() override {
|
||||||
|
AddInput("X", "The input Tensor of Dist Op.");
|
||||||
|
AddInput("Y", "The Right-hand-side input Tensor of Dist Op.");
|
||||||
|
AddOutput("Out",
|
||||||
|
"The output of Dist Op, "
|
||||||
|
"which is the p-norm of (X - Y)");
|
||||||
|
AddAttr<float>("p", "the norm to be computed.").SetDefault(2.0f);
|
||||||
|
AddComment(R"DOC(
|
||||||
|
Dist Operator.
|
||||||
|
Given two tensors X and Y, compute Lp-norm of (X-Y). It is not a norm in a strict sense,
|
||||||
|
only as a measure of distance. The shapes of X and Y must be broadcastable. Where, Z = X - Y,
|
||||||
|
|
||||||
|
When p = 0, defining $0^0 = 0$, the zero-norm of Z is simply the number of non-zero elements of z.
|
||||||
|
$$
|
||||||
|
||Z||_{0} = \lim_{p \rightarrow 0} \sum_{i=1}^{m} |z_i|^p
|
||||||
|
$$
|
||||||
|
|
||||||
|
When p = inf, the inf-norm of Z is the maximum element of Z.
|
||||||
|
$$
|
||||||
|
||Z||_\infty=\max_i |z_i|
|
||||||
|
$$
|
||||||
|
|
||||||
|
When p = -inf, the negative-inf-norm of Z is the minimum element of Z.
|
||||||
|
$$
|
||||||
|
||Z||_{-\infty}=\min_i |z_i|
|
||||||
|
$$
|
||||||
|
|
||||||
|
Otherwise, the p-norm of Z follows the formula,
|
||||||
|
$$
|
||||||
|
||Z||_{p} = (\sum_{i=i}^{m} |z_i|^p)^{1/p}
|
||||||
|
$$
|
||||||
|
)DOC");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class DistOpGrad : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||||
|
auto x_dims = ctx->GetInputDim("X");
|
||||||
|
auto y_dims = ctx->GetInputDim("Y");
|
||||||
|
if (ctx->HasOutput(framework::GradVarName("X"))) {
|
||||||
|
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
|
||||||
|
}
|
||||||
|
if (ctx->HasOutput(framework::GradVarName("Y"))) {
|
||||||
|
ctx->SetOutputDim(framework::GradVarName("Y"), y_dims);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class DistGradOpMaker : public framework::SingleGradOpMaker<T> {
|
||||||
|
public:
|
||||||
|
using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
|
||||||
|
|
||||||
|
protected:
|
||||||
|
void Apply(GradOpPtr<T> op) const override {
|
||||||
|
op->SetType(this->ForwardOpType() + "_grad");
|
||||||
|
op->SetInput("X", this->Input("X"));
|
||||||
|
op->SetInput("Y", this->Input("Y"));
|
||||||
|
op->SetInput("Out", this->Output("Out"));
|
||||||
|
op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
|
||||||
|
|
||||||
|
op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
|
||||||
|
op->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
|
||||||
|
op->SetAttrMap(this->Attrs());
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OPERATOR(dist, ops::DistOp, ops::DistOpMaker,
|
||||||
|
ops::DistGradOpMaker<paddle::framework::OpDesc>,
|
||||||
|
ops::DistGradOpMaker<paddle::imperative::OpBase>);
|
||||||
|
REGISTER_OPERATOR(dist_grad, ops::DistOpGrad);
|
||||||
|
REGISTER_OP_CPU_KERNEL(
|
||||||
|
dist, ops::DistKernel<paddle::platform::CPUDeviceContext, float>,
|
||||||
|
ops::DistKernel<paddle::platform::CPUDeviceContext, double>);
|
||||||
|
REGISTER_OP_CPU_KERNEL(
|
||||||
|
dist_grad, ops::DistGradKernel<paddle::platform::CPUDeviceContext, float>,
|
||||||
|
ops::DistGradKernel<paddle::platform::CPUDeviceContext, double>)
|
||||||
@ -0,0 +1,23 @@
|
|||||||
|
// Copyright (c) 2020 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/dist_op.h"
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
dist, ops::DistKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::DistKernel<paddle::platform::CUDADeviceContext, double>);
|
||||||
|
REGISTER_OP_CUDA_KERNEL(
|
||||||
|
dist_grad, ops::DistGradKernel<paddle::platform::CUDADeviceContext, float>,
|
||||||
|
ops::DistGradKernel<paddle::platform::CUDADeviceContext, double>);
|
||||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,165 @@
|
|||||||
|
# Copyright (c) 2020 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 OpTest
|
||||||
|
import paddle
|
||||||
|
import paddle.fluid as fluid
|
||||||
|
import paddle.fluid.core as core
|
||||||
|
|
||||||
|
|
||||||
|
def dist(x, y, p):
|
||||||
|
if p == 0.:
|
||||||
|
out = np.count_nonzero(x - y)
|
||||||
|
elif p == float("inf"):
|
||||||
|
out = np.max(np.abs(x - y))
|
||||||
|
elif p == float("-inf"):
|
||||||
|
out = np.min(np.abs(x - y))
|
||||||
|
else:
|
||||||
|
out = np.power(np.sum(np.power(np.abs(x - y), p)), 1.0 / p)
|
||||||
|
return np.array(out).astype(x.dtype)
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOp(OpTest):
|
||||||
|
def setUp(self):
|
||||||
|
self.op_type = 'dist'
|
||||||
|
self.attrs = {}
|
||||||
|
self.init_case()
|
||||||
|
self.inputs = {
|
||||||
|
"X": np.random.random(self.x_shape).astype("float64"),
|
||||||
|
"Y": np.random.random(self.y_shape).astype("float64")
|
||||||
|
}
|
||||||
|
|
||||||
|
self.attrs["p"] = self.p
|
||||||
|
self.outputs = {
|
||||||
|
"Out": dist(self.inputs["X"], self.inputs["Y"], self.attrs["p"])
|
||||||
|
}
|
||||||
|
self.gradient = self.calc_gradient()
|
||||||
|
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (120)
|
||||||
|
self.y_shape = (120)
|
||||||
|
self.p = 0.
|
||||||
|
|
||||||
|
def calc_gradient(self):
|
||||||
|
x = self.inputs["X"]
|
||||||
|
y = self.inputs["Y"]
|
||||||
|
p = self.attrs["p"]
|
||||||
|
if p == 0:
|
||||||
|
grad = np.zeros(x.shape)
|
||||||
|
elif p in [float("inf"), float("-inf")]:
|
||||||
|
norm = dist(x, y, p)
|
||||||
|
x_minux_y_abs = np.abs(x - y)
|
||||||
|
grad = np.sign(x - y)
|
||||||
|
grad[x_minux_y_abs != norm] = 0
|
||||||
|
else:
|
||||||
|
norm = dist(x, y, p)
|
||||||
|
grad = np.power(norm, 1 - p) * np.power(np.abs(x - y),
|
||||||
|
p - 1) * np.sign(x - y)
|
||||||
|
|
||||||
|
def get_reduce_dims(x, y):
|
||||||
|
x_reduce_dims = []
|
||||||
|
y_reduce_dims = []
|
||||||
|
|
||||||
|
if x.ndim >= y.ndim:
|
||||||
|
y_reshape = tuple([1] * (x.ndim - y.ndim) + list(y.shape))
|
||||||
|
y = y.reshape(y_reshape)
|
||||||
|
else:
|
||||||
|
x_reshape = tuple([1] * (y.ndim - x.ndim) + list(x.shape))
|
||||||
|
x = x.reshape(x_reshape)
|
||||||
|
for i in range(x.ndim):
|
||||||
|
if x.shape[i] > y.shape[i]:
|
||||||
|
y_reduce_dims.append(i)
|
||||||
|
elif x.shape[i] < y.shape[i]:
|
||||||
|
x_reduce_dims.append(i)
|
||||||
|
return x_reduce_dims, y_reduce_dims
|
||||||
|
|
||||||
|
x_reduce_dims, y_reduce_dims = get_reduce_dims(x, y)
|
||||||
|
if len(x_reduce_dims) != 0:
|
||||||
|
x_grad = np.sum(grad, tuple(x_reduce_dims)).reshape(x.shape)
|
||||||
|
else:
|
||||||
|
x_grad = grad
|
||||||
|
if len(y_reduce_dims) != 0:
|
||||||
|
y_grad = -np.sum(grad, tuple(y_reduce_dims)).reshape(y.shape)
|
||||||
|
else:
|
||||||
|
y_grad = -grad
|
||||||
|
|
||||||
|
return x_grad, y_grad
|
||||||
|
|
||||||
|
def test_check_output(self):
|
||||||
|
self.check_output()
|
||||||
|
|
||||||
|
def test_check_grad(self):
|
||||||
|
self.check_grad(["X", "Y"], "Out", user_defined_grads=self.gradient)
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOpCase1(TestDistOp):
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (3, 5, 5, 6)
|
||||||
|
self.y_shape = (5, 5, 6)
|
||||||
|
self.p = 1.
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOpCase2(TestDistOp):
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (10, 10)
|
||||||
|
self.y_shape = (4, 10, 10)
|
||||||
|
self.p = 2.
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOpCase3(TestDistOp):
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (15, 10)
|
||||||
|
self.y_shape = (15, 10)
|
||||||
|
self.p = float("inf")
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOpCase4(TestDistOp):
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (2, 3, 4, 5, 8)
|
||||||
|
self.y_shape = (3, 1, 5, 8)
|
||||||
|
self.p = float("-inf")
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistOpCase5(TestDistOp):
|
||||||
|
def init_case(self):
|
||||||
|
self.x_shape = (4, 1, 4, 8)
|
||||||
|
self.y_shape = (2, 2, 1, 4, 4, 8)
|
||||||
|
self.p = 1.5
|
||||||
|
|
||||||
|
|
||||||
|
class TestDistAPI(unittest.TestCase):
|
||||||
|
def test_api(self):
|
||||||
|
main_program = fluid.Program()
|
||||||
|
startup_program = fluid.Program()
|
||||||
|
with fluid.program_guard(main_program, startup_program):
|
||||||
|
x = fluid.data(name='x', shape=[2, 3, 4, 5], dtype='float64')
|
||||||
|
y = fluid.data(name='y', shape=[3, 1, 5], dtype='float64')
|
||||||
|
p = 2
|
||||||
|
x_i = np.random.random((2, 3, 4, 5)).astype("float64")
|
||||||
|
y_i = np.random.random((3, 1, 5)).astype("float64")
|
||||||
|
result = paddle.dist(x, y, p)
|
||||||
|
place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
|
||||||
|
) else fluid.CPUPlace()
|
||||||
|
exe = fluid.Executor(place)
|
||||||
|
out = exe.run(fluid.default_main_program(),
|
||||||
|
feed={'x': x_i,
|
||||||
|
'y': y_i},
|
||||||
|
fetch_list=[result])
|
||||||
|
self.assertTrue(np.allclose(dist(x_i, y_i, p), out[0]))
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
unittest.main()
|
||||||
Loading…
Reference in new issue