提交 fd66089d 编写于 作者: D dengkaipeng 提交者: ceci3

add spectral_norm forwarn kenel

上级 cab46b62
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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/spectral_norm_op.h"
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
using framework::Tensor;
class SpectralNormOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Weight"),
"Input(Weight) of SpectralNormOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("U"),
"Input(U) of SpectralNormOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("V"),
"Input(V) of SpectralNormOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of SpectralNormOp should not be null.");
auto dim_weight = ctx->GetInputDim("Weight");
auto weight_dimsize = dim_weight.size();
PADDLE_ENFORCE(weight_dimsize >= 2 && weight_dimsize <= 5,
"The size of dims of Input(Weights) can only be 2, 3,"
"4, 5 for fc, conv1d, conv2d, conv3d layers.");
int dim = ctx->Attrs().Get<int>("dim");
int power_iters = ctx->Attrs().Get<int>("power_iters");
PADDLE_ENFORCE(dim >= 0 && dim < weight_dimsize - 1,
"Attr(dim) should be larger equal 0 and less then the"
"size of dims of Input(Weights) - 1,");
PADDLE_ENFORCE(power_iters >= 0,
"Attr(power_iters) should be larger equal then 0");
ctx->SetOutputDim("Out", dim_weight);
ctx->ShareLoD("Weight", /*->*/ "Out");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(ctx.Input<Tensor>("Weight")->type(),
ctx.GetPlace());
}
};
class SpectralNormOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Weight",
"The input weight tensor of spectral_norm operator, "
"This can be a 2-D, 3-D, 4-D, 5-D tensor which is the"
"weights of fc, conv1d, conv2d, conv3d layer.");
AddInput("U",
"The weight_u tensor of spectral_norm operator, "
"This can be a 1-D tensor in shape [H, 1],"
"H is the 1st dimentions of Weight after reshape"
"corresponding by Attr(dim).");
AddInput("V",
"The weight_u tensor of spectral_norm operator, "
"This can be a 1-D tensor in shape [W, 1],"
"W is the 2nd dimentions of Weight after reshape"
"corresponding by Attr(dim).");
AddOutput("Out",
"The output weight tensor of spectral_norm operator, "
"This tensor is in same shape with Input(Weight).");
AddAttr<int>("dim",
"dimension corresponding to number of outputs,"
"default 0 for fc layer, and 1 for conv1d, conv2d, conv3d"
"layers")
.SetDefault(0);
AddAttr<int>("power_iters",
"number of power iterations to calculate"
"spectral norm, default is 1.")
.SetDefault(1);
AddAttr<float>("eps",
"epsilob for numerical stability in"
"calculating norms")
.SetDefault(1e-12);
AddComment(R"DOC(
This operator samples input X to given output shape by using specified
)DOC");
}
};
class SpectralNormOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Weight"), "Input(Weight) should not be null");
PADDLE_ENFORCE(ctx->HasInput("U"), "Input(U) should not be null");
PADDLE_ENFORCE(ctx->HasInput("V"), "Input(V) should not be null");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
"Input(Out@GRAD) should not be null");
auto dim_x = ctx->GetInputDim("Weight");
if (ctx->HasOutput(framework::GradVarName("Weight"))) {
ctx->SetOutputDim(framework::GradVarName("Weight"), dim_x);
}
}
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(ctx.Input<Tensor>("Weight")->type(),
ctx.GetPlace());
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(spectral_norm, ops::SpectralNormOp, ops::SpectralNormOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(spectral_norm_grad, ops::SpectralNormOpGrad);
REGISTER_OP_CPU_KERNEL(
spectral_norm,
ops::SpectralNormKernel<paddle::platform::CPUDeviceContext, float>,
ops::SpectralNormKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
spectral_norm_grad,
ops::SpectralNormGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::SpectralNormGradKernel<paddle::platform::CPUDeviceContext, double>);
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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 "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
using Tensor = framework::Tensor;
using Array1 = Eigen::DSizes<int64_t, 1>;
using Array2 = Eigen::DSizes<int64_t, 2>;
using IndexPair = Eigen::IndexPair<int>;
static inline void ResizeWeight(Tensor* weight_mat, const int dim) {
auto weight_dims = weight_mat->dims();
int h = 1;
int w = 1;
for (int i = 0; i < weight_dims.size(); i++) {
if (i <= dim) {
h *= weight_dims[i];
} else {
w *= weight_dims[i];
}
}
*weight_mat = weight_mat->Resize({h, w});
}
template <typename DeviceContext, typename T>
static inline void CalcMatrixSigmaAndNormWeight(
Tensor* sigma, Tensor* u, Tensor* v, Tensor* weight, const int power_iters,
const float eps, const framework::ExecutionContext& ctx) {
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto sigma_t = EigenTensor<T, 2>::From(*sigma);
auto weight_t = EigenTensor<T, 2>::From(*weight);
auto u_t = EigenTensor<T, 1>::From(*u);
auto v_t = EigenTensor<T, 1>::From(*v);
const int h = weight->dims()[0];
const int w = weight->dims()[1];
Eigen::array<int, 2> perm = {1, 0};
Eigen::array<IndexPair, 1> product_dims = {IndexPair(1, 0)};
auto weight_trans_t = weight_t.shuffle(perm);
LOG(ERROR) << "weight: " << weight_t;
LOG(ERROR) << "weight_trans: " << weight_trans_t;
for (int i = 0; i < power_iters; i++) {
v_t.device(place) = weight_trans_t.contract(u_t, product_dims);
LOG(ERROR) << "iter v: " << v_t;
auto v_t_norm =
v_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast(
Array1(w));
LOG(ERROR) << "iter v_norm: " << v_t_norm;
v_t.device(place) = v_t / (v_t_norm + v_t_norm.constant(eps));
LOG(ERROR) << "iter norm v: " << v_t;
u_t.device(place) = weight_t.contract(v_t, product_dims);
LOG(ERROR) << "iter u: " << u_t;
auto u_t_norm =
u_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast(
Array1(h));
u_t.device(place) = u_t / (u_t_norm + u_t_norm.constant(eps));
LOG(ERROR) << "iter norm u: " << u_t;
}
LOG(ERROR) << "h" << h << "w" << w;
LOG(ERROR) << "u: " << u_t;
LOG(ERROR) << "v: " << v_t;
LOG(ERROR) << "weight_v: " << weight_t.contract(v_t, product_dims);
sigma_t.device(place) = (u_t * weight_t.contract(v_t, product_dims))
.sum()
.eval()
.reshape(Array2(1, 1))
.broadcast(Array2(h, w));
LOG(ERROR) << "weight: " << weight_t;
LOG(ERROR) << "sigma: " << sigma_t;
weight_t.device(place) = weight_t / sigma_t;
}
template <typename DeviceContext, typename T>
class SpectralNormKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto weight = ctx.Input<Tensor>("Weight");
auto u = ctx.Input<Tensor>("U");
auto v = ctx.Input<Tensor>("V");
auto out = ctx.Output<Tensor>("Out");
int dim = ctx.Attr<int>("dim");
int power_iters = ctx.Attr<int>("power_iters");
float eps = ctx.Attr<float>("eps");
Tensor weight_mat;
TensorCopySync(*weight, ctx.GetPlace(), &weight_mat);
ResizeWeight(&weight_mat, dim);
Tensor sigma;
sigma.mutable_data<T>(weight->dims(), ctx.GetPlace());
Tensor uu, vv;
TensorCopySync(*u, ctx.GetPlace(), &uu);
TensorCopySync(*v, ctx.GetPlace(), &vv);
CalcMatrixSigmaAndNormWeight<DeviceContext, T>(
&sigma, &uu, &vv, &weight_mat, power_iters, eps, ctx);
TensorCopySync(weight_mat, ctx.GetPlace(), out);
}
};
template <typename DeviceContext, typename T>
class SpectralNormGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {}
};
} // namespace operators
} // namespace paddle
# Copyright (c) 2018 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 division
import unittest
import numpy as np
from op_test import OpTest
from paddle.fluid import core
class TestSpectralNormOp(OpTest):
def setUp(self):
self.initTestCase()
self.op_type = 'spectral_norm'
# weight = np.random.random(self.weight_shape).astype('float32')
# u = np.random.random(self.u_shape).astype('float32')
# v = np.random.random(self.u_shape).astype('float32')
weight = np.ones(self.weight_shape).astype('float32')
weight[1, :] = 2.
u = np.ones(self.u_shape).astype('float32')
v = np.ones(self.v_shape).astype('float32')
self.attrs = {
"dim": self.dim,
"power_iters": self.power_iters,
"eps": self.eps,
}
self.inputs = {
"Weight": weight,
"U": u,
"V": v,
}
output = weight
self.outputs = {"Out": weight, }
def test_check_output(self):
self.check_output()
def initTestCase(self):
self.weight_shape = (2, 3)
self.u_shape = (2, )
self.v_shape = (3, )
self.dim = 0
self.power_iters = 1
self.eps = 1e-12
if __name__ == "__main__":
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册