提交 8fdc315a 编写于 作者: Z Zhuoyuan 提交者: GitHub

Merge pull request #4739 from zchen0211/develop

deconv op implementing ...
/* Copyright (c) 2016 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/operators/conv2dtranspose_op.h"
namespace paddle {
namespace operators {
void Conv2DTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of Conv2DTransposeOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Filter"),
"Input(Filter) of Conv2DTransposeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Output"),
"Output(Output) of Conv2DTransposeOp should not be null.");
auto in_dims = ctx->GetInputDim("Input");
auto filter_dims = ctx->GetInputDim("Filter");
std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
for (size_t i = 0; i < paddings.size(); ++i) {
PADDLE_ENFORCE_EQ(paddings[i], 0,
"No Padding allowed in conv transpose op.");
}
PADDLE_ENFORCE_EQ(in_dims.size(), 4,
"Conv2DTransposeOp input should be 4-D tensor.");
PADDLE_ENFORCE_EQ(filter_dims.size(), 4,
"Conv2DTransposeOp filter should be 4-D tensor.");
PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[0],
"input and kernel input dimension should be equal.");
auto output_height = (in_dims[2] - 1) * strides[0] + filter_dims[2];
auto output_width = (in_dims[3] - 1) * strides[1] + filter_dims[3];
ctx->SetOutputDim("Output",
{in_dims[0], filter_dims[1], output_height, output_width});
}
Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(
framework::OpProto* proto, framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput(
"Input",
"(Tensor) The input tensor of convolution transpose operator. "
"The format of input tensor is NCHW. Where N is batch size, C is the "
"number of input channels, H and W is the height and width of image.");
AddInput("Filter",
"(Tensor) The filter tensor of convolution transpose operator."
"The format of the filter tensor is CMHW, where C is the number of "
"output image channels, M is the number of input image channels, "
"H and W is height and width of filter. "
"We enforce groups number == 1 and padding == 0 in "
"convolution transpose Scenario.");
AddOutput("Output",
"(Tensor) The output tensor of convolution transpose operator."
"The format of output tensor is also NCHW.");
AddAttr<std::vector<int>>("strides",
"strides of convolution transpose operator.")
.SetDefault({1, 1});
AddAttr<std::vector<int>>("paddings",
"paddings of convolution transpose operator.")
.SetDefault({0, 0});
AddComment(R"DOC(
The convolution transpose operation calculates the output based on the input, filter
and strides, paddings, groups parameters. The size of each dimension of the
parameters is checked in the infer-shape.
)DOC");
}
void Conv2DTransposeOpGrad::InferShape(
framework::InferShapeContext* ctx) const {
auto in_dims = ctx->GetInputDim("Input");
auto filter_dims = ctx->GetInputDim("Filter");
if (ctx->HasOutput(framework::GradVarName("Input"))) {
ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
}
if (ctx->HasOutput(framework::GradVarName("Filter"))) {
ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
}
}
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(conv2dtranspose, ops::Conv2DTransposeOp,
ops::Conv2DTransposeOpMaker, conv2dtranspose_grad,
ops::Conv2DTransposeOpGrad);
REGISTER_OP_CPU_KERNEL(
conv2dtranspose,
ops::GemmConv2DTransposeKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
conv2dtranspose_grad,
ops::GemmConv2DTransposeGradKernel<paddle::platform::CPUPlace, float>);
/* Copyright (c) 2016 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/operators/conv2dtranspose_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
conv2dtranspose,
ops::GemmConv2DTransposeKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
conv2dtranspose_grad,
ops::GemmConv2DTransposeGradKernel<paddle::platform::GPUPlace, float>);
/* Copyright (c) 2016 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/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/im2col.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using DDim = framework::DDim;
// Define Op classes in .h file so that other conv transpose
// operator implementations can reuse the code.
class Conv2DTransposeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
Conv2DTransposeOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker);
};
class Conv2DTransposeOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override;
};
class Conv2DTransposeOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override;
};
template <typename Place, typename T>
class GemmConv2DTransposeKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
// The filter will be reshaped, so it should not be constant pointer
Tensor filter = *context.Input<Tensor>("Filter");
Tensor* output = context.Output<Tensor>("Output");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
// TODO(Zhuoyuan): Paddings can be added in future.
// groups will alway be disabled in conv2dtranspose.
const int batch_size = input->dims()[0];
const int m = input->dims()[1];
const int h = input->dims()[2];
const int w = input->dims()[3];
const int k_h = filter.dims()[2];
const int k_w = filter.dims()[3];
const int c = output->dims()[1]; // output channels
const int o_h = output->dims()[2];
const int o_w = output->dims()[3];
paddle::operators::math::Col2ImFunctor<
paddle::operators::math::ColFormat::kCFO, Place, T>
col2im;
// use col_shape in the im2col and col2im calculation
DDim col_shape = {c, k_h, k_w, h, w};
// use col_matrix_shape in the gemm calculation
DDim col_matrix_shape = {c * k_h * k_w, h * w};
Tensor col;
col.mutable_data<T>(col_shape, context.GetPlace());
// col_matrix shares the same piece of data with col,
// but will be reshaped into a two-dimensional matrix shape
// to call the matrix multiplication interface.
Tensor col_matrix;
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
DDim output_shape = {c, o_h, o_w};
DDim input_matrix_shape = {m, h * w};
DDim filter_matrix_shape = {m, c * k_h * k_w};
filter.Resize(filter_matrix_shape);
// convolution transpose: gemm + col2im (similar to conv-backward on input)
output->mutable_data<T>(context.GetPlace());
auto t = framework::EigenVector<T>::Flatten(*output);
t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
for (int i = 0; i < batch_size; i++) {
// batch with size (M, h * w)
Tensor input_batch = input->Slice(i, i + 1).Resize(input_matrix_shape);
// filter size: (M, c * k_h * k_w)
// output size: (c, o_h, o_w)
Tensor output_batch = output->Slice(i, i + 1).Resize(output_shape);
// col_matrix = filter * input_batch
// of shape (c * k_h * k_w, h * w)
math::matmul<Place, T>(context.device_context(), filter, true,
input_batch, false, T(1.0), &col_matrix, T(0.0));
col2im(context.device_context(), output_batch, col, strides[0],
strides[1], 0, 0);
}
}
};
template <typename Place, typename T>
class GemmConv2DTransposeGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
const Tensor* output_grad =
context.Input<Tensor>(framework::GradVarName("Output"));
// For filter, we do not use const pointer b/c we will do reshape,
// but we should avoid modifying its value.
Tensor filter = *context.Input<Tensor>("Filter");
Tensor* input_grad =
context.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad =
context.Output<Tensor>(framework::GradVarName("Filter"));
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
// Actually, no paddings and groups allowed in conv transpose.
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
const int batch_size = input->dims()[0];
const int m = input->dims()[1];
const int h = input->dims()[2];
const int w = input->dims()[3];
const int k_h = filter.dims()[2];
const int k_w = filter.dims()[3];
const int c = output_grad->dims()[1]; // output channels
const int o_h = output_grad->dims()[2];
const int o_w = output_grad->dims()[3];
// Only im2col functor required for bp to get to the right shape
paddle::operators::math::Im2ColFunctor<
paddle::operators::math::ColFormat::kCFO, Place, T>
im2col;
// use col_shape in the im2col and col2im calculation
DDim col_shape = {c, k_h, k_w, h, w};
// use col_matrix_shape in the gemm calculation
DDim col_matrix_shape_f = {c * h * w, k_h * k_w};
Tensor col;
col.mutable_data<T>(col_shape, context.GetPlace());
// col_matrix shares the same piece of data with col,
// but will be reshaped into a two-dimensional matrix shape
// to call the matrix multiplication interface.
DDim output_shape = {c, o_h, o_w};
DDim input_matrix_shape = {m, h * w};
DDim filter_matrix_shape = {m, c * k_h * k_w};
filter.Resize(filter_matrix_shape);
// convolution transpose grad on input:
// im2col + gemm (similar to conv-forward)
// input need to compute gradient
if (input_grad) {
Tensor col_matrix;
col_matrix.ShareDataWith(col);
DDim col_matrix_shape = {c * k_h * k_w, h * w};
col_matrix.Resize(col_matrix_shape);
input_grad->mutable_data<T>(context.GetPlace());
auto t = framework::EigenVector<T>::Flatten(*input_grad);
t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
for (int i = 0; i < batch_size; i++) {
// batch with size (c, o_h * o_w)
Tensor output_grad_batch =
output_grad->Slice(i, i + 1).Resize(output_shape);
// filter of size (m, c * k_h * k_w)
// batch with size (m, h, w)
Tensor input_grad_batch =
input_grad->Slice(i, i + 1).Resize(input_matrix_shape);
// im2col: dy from (c, o_h, o_w) -> (c * k_h * k_w, h * w)
im2col(context.device_context(), output_grad_batch, col, strides[0],
strides[1], paddings[0], paddings[1]);
// gemm: dx = filter * dy
// (m, c * k_h * k_w) * (c * k_h * k_w, h * w) -> (m, c, h)
math::matmul<Place, T>(context.device_context(), filter, false,
col_matrix, false, T(1.0), &input_grad_batch,
T(0.0));
}
}
// filter gradient required
if (filter_grad) {
Tensor col_matrix_f;
col_matrix_f.ShareDataWith(col);
DDim col_matrix_shape_f = {c * h * w, k_h * k_w};
col_matrix_f.Resize(col_matrix_shape_f);
filter_grad->mutable_data<T>(context.GetPlace());
Tensor filter_grad_ = *filter_grad;
filter_grad_.Resize(filter_matrix_shape);
auto t = framework::EigenVector<T>::Flatten(filter_grad_);
t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
for (int i = 0; i < batch_size; ++i) {
// batch with size (c, o_h, o_w)
Tensor output_grad_batch =
output_grad->Slice(i, i + 1).Resize(output_shape);
// input batch
Tensor in_batch = input->Slice(i, i + 1).Resize(input_matrix_shape);
// im2col: (c * h * w, k_h * k_w)
im2col(context.device_context(), output_grad_batch, col, strides[0],
strides[1], paddings[0], paddings[1]);
// gemm: d_filter = x * y_grad^T
// (m, c * h * w) * (k_h * k_w, c * h * w) -> (m, c, h)
math::matmul<Place, T>(context.device_context(), in_batch, false,
col_matrix_f, true, T(1.0), &filter_grad_,
T(1.0));
}
}
}
};
} // namespace operators
} // namespace paddle
import unittest
import numpy as np
from op_test import OpTest
def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param):
# [2, 3, 5, 5]
in_n, in_c, in_h, in_w = input_.shape
# [3, 6, 3, 3]
f_c, out_c, f_h, f_w = filter_.shape
assert in_c == f_c
stride, pad = conv2dtranspose_param['stride'], conv2dtranspose_param['pad']
out_h = (in_h - 1) * stride[0] + f_h
out_w = (in_w - 1) * stride[1] + f_w
out = np.zeros((in_n, out_c, out_h, out_w))
for n in range(in_n):
for i in range(in_h):
for j in range(in_w):
input_masked = input_[n, :, i, j] # (c)
input_masked = np.reshape(input_masked, (in_c, 1, 1))
input_masked = np.tile(input_masked, (1, f_h, f_w))
for k in range(out_c):
tmp_out = np.sum(input_masked * filter_[:, k, :, :], axis=0)
i1, i2 = i * stride[0], i * stride[0] + f_h
j1, j2 = j * stride[0], j * stride[0] + f_w
out[n, k, i1:i2, j1:j2] += tmp_out
return out
class TestConv2dTransposeOp(OpTest):
def setUp(self):
# init as conv transpose
self.init_op_type()
# [2, 3, 5, 5] -> kernel [3, 6, 3, 3] -> output [2, 6, 7, 7]
self.init_test_case()
conv2dtranspose_param = {'stride': self.stride, 'pad': self.pad}
input_ = np.random.random(self.input_size).astype("float32")
filter_ = np.random.random(self.filter_size).astype("float32")
output = conv2dtranspose_forward_naive(input_, filter_,
conv2dtranspose_param)
# print 'deconv output py', output, output.shape
self.inputs = {'Input': input_, 'Filter': filter_}
self.attrs = {
'strides': self.stride,
'paddings': self.pad,
# 'dilations': self.dilations
}
self.outputs = {'Output': output}
def test_check_output(self):
print 'check output here'
self.check_output()
def test_check_grad(self):
self.check_grad(
set(['Input', 'Filter']), 'Output', max_relative_error=0.05)
def test_check_grad_no_filter(self):
self.check_grad(
['Input'],
'Output',
max_relative_error=0.05,
no_grad_set=set(['Filter']))
def test_check_grad_no_input(self):
self.check_grad(
['Filter'],
'Output',
max_relative_error=0.05,
no_grad_set=set(['Input']))
def init_test_case(self):
self.pad = [0, 0]
self.stride = [1, 1]
self.dilations = [1, 1]
self.input_size = [2, 3, 5, 5] # NCHW
f_c = self.input_size[1]
self.filter_size = [f_c, 6, 3, 3]
def init_op_type(self):
self.op_type = "conv2dtranspose"
"""
class TestCudnn(TestConv2dOp):
def init_group(self):
self.groups = 1
def init_op_type(self):
self.op_type = "conv_cudnn"
"""
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