Merge pull request #5118 from chengduoZH/Add_deconv3d_op
Add 3D convolution transposed operator.mobile_baidu
qingqing01
7 years ago
committed by
GitHub
commit
43a64a7655
@ -1,111 +0,0 @@
|
||||
/* 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/conv2d_transpose_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 is the height of the image, and "
|
||||
"W is the width of the 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 is the height of the filter, and W is the width of the filter. "
|
||||
"We enforce groups number == 1 and padding == 0 in "
|
||||
"the 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(
|
||||
Convolution Transpose Operator.
|
||||
|
||||
The convolution transpose operation calculates the output based on the input,
|
||||
filter, strides, paddings, and groups parameters. The size of each dimension
|
||||
of the parameters is checked in the infer-shape method.
|
||||
|
||||
)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(conv2d_transpose, ops::Conv2DTransposeOp,
|
||||
ops::Conv2DTransposeOpMaker, conv2d_transpose_grad,
|
||||
ops::Conv2DTransposeOpGrad);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv2d_transpose,
|
||||
ops::GemmConv2DTransposeKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv2d_transpose_grad,
|
||||
ops::GemmConv2DTransposeGradKernel<paddle::platform::CPUPlace, float>);
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,203 @@
|
||||
/* 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/conv_transpose_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
|
||||
PADDLE_ENFORCE(ctx->HasInput("Input"),
|
||||
"Input(Input) of ConvTransposeOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("Filter"),
|
||||
"Input(Filter) of ConvTransposeOp should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Output"),
|
||||
"Output(Output) of ConvTransposeOp 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(in_dims.size() == 4 || in_dims.size() == 5,
|
||||
"ConvTransposeOp intput should be 4-D or 5-D tensor.");
|
||||
PADDLE_ENFORCE_EQ(in_dims.size(), filter_dims.size(),
|
||||
"ConvTransposeOp input dimension and filter dimension "
|
||||
"should be the same.");
|
||||
PADDLE_ENFORCE(in_dims.size() - strides.size() == 2U,
|
||||
"ConvTransposeOp input dimension and strides dimension should "
|
||||
"be consistent.");
|
||||
PADDLE_ENFORCE_EQ(paddings.size(), strides.size(),
|
||||
"ConvTransposeOp paddings dimension and Conv strides "
|
||||
"dimension should be the same.");
|
||||
PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[0],
|
||||
"In ConvTransposeOp, The input channel should be the same "
|
||||
"as the number of filters.");
|
||||
|
||||
std::vector<int64_t> output_shape({in_dims[0], filter_dims[1]});
|
||||
for (size_t i = 0; i < paddings.size(); ++i) {
|
||||
output_shape.push_back((in_dims[i + 2] - 1) * strides[i] +
|
||||
filter_dims[i + 2]);
|
||||
}
|
||||
ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
|
||||
}
|
||||
|
||||
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 is the height of the feature, and "
|
||||
"W is the width of the feature.");
|
||||
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 is the height of the filter, and W is the width of the filter. "
|
||||
"We enforce groups number == 1 and padding == 0 in "
|
||||
"the 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",
|
||||
"(vector defalut:{1, 1}), strides of convolution transpose operator.")
|
||||
.SetDefault({1, 1});
|
||||
AddAttr<std::vector<int>>(
|
||||
"paddings",
|
||||
"(vector defalut:{0, 0}), paddings of convolution transpose operator.")
|
||||
.SetDefault({0, 0});
|
||||
AddComment(R"DOC(
|
||||
Convolution2D Transpose Operator.
|
||||
|
||||
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.
|
||||
|
||||
Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch
|
||||
size, C is the number of channels, H is the height of the feature, and
|
||||
W is the width of the feature. Parameters(ksize, strides, paddings) are two elements.
|
||||
These two elements represent height and width, respectively.
|
||||
The input(X) size and output(Out) size may be different.
|
||||
Example:
|
||||
Input:
|
||||
Input shape: (N, C_in, H_in, W_in)
|
||||
Filter shape: (C_in, C_out, H_f, W_f)
|
||||
Output:
|
||||
Output shape: (N, C_out, H_out, W_out)
|
||||
where
|
||||
H_out = (H_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
|
||||
W_out = (W_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
|
||||
)DOC");
|
||||
}
|
||||
|
||||
Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(
|
||||
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 NCDHW. Where N is batch size, C is "
|
||||
"the number of channels, D is the depth of the feature, H is the "
|
||||
"height of the feature, and "
|
||||
"W is the width of the feature.");
|
||||
AddInput("Filter",
|
||||
"(Tensor) The filter tensor of convolution transpose operator."
|
||||
"The format of the filter tensor is CMDHW, where C is the number of "
|
||||
"output image channels, M is the number of input image channels, D "
|
||||
"is the depth of the filter, H is the height of the filter, and "
|
||||
"W is the width of the filter."
|
||||
"We enforce groups number == 1 and padding == 0 in "
|
||||
"the convolution3d transpose scenario.");
|
||||
AddOutput("Output",
|
||||
"(Tensor) The output tensor of convolution transpose operator."
|
||||
"The format of output tensor is also NCDHW."
|
||||
"Where N is batch size, C is "
|
||||
"the number of channels, D is the depth of the feature, H is the "
|
||||
"height of the feature, and W is the width of the feature.");
|
||||
AddAttr<std::vector<int>>(
|
||||
"strides",
|
||||
"(vector defalut:{1, 1, 1}), strides of convolution transpose operator.")
|
||||
.SetDefault({1, 1, 1});
|
||||
AddAttr<std::vector<int>>(
|
||||
"paddings",
|
||||
"(vector defalut:{0, 0, 0}), paddings of convolution transpose operator.")
|
||||
.SetDefault({0, 0, 0});
|
||||
AddComment(R"DOC(
|
||||
Convolution3D Transpose Operator.
|
||||
|
||||
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.
|
||||
|
||||
Input(Input, Filter) and output(Output) are in NCDHW format. Where N is batch
|
||||
size, C is the number of channels, D is the depth of the feature,
|
||||
H is the height of the feature, and W is the width of the feature.
|
||||
Parameters(ksize, strides, paddings) are three elements.
|
||||
These three elements represent depth, height and width, respectively.
|
||||
The input(X) size and output(Out) size may be different.
|
||||
Example:
|
||||
Input:
|
||||
Input shape: (N, C_in, D_in, H_in, W_in)
|
||||
Filter shape: (C_in, C_out, D_f, H_f, W_f)
|
||||
Output:
|
||||
Output shape: (N, C_out, D_out, H_out, W_out)
|
||||
where
|
||||
D_out = (D_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
|
||||
H_out = (H_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
|
||||
W_out = (W_in - 1) * strides[2] - 2 * paddings[2] + filter_size[2];
|
||||
)DOC");
|
||||
}
|
||||
|
||||
void ConvTransposeOpGrad::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(conv2d_transpose, ops::ConvTransposeOp, ops::Conv2DTransposeOpMaker,
|
||||
conv2d_transpose_grad, ops::ConvTransposeOpGrad);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv2d_transpose,
|
||||
ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv2d_transpose_grad,
|
||||
ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>);
|
||||
|
||||
REGISTER_OP(conv3d_transpose, ops::ConvTransposeOp, ops::Conv3DTransposeOpMaker,
|
||||
conv3d_transpose_grad, ops::ConvTransposeOpGrad);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv3d_transpose,
|
||||
ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
conv3d_transpose_grad,
|
||||
ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>);
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,97 @@
|
||||
import unittest
|
||||
import numpy as np
|
||||
from op_test import OpTest
|
||||
|
||||
|
||||
def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
|
||||
# [2, 3, 5, 5, 5]
|
||||
in_n, in_c, in_d, in_h, in_w = input_.shape
|
||||
# [3, 6, 3, 3, 3]
|
||||
f_c, out_c, f_d, f_h, f_w = filter_.shape
|
||||
assert in_c == f_c
|
||||
|
||||
stride, pad = conv3dtranspose_param['stride'], conv3dtranspose_param['pad']
|
||||
out_d = (in_d - 1) * stride[0] + f_d
|
||||
out_h = (in_h - 1) * stride[1] + f_h
|
||||
out_w = (in_w - 1) * stride[2] + f_w
|
||||
|
||||
out = np.zeros((in_n, out_c, out_d, out_h, out_w))
|
||||
|
||||
for n in range(in_n):
|
||||
for d in range(in_d):
|
||||
for i in range(in_h):
|
||||
for j in range(in_w):
|
||||
input_masked = input_[n, :, d, i, j] # (c)
|
||||
input_masked = np.reshape(input_masked, (in_c, 1, 1, 1))
|
||||
input_masked = np.tile(input_masked, (1, f_d, f_h, f_w))
|
||||
|
||||
for k in range(out_c):
|
||||
tmp_out = np.sum(input_masked * filter_[:, k, :, :, :],
|
||||
axis=0)
|
||||
d1, d2 = d * stride[0], d * stride[0] + f_d
|
||||
i1, i2 = i * stride[1], i * stride[1] + f_h
|
||||
j1, j2 = j * stride[2], j * stride[2] + f_w
|
||||
out[n, k, d1:d2, i1:i2, j1:j2] += tmp_out
|
||||
|
||||
return out
|
||||
|
||||
|
||||
class TestConv3dTransposeOp(OpTest):
|
||||
def setUp(self):
|
||||
# init as conv transpose
|
||||
self.init_op_type()
|
||||
|
||||
# [2, 3, 5, 5, 5] -> kernel [3, 6, 3, 3, 3] -> output [2, 6, 7, 7, 7]
|
||||
self.init_test_case()
|
||||
|
||||
conv3dtranspose_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 = conv3dtranspose_forward_naive(
|
||||
input_, filter_, conv3dtranspose_param).astype("float32")
|
||||
# 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=0.02)
|
||||
|
||||
def test_check_grad_no_filter(self):
|
||||
self.check_grad(
|
||||
['Input'],
|
||||
'Output',
|
||||
max_relative_error=0.02,
|
||||
no_grad_set=set(['Filter']))
|
||||
|
||||
def test_check_grad_no_input(self):
|
||||
self.check_grad(
|
||||
['Filter'],
|
||||
'Output',
|
||||
max_relative_error=0.02,
|
||||
no_grad_set=set(['Input']))
|
||||
|
||||
def init_test_case(self):
|
||||
self.pad = [0, 0, 0]
|
||||
self.stride = [1, 1, 1]
|
||||
self.dilations = [1, 1, 1]
|
||||
self.input_size = [2, 3, 5, 5, 5] # NCHW
|
||||
f_c = self.input_size[1]
|
||||
self.filter_size = [f_c, 6, 3, 3, 3]
|
||||
|
||||
def init_op_type(self):
|
||||
self.op_type = "conv3d_transpose"
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Loading…
Reference in new issue