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

add dilation for im2col

Browse Source
updateWriteDocsCN
chengduoZH 8 years ago
parent 91b724821c
commit 97e9dd7237
9 changed files with 395 additions and 351 deletions
Show Stats Download Patch File Download Diff File

2
paddle/operators/conv_cudnn_op.cc
Unescape View File

@ -22,8 +22,6 @@ class CudnnConvOpMaker : public Conv2DOpMaker {
CudnnConvOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: Conv2DOpMaker(proto, op_checker) {
AddAttr<std::vector<int>>("dilations", "dilations of convolution operator.")
.SetDefault(std::vector<int>{1, 1});
AddAttr<int>("workspace_size_MB",
"workspace size for cudnn, in MB, "
"workspace is a section of GPU memory which will be "

13
paddle/operators/conv_op.cc
Unescape View File

@ -30,6 +30,7 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
int groups = ctx->Attrs().Get<int>("groups");
std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
int input_channels = in_dims[1];
int output_channels = filter_dims[0];
@ -54,7 +55,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
for (size_t i = 0; i < paddings.size(); ++i) {
output_shape.push_back(OutputSize(in_dims[i + 2], filter_dims[i + 2],
paddings[i], strides[i]));
dilations[i], paddings[i], paddings[i],
strides[i]));
}
ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
}
@ -90,6 +92,10 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto,
"first half of the input channels, while the second half of the filters "
"is only connected to the second half of the input channels.")
.SetDefault(1);
AddAttr<std::vector<int>>("dilations",
"(vector default:{1, 1}), the dilations of "
"convolution operator.")
.SetDefault(std::vector<int>{1, 1});
AddComment(R"DOC(
Convolution Operator.
@ -151,6 +157,11 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto,
"first half of the input channels, while the second half of the filters "
"is only connected to the second half of the input channels.")
.SetDefault(1);
AddAttr<std::vector<int>>("dilations",
"(vector default:{1, 1, 1}), the dilations of "
"convolution operator. Currently, conv3d doesn't "
"support dilation.")
.SetDefault(std::vector<int>{1, 1, 1});
AddComment(R"DOC(
Convolution3D Operator.

29
paddle/operators/conv_op.h
Unescape View File

@ -27,9 +27,12 @@ using Tensor = framework::Tensor;
// Base convolution operator definations for other conv
// like operators to reuse the implementation.
inline int OutputSize(int input_size, int filter_size, int padding,
int stride) {
int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
inline int OutputSize(int input_size, int filter_size, int dilation,
int padding_up, int padding_down, int stride) {
int output_size = (input_size + padding_up + padding_down -
(dilation * (filter_size - 1) + 1)) /
stride +
1;
return output_size;
}
@ -76,6 +79,7 @@ class GemmConvKernel : public framework::OpKernel<T> {
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
int groups = context.Attr<int>("groups");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
const int batch_size = static_cast<int>(input->dims()[0]);
@ -139,9 +143,9 @@ class GemmConvKernel : public framework::OpKernel<T> {
if (filter_shape_vec.size() == 2) {
// im2col
math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
im2col(context.device_context(), in_slice, col, strides[0],
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);
im2col(context.device_context(), in_slice, col, dilations[0],
dilations[1], strides[0], strides[1], paddings[0], paddings[0],
paddings[1], paddings[1]);
} else if (filter_shape_vec.size() == 3) {
// vol2col
math::Vol2ColFunctor<Place, T> vol2col;
@ -181,6 +185,7 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
int groups = context.Attr<int>("groups");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
const int batch_size = static_cast<int>(input->dims()[0]);
@ -263,9 +268,9 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
if (filter_shape_vec.size() == 2) {
math::Col2ImFunctor<math::ColFormat::kCFO, Place, T> col2im;
col2im(context.device_context(), in_grad_slice, col, strides[0],
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);
col2im(context.device_context(), in_grad_slice, col, dilations[0],
dilations[1], strides[0], strides[1], paddings[0],
paddings[0], paddings[1], paddings[1]);
} else if (filter_shape_vec.size() == 3) {
math::Col2VolFunctor<Place, T> col2vol;
@ -295,9 +300,9 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
if (filter_shape_vec.size() == 2) {
math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
im2col(context.device_context(), in_slice, col, strides[0],
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);
im2col(context.device_context(), in_slice, col, dilations[0],
dilations[1], strides[0], strides[1], paddings[0],
paddings[0], paddings[1], paddings[1]);
} else if (filter_shape_vec.size() == 3) {
math::Vol2ColFunctor<Place, T> vol2col;
vol2col(context.device_context(), in_slice, col, strides[0],

16
paddle/operators/conv_transpose_op.h
Unescape View File

@ -69,6 +69,9 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
// TODO(Zhuoyuan): Paddings can be added in future.
// groups will alway be disabled in conv2dtranspose.
int dilation_h = 1;
int dilation_w = 1;
const int batch_size = static_cast<int>(input->dims()[0]);
// input_shape_vec: {h, w} or {d, h, w}
@ -140,8 +143,8 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
// from (c * k_h * k_w, h * w) to (c, o_h, o_w)
math::Col2ImFunctor<math::ColFormat::kCFO, Place, T> col2im;
col2im(context.device_context(), output_batch, col, strides[0],
strides[1], 0, 0, 0, 0);
col2im(context.device_context(), output_batch, col, dilation_h,
dilation_w, strides[0], strides[1], 0, 0, 0, 0);
} else if (filter_shape_vec.size() == 3) {
// col2vol: col_matrix -> dy
// from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w)
@ -174,6 +177,9 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
// Actually, no paddings and groups allowed in conv transpose.
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
int dilation_h = 1;
int dilation_w = 1;
const int batch_size = static_cast<int>(input->dims()[0]);
// input_shape_vec: {h, w} or {d, h, w}
@ -248,9 +254,9 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
// im2col: dy -> col matrix
// from (c, o_h, o_w) to (c * k_h * k_w, h * w)
math::Im2ColFunctor<math::ColFormat::kCFO, Place, T> im2col;
im2col(context.device_context(), output_grad_batch, col, strides[0],
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);
im2col(context.device_context(), output_grad_batch, col, dilation_h,
dilation_w, strides[0], strides[1], paddings[0], paddings[0],
paddings[1], paddings[1]);
} else if (filter_shape_vec.size() == 3) {
// vol2col: dy -> col_matrix
// from (c, o_d, o_h, o_w) to (c * k_d * k_h * k_w, d * h * w)

10
paddle/operators/math/context_project.h
Unescape View File

@ -95,6 +95,9 @@ class ContextProjectFunctor {
math::Im2ColFunctor<math::ColFormat::kOCF, Place, float> im2col_ocf;
int dilation_h = 1;
int dilation_w = 1;
int input_row_begin, input_row_end;
int sequence_height, sequence_width;
sequence_width = in.dims()[1];
@ -124,7 +127,7 @@ class ContextProjectFunctor {
sequence_width}); // input_channels, input_height, input_width
in_t.Resize(framework::make_ddim(input_shape));
im2col_ocf(context, in_t, out_t,
im2col_ocf(context, in_t, out_t, dilation_h, dilation_w,
/*stride_height*/ context_stride, /*stride_width*/ 1, up_pad,
down_pad, 0, 0);
out_t.Resize({sequence_height, context_length * sequence_width});
@ -204,6 +207,9 @@ class ContextProjectGradFunctor {
math::Col2ImFunctor<math::ColFormat::kOCF, Place, float> col2im_ocf;
int dilation_h = 1;
int dilation_w = 1;
int input_row_begin, input_row_end;
int sequence_height, sequence_width;
sequence_width = in.dims()[1];
@ -234,7 +240,7 @@ class ContextProjectGradFunctor {
sequence_width}); // input_channels, input_height, input_width
in_t.Resize(framework::make_ddim(input_shape));
col2im_ocf(context, in_t, out_t,
col2im_ocf(context, in_t, out_t, dilation_h, dilation_w,
/*stride_height*/ context_stride, /*stride_width*/ 1,
up_pad, down_pad, 0, 0);
out_t.Resize({sequence_height, context_length * sequence_width});

281
paddle/operators/math/im2col.cc
View File

File diff suppressed because it is too large Load Diff

366
paddle/operators/math/im2col.cu
View File

File diff suppressed because it is too large Load Diff

11
paddle/operators/math/im2col.h
Unescape View File

@ -74,17 +74,18 @@ class Im2ColFunctor {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& im, framework::Tensor& col,
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right);
int dilation_h, int dilation_w, int stride_height,
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right);
};
template <ColFormat Format, typename Place, typename T>
class Col2ImFunctor {
public:
void operator()(const platform::DeviceContext& context, framework::Tensor& im,
const framework::Tensor& col, int stride_height,
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right);
const framework::Tensor& col, int dilation_h, int dilation_w,
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right);
};
} // namespace math

18
paddle/operators/math/im2col_test.cc
Unescape View File

@ -47,6 +47,8 @@ void testIm2col() {
int filter_size = 2;
int stride = 1;
int padding = 0;
int dilation_h = 1;
int dilation_w = 1;
int output_height = (input_height - filter_size + 2 * padding) / stride + 1;
int output_width = (input_width - filter_size + 2 * padding) / stride + 1;
float* input_ptr = input_tmp.mutable_data<float>(
@ -85,10 +87,10 @@ void testIm2col() {
paddle::operators::math::ColFormat::kOCF, Place, float>
im2col_ocf;
im2col(*context, input, output_cfo, stride, stride, padding, padding, padding,
padding);
im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding,
padding, padding);
im2col(*context, input, output_cfo, dilation_h, dilation_w, stride, stride,
padding, padding, padding, padding);
im2col_ocf(*context, input, output_ocf, dilation_h, dilation_w, stride,
stride, padding, padding, padding, padding);
float out_cfo_data[] = {0, 1, 1, 2, 3, 4, 4, 5};
float out_ocf_data[] = {0, 1, 3, 4, 1, 2, 4, 5};
@ -131,8 +133,8 @@ void testIm2col() {
input.CopyFrom(input_tmp, *place, *context);
}
col2im(*context, input, output_cfo, stride, stride, padding, padding, padding,
padding);
col2im(*context, input, output_cfo, dilation_h, dilation_w, stride, stride,
padding, padding, padding, padding);
float* in_ptr;
if (paddle::platform::is_cpu_place(*place)) {
@ -153,8 +155,8 @@ void testIm2col() {
input.CopyFrom(input_tmp, *place, *context);
}
col2im_ocf(*context, input, output_ocf, stride, stride, padding, padding,
padding, padding);
col2im_ocf(*context, input, output_ocf, dilation_h, dilation_w, stride,
stride, padding, padding, padding, padding);
if (paddle::platform::is_cpu_place(*place)) {
in_ptr = input.data<float>();

Write Preview
Loading…
Cancel
Save