Skip to content

Add up, down, left and right padding for im2col. #4887

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Oct 24, 2017
Merged

Add up, down, left and right padding for im2col. #4887

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
d2c1408
fix im2col kocf for sequence projection
chengduoZH Oct 16, 2017
dc7d073
add padding up, down, left, right
chengduoZH Oct 21, 2017
2947f56
follow comments
chengduoZH Oct 23, 2017
811a317
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
chengduoZH Oct 23, 2017
09662da
follow comments
chengduoZH Oct 23, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions paddle/operators/conv2d_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,7 @@ class GemmConv2DKernel : public framework::OpKernel<T> {
// im2col
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
im2col(context.device_context(), in_slice, col, strides[0], strides[1],
paddings[0], paddings[1]);
paddings[0], paddings[0], paddings[1], paddings[1]);

// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Expand Down Expand Up @@ -213,7 +213,8 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
Tensor in_grad_slice =
in_grad_batch.Slice(g * in_step, (g + 1) * in_step);
col2im(context.device_context(), in_grad_slice, col, strides[0],
strides[1], paddings[0], paddings[1]);
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);
}
}
}
Expand All @@ -235,7 +236,8 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
im2col(context.device_context(), in_slice, col, strides[0],
strides[1], paddings[0], paddings[1]);
strides[1], paddings[0], paddings[0], paddings[1],
paddings[1]);

// gemm
Tensor filter_grad_slice =
Expand Down
114 changes: 87 additions & 27 deletions paddle/operators/math/im2col.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,8 +29,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& im, framework::Tensor& col,
int stride_height, int stride_width, int padding_height,
int padding_width) {
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);

Expand All @@ -41,6 +41,22 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
int filter_width = col.dims()[2];
int output_height = col.dims()[3];
int output_width = col.dims()[4];

PADDLE_ENFORCE_EQ(
(input_height + padding_up + padding_down - filter_height) /
stride_height +
1,
output_height,
"Output_height and padding(padding_up, padding_down) are "
"inconsistent.");
PADDLE_ENFORCE_EQ(
(input_width + padding_left + padding_right - filter_width) /
stride_width +
1,
output_width,
"output_width and padding(padding_left, padding_right) are "
"inconsistent.");

int channels_col = input_channels * filter_height * filter_width;

const T* im_data = im.data<T>();
Expand All @@ -52,16 +68,14 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
int c_im = c / filter_width / filter_height;
for (int h = 0; h < output_height; ++h) {
for (int w = 0; w < output_width; ++w) {
int im_row_idx = h * stride_height + h_offset;
int im_col_idx = w * stride_width + w_offset;
if ((im_row_idx - padding_height) < 0 ||
(im_row_idx - padding_height) >= input_height ||
(im_col_idx - padding_width) < 0 ||
(im_col_idx - padding_width) >= input_width) {
int im_row_idx = h * stride_height + h_offset - padding_up;
int im_col_idx = w * stride_width + w_offset - padding_left;

if (im_row_idx < 0 || im_row_idx >= input_height || im_col_idx < 0 ||
im_col_idx >= input_width) {
col_data[(c * output_height + h) * output_width + w] = T(0);
} else {
im_row_idx += c_im * input_height - padding_height;
im_col_idx -= padding_width;
im_row_idx += c_im * input_height;
col_data[(c * output_height + h) * output_width + w] =
im_data[im_row_idx * input_width + im_col_idx];
}
Expand All @@ -82,7 +96,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
public:
void operator()(const platform::DeviceContext& context, framework::Tensor& im,
const framework::Tensor& col, int stride_height,
int stride_width, int padding_height, int padding_width) {
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);
int input_channels = im.dims()[0];
Expand All @@ -92,6 +107,22 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
int filter_width = col.dims()[2];
int output_height = col.dims()[3];
int output_width = col.dims()[4];

PADDLE_ENFORCE_EQ(
(input_height + padding_up + padding_down - filter_height) /
stride_height +
1,
output_height,
"Output_height and padding(padding_up, padding_down) are "
"inconsistent.");
PADDLE_ENFORCE_EQ(
(input_width + padding_left + padding_right - filter_width) /
stride_width +
1,
output_width,
"output_width and padding(padding_left, padding_right) are "
"inconsistent.");

int channels_col = input_channels * filter_height * filter_width;

T* im_data = im.data<T>();
Expand All @@ -103,14 +134,12 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
int c_im = c / filter_width / filter_height;
for (int h = 0; h < output_height; ++h) {
for (int w = 0; w < output_width; ++w) {
int im_row_idx = h * stride_height + h_offset;
int im_col_idx = w * stride_width + w_offset;
if ((im_row_idx - padding_height) >= 0 &&
(im_row_idx - padding_height) < input_height &&
(im_col_idx - padding_width) >= 0 &&
(im_col_idx - padding_width) < input_width) {
im_row_idx += c_im * input_height - padding_height;
im_col_idx -= padding_width;
int im_row_idx = h * stride_height + h_offset - padding_up;
int im_col_idx = w * stride_width + w_offset - padding_left;

if ((im_row_idx) >= 0 && (im_row_idx) < input_height &&
(im_col_idx) >= 0 && (im_col_idx) < input_width) {
im_row_idx += c_im * input_height;
im_data[im_row_idx * input_width + im_col_idx] +=
col_data[(c * output_height + h) * output_width + w];
}
Expand Down Expand Up @@ -140,8 +169,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& im, framework::Tensor& col,
int stride_height, int stride_width, int padding_height,
int padding_width) {
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);
int input_channels = im.dims()[0];
Expand All @@ -152,6 +181,21 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
int output_height = col.dims()[0];
int output_width = col.dims()[1];

PADDLE_ENFORCE_EQ(
(input_height + padding_up + padding_down - filter_height) /
stride_height +
1,
output_height,
"Output_height and padding(padding_up, padding_down) are "
"inconsistent.");
PADDLE_ENFORCE_EQ(
(input_width + padding_left + padding_right - filter_width) /
stride_width +
1,
output_width,
"output_width and padding(padding_left, padding_right) are "
"inconsistent.");

const T* im_data = im.data<T>();
T* col_data = col.data<T>();

Expand All @@ -163,10 +207,10 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
for (int filter_col_idx = 0; filter_col_idx < filter_width;
++filter_col_idx) {
int im_row_offset =
col_row_idx * stride_height + filter_row_idx - padding_height;
col_row_idx * stride_height + filter_row_idx - padding_up;
int im_col_offset =
col_col_idx * stride_width + filter_col_idx - padding_width;
int col_offset = (((col_row_idx * output_width + col_col_idx) *
col_col_idx * stride_width + filter_col_idx - padding_left;
int col_offset = ((((col_row_idx)*output_width + col_col_idx) *
input_channels +
channel) *
filter_height +
Expand Down Expand Up @@ -201,7 +245,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
public:
void operator()(const platform::DeviceContext& context, framework::Tensor& im,
const framework::Tensor& col, int stride_height,
int stride_width, int padding_height, int padding_width) {
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);
int input_channels = im.dims()[0];
Expand All @@ -212,6 +257,21 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
int output_height = col.dims()[0];
int output_width = col.dims()[1];

PADDLE_ENFORCE_EQ(
(input_height + padding_up + padding_down - filter_height) /
stride_height +
1,
output_height,
"Output_height and padding(padding_up, padding_down) are "
"inconsistent.");
PADDLE_ENFORCE_EQ(
(input_width + padding_left + padding_right - filter_width) /
stride_width +
1,
output_width,
"output_width and padding(padding_left, padding_right) are "
"inconsistent.");

T* im_data = im.data<T>();
const T* col_data = col.data<T>();

Expand All @@ -223,9 +283,9 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
for (int filter_col_idx = 0; filter_col_idx < filter_width;
++filter_col_idx) {
int im_row_offset =
col_row_idx * stride_height + filter_row_idx - padding_height;
col_row_idx * stride_height + filter_row_idx - padding_up;
int im_col_offset =
col_col_idx * stride_width + filter_col_idx - padding_width;
col_col_idx * stride_width + filter_col_idx - padding_left;
int col_offset = (((col_row_idx * output_width + col_col_idx) *
input_channels +
channel) *
Expand Down
75 changes: 57 additions & 18 deletions paddle/operators/math/im2col.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,8 +66,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& im, framework::Tensor& col,
int stride_height, int stride_width, int padding_height,
int padding_width) {
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);

Expand All @@ -79,6 +79,15 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
int output_height = col.dims()[3];
int output_width = col.dims()[4];

PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
stride_height +
1 ==
output_height);
PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
stride_width +
1 ==
output_width);

int num_outputs = input_channels * output_height * output_width;
int blocks = (num_outputs + 1024 - 1) / 1024;
int block_x = 512;
Expand All @@ -89,8 +98,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(
im.data<T>(), num_outputs, input_height, input_width, filter_height,
filter_width, stride_height, stride_width, padding_height,
padding_width, output_height, output_width, col.data<T>());
filter_width, stride_height, stride_width, padding_up, padding_left,
output_height, output_width, col.data<T>());
}
};

Expand Down Expand Up @@ -152,7 +161,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
public:
void operator()(const platform::DeviceContext& context, framework::Tensor& im,
const framework::Tensor& col, int stride_height,
int stride_width, int padding_height, int padding_width) {
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);

Expand All @@ -164,8 +174,18 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
int output_height = col.dims()[3];
int output_width = col.dims()[4];

size_t num_kernels = input_channels * (input_height + 2 * padding_height) *
(input_width + 2 * padding_width);
PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
stride_height +
1 ==
output_height);
PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
stride_width +
1 ==
output_width);

size_t num_kernels = input_channels *
(input_height + padding_up + padding_down) *
(input_width + padding_left + padding_right);

size_t blocks = (num_kernels + 1024 - 1) / 1024;
size_t block_x = 512;
Expand All @@ -178,10 +198,10 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
col2im<T><<<grid, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(
num_kernels, col.data<T>(), input_height + 2 * padding_height,
input_width + 2 * padding_width, input_channels, filter_height,
filter_width, stride_height, stride_width, padding_height,
padding_width, output_height, output_width, im.data<T>());
num_kernels, col.data<T>(), input_height + padding_up + padding_down,
input_width + padding_left + padding_left, input_channels,
filter_height, filter_width, stride_height, stride_width, padding_up,
padding_left, output_height, output_width, im.data<T>());
}
};

Expand Down Expand Up @@ -238,8 +258,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& im, framework::Tensor& col,
int stride_height, int stride_width, int padding_height,
int padding_width) {
int stride_height, int stride_width, int padding_up,
int padding_down, int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);
int input_channels = im.dims()[0];
Expand All @@ -250,6 +270,15 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
int output_height = col.dims()[0];
int output_width = col.dims()[1];

PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
stride_height +
1 ==
output_height);
PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
stride_width +
1 ==
output_width);

int block_dim_x = 0;
int block_dim_y = 0;
if (filter_height <= 4 && filter_width <= 4) {
Expand All @@ -274,8 +303,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(
im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
filter_height, filter_width, stride_height, stride_width,
padding_height, padding_width, output_height, output_width);
filter_height, filter_width, stride_height, stride_width, padding_up,
padding_left, output_height, output_width);
}
};

Expand Down Expand Up @@ -322,7 +351,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
public:
void operator()(const platform::DeviceContext& context, framework::Tensor& im,
const framework::Tensor& col, int stride_height,
int stride_width, int padding_height, int padding_width) {
int stride_width, int padding_up, int padding_down,
int padding_left, int padding_right) {
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col.dims().size() == 5);
int input_channels = im.dims()[0];
Expand All @@ -333,6 +363,15 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
int output_height = col.dims()[0];
int output_width = col.dims()[1];

PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
stride_height +
1 ==
output_height);
PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
stride_width +
1 ==
output_width);

int block_dim_x = 0;
int block_dim_y = 0;
if (filter_height <= 4 && filter_width <= 4) {
Expand All @@ -357,8 +396,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(
im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
filter_height, filter_width, stride_height, stride_width,
padding_height, padding_width, output_height, output_width);
filter_height, filter_width, stride_height, stride_width, padding_up,
padding_left, output_height, output_width);
}
};

Expand Down
Loading