Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 9437ce36 编写于 作者: J Jack Zhou 提交者： GitHub
浏览文件
操作

Error description optimize for math dir 

Error description optimize for math dir
上级 5c1bafbb
隐藏空白更改
内联 并排
Showing with 271 addition and 123 deletion
paddle/fluid/operators/math/math_function.cc
浏览文件 @ 9437ce36
...@@ -128,9 +128,23 @@ struct RowwiseAdd<platform::CPUDeviceContext, T> { ...@@ -128,9 +128,23 @@ struct RowwiseAdd<platform::CPUDeviceContext, T> {
const framework::Tensor& input, const framework::Tensor& input,
const framework::Tensor& vector, framework::Tensor* output) { const framework::Tensor& vector, framework::Tensor* output) {
auto in_dims = input.dims(); auto in_dims = input.dims();
auto out_dims = output->dims();
auto size = input.numel() / in_dims[0]; auto size = input.numel() / in_dims[0];
PADDLE_ENFORCE_EQ(vector.numel(), size); PADDLE_ENFORCE_EQ(
PADDLE_ENFORCE_EQ(output->dims(), in_dims); vector.numel(), size,
platform::errors::InvalidArgument(
"The input vector size"
" should be equal to the size of each row of input tensor."
" Expected vector size=%d, but received %d",
size, vector.numel()));
const char* in_dims_cstr = in_dims.to_str().c_str();
const char* out_dims_cstr = out_dims.to_str().c_str();
PADDLE_ENFORCE_EQ(out_dims, in_dims,
platform::errors::InvalidArgument(
"The output tensor shape should be same as the input"
" tensor shape. Expected output tensor shape: %s,"
" but received %s",
in_dims_cstr, out_dims_cstr));
auto in = framework::EigenMatrix<T>::From(input); auto in = framework::EigenMatrix<T>::From(input);
auto vec = framework::EigenVector<T>::Flatten(vector); auto vec = framework::EigenVector<T>::Flatten(vector);
......
paddle/fluid/operators/math/math_function.cu
浏览文件 @ 9437ce36
...@@ -88,9 +88,24 @@ struct RowwiseAdd<platform::CUDADeviceContext, T> { ...@@ -88,9 +88,24 @@ struct RowwiseAdd<platform::CUDADeviceContext, T> {
const framework::Tensor& input, const framework::Tensor& input,
const framework::Tensor& vector, framework::Tensor* output) { const framework::Tensor& vector, framework::Tensor* output) {
auto in_dims = input.dims(); auto in_dims = input.dims();
auto out_dims = output->dims();
auto size = input.numel() / in_dims[0]; auto size = input.numel() / in_dims[0];
PADDLE_ENFORCE_EQ(vector.numel(), size); PADDLE_ENFORCE_EQ(
PADDLE_ENFORCE_EQ(output->dims(), in_dims); vector.numel(), size,
platform::errors::InvalidArgument(
"The input vector size"
" should be equal to the size of each row of input tensor."
" Expected vector size=%d, but received %d",
size, vector.numel()));
const char* in_dims_cstr = in_dims.to_str().c_str();
const char* out_dims_cstr = out_dims.to_str().c_str();
PADDLE_ENFORCE_EQ(
out_dims, in_dims,
platform::errors::InvalidArgument(
"The output tensor shape should be same as the input tensor"
" shape. Expected output tensor shape: %s,"
" but received %s",
in_dims_cstr, out_dims_cstr));
int blocks = 512; int blocks = 512;
int grids = (input.numel() + blocks - 1) / blocks; int grids = (input.numel() + blocks - 1) / blocks;
RowwiseAddKernel<T><<<grids, blocks, 0, context.stream()>>>( RowwiseAddKernel<T><<<grids, blocks, 0, context.stream()>>>(
...@@ -113,7 +128,12 @@ void ColwiseSum<platform::CUDADeviceContext, double>::operator()( ...@@ -113,7 +128,12 @@ void ColwiseSum<platform::CUDADeviceContext, double>::operator()(
framework::Tensor* vector) { framework::Tensor* vector) {
auto in_dims = input.dims(); auto in_dims = input.dims();
auto size = input.numel() / in_dims[0]; auto size = input.numel() / in_dims[0];
PADDLE_ENFORCE_EQ(vector->numel(), size); PADDLE_ENFORCE_EQ(vector->numel(), size,
platform::errors::InvalidArgument(
"The size of input vector"
" should be equal to the size of input tensor column"
" dimension. Expected vector size=%d, but received %d",
size, vector->numel()));
framework::Tensor one; framework::Tensor one;
one.mutable_data<double>({in_dims[0]}, context.GetPlace()); one.mutable_data<double>({in_dims[0]}, context.GetPlace());
SetConstant<platform::CUDADeviceContext, double> set; SetConstant<platform::CUDADeviceContext, double> set;
...@@ -134,7 +154,12 @@ void RowwiseSum<platform::CUDADeviceContext, double>::operator()( ...@@ -134,7 +154,12 @@ void RowwiseSum<platform::CUDADeviceContext, double>::operator()(
framework::Tensor* vector) { framework::Tensor* vector) {
auto in_dims = input.dims(); auto in_dims = input.dims();
auto size = input.numel() / in_dims[0]; auto size = input.numel() / in_dims[0];
PADDLE_ENFORCE_EQ(vector->numel(), in_dims[0]); PADDLE_ENFORCE_EQ(vector->numel(), in_dims[0],
platform::errors::InvalidArgument(
"The size of input vector"
" should be equal to the size of input tensor row"
" dimension. Expected vector size=%d, but received %d",
in_dims[0], vector->numel()));
framework::Tensor one; framework::Tensor one;
one.mutable_data<double>({size}, context.GetPlace()); one.mutable_data<double>({size}, context.GetPlace());
SetConstant<platform::CUDADeviceContext, double> set; SetConstant<platform::CUDADeviceContext, double> set;
......
paddle/fluid/operators/math/math_function_impl.h
浏览文件 @ 9437ce36
...@@ -59,7 +59,12 @@ void ColwiseSum<DeviceContext, T>::operator()(const DeviceContext& context, ...@@ -59,7 +59,12 @@ void ColwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
framework::Tensor* out) { framework::Tensor* out) {
auto in_dims = input.dims(); auto in_dims = input.dims();
auto size = input.numel() / in_dims[0]; auto size = input.numel() / in_dims[0];
PADDLE_ENFORCE_EQ(out->numel(), size); PADDLE_ENFORCE_EQ(out->numel(), size,
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor column"
" dimension. Expected output size=%d, but received %d",
size, out->numel()));
auto in = framework::EigenMatrix<T>::From(input); auto in = framework::EigenMatrix<T>::From(input);
auto vec = framework::EigenVector<T>::Flatten(*out); auto vec = framework::EigenVector<T>::Flatten(*out);
...@@ -78,7 +83,13 @@ class ColwiseSum<platform::CPUDeviceContext, T> { ...@@ -78,7 +83,13 @@ class ColwiseSum<platform::CPUDeviceContext, T> {
auto& in_dims = input.dims(); auto& in_dims = input.dims();
auto height = in_dims[0]; auto height = in_dims[0];
auto size = in_dims[1]; auto size = in_dims[1];
PADDLE_ENFORCE_EQ(out->numel(), size); PADDLE_ENFORCE_EQ(
out->numel(), size,
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor column"
" dimension. Expected output size=%d, but received %d",
size, out->numel()));
T* out_buf = out->mutable_data<T>(out->place()); T* out_buf = out->mutable_data<T>(out->place());
const T* in_buf = input.data<T>(); const T* in_buf = input.data<T>();
...@@ -100,8 +111,16 @@ void RowwiseMean<DeviceContext, T>::operator()(const DeviceContext& context, ...@@ -100,8 +111,16 @@ void RowwiseMean<DeviceContext, T>::operator()(const DeviceContext& context,
const framework::Tensor& input, const framework::Tensor& input,
framework::Tensor* out) { framework::Tensor* out) {
auto in_dims = input.dims(); auto in_dims = input.dims();
PADDLE_ENFORCE_EQ(in_dims.size(), 2U); PADDLE_ENFORCE_EQ(in_dims.size(), 2U, platform::errors::InvalidArgument(
PADDLE_ENFORCE_EQ(out->numel(), in_dims[0]); "The rank of input tensor "
"should be 2, but received %d",
in_dims.size()));
PADDLE_ENFORCE_EQ(out->numel(), in_dims[0],
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor row"
" dimension. Expected output size=%d, but received %d",
in_dims[0], out->numel()));
auto in = framework::EigenMatrix<T>::From(input); auto in = framework::EigenMatrix<T>::From(input);
auto vec = framework::EigenVector<T>::Flatten(*out); auto vec = framework::EigenVector<T>::Flatten(*out);
...@@ -118,10 +137,19 @@ class RowwiseMean<platform::CPUDeviceContext, T> { ...@@ -118,10 +137,19 @@ class RowwiseMean<platform::CPUDeviceContext, T> {
void operator()(const platform::CPUDeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
const framework::Tensor& input, framework::Tensor* out) { const framework::Tensor& input, framework::Tensor* out) {
auto& in_dims = input.dims(); auto& in_dims = input.dims();
PADDLE_ENFORCE_EQ(in_dims.size(), 2U); PADDLE_ENFORCE_EQ(in_dims.size(), 2U, platform::errors::InvalidArgument(
"The rank of input tensor "
"should be 2, but received %d",
in_dims.size()));
auto height = in_dims[0]; auto height = in_dims[0];
auto size = in_dims[1]; auto size = in_dims[1];
PADDLE_ENFORCE_EQ(out->numel(), height); PADDLE_ENFORCE_EQ(
out->numel(), height,
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor row"
" dimension. Expected output size=%d, but received %d",
height, out->numel()));
auto inv_size = 1.0 / size; auto inv_size = 1.0 / size;
T* out_buf = out->mutable_data<T>(out->place()); T* out_buf = out->mutable_data<T>(out->place());
const T* in_buf = input.data<T>(); const T* in_buf = input.data<T>();
...@@ -141,8 +169,16 @@ void RowwiseSum<DeviceContext, T>::operator()(const DeviceContext& context, ...@@ -141,8 +169,16 @@ void RowwiseSum<DeviceContext, T>::operator()(const DeviceContext& context,
const framework::Tensor& input, const framework::Tensor& input,
framework::Tensor* out) { framework::Tensor* out) {
auto in_dims = input.dims(); auto in_dims = input.dims();
PADDLE_ENFORCE_EQ(in_dims.size(), 2U); PADDLE_ENFORCE_EQ(in_dims.size(), 2U, platform::errors::InvalidArgument(
PADDLE_ENFORCE_EQ(out->numel(), in_dims[0]); "The rank of input tensor "
"should be 2, but received %d",
in_dims.size()));
PADDLE_ENFORCE_EQ(out->numel(), in_dims[0],
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor row"
" dimension. Expected output size=%d, but received %d",
in_dims[0], out->numel()));
auto in = framework::EigenMatrix<T>::From(input); auto in = framework::EigenMatrix<T>::From(input);
auto vec = framework::EigenVector<T>::Flatten(*out); auto vec = framework::EigenVector<T>::Flatten(*out);
...@@ -159,10 +195,19 @@ class RowwiseSum<platform::CPUDeviceContext, T> { ...@@ -159,10 +195,19 @@ class RowwiseSum<platform::CPUDeviceContext, T> {
void operator()(const platform::CPUDeviceContext& context, void operator()(const platform::CPUDeviceContext& context,
const framework::Tensor& input, framework::Tensor* out) { const framework::Tensor& input, framework::Tensor* out) {
auto& in_dims = input.dims(); auto& in_dims = input.dims();
PADDLE_ENFORCE_EQ(in_dims.size(), 2U); PADDLE_ENFORCE_EQ(in_dims.size(), 2U, platform::errors::InvalidArgument(
"The rank of input tensor "
"should be 2, but received %d",
in_dims.size()));
auto height = in_dims[0]; auto height = in_dims[0];
auto size = in_dims[1]; auto size = in_dims[1];
PADDLE_ENFORCE_EQ(out->numel(), height); PADDLE_ENFORCE_EQ(
out->numel(), height,
platform::errors::InvalidArgument(
"The size of output tensor "
"should be equal to the size of input tensor row"
" dimension. Expected output size=%d, but received %d",
height, out->numel()));
T* out_buf = out->mutable_data<T>(out->place()); T* out_buf = out->mutable_data<T>(out->place());
const T* in_buf = input.data<T>(); const T* in_buf = input.data<T>();
......
paddle/fluid/operators/math/math_function_test.cc
浏览文件 @ 9437ce36
...@@ -224,7 +224,11 @@ TEST(math_funciton, set_constant) { ...@@ -224,7 +224,11 @@ TEST(math_funciton, set_constant) {
auto* ctx = new paddle::platform::CPUDeviceContext(); auto* ctx = new paddle::platform::CPUDeviceContext();
paddle::operators::math::set_constant(*ctx, &t, 10); paddle::operators::math::set_constant(*ctx, &t, 10);
for (int64_t i = 0; i < t.numel(); ++i) { for (int64_t i = 0; i < t.numel(); ++i) {
PADDLE_ENFORCE_EQ(10, t.data<int>()[i]); PADDLE_ENFORCE_EQ(10, t.data<int>()[i],
paddle::platform::errors::InvalidArgument(
"Each value of input"
"tensor should be 10, but received %d.",
t.data<int>()[i]));
} }
delete ctx; delete ctx;
} }
......
paddle/fluid/operators/math/math_function_test.cu
浏览文件 @ 9437ce36
...@@ -18,7 +18,12 @@ ...@@ -18,7 +18,12 @@
void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size, void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size,
const std::vector<float>& data) { const std::vector<float>& data) {
PADDLE_ENFORCE_EQ(size, data.size()); PADDLE_ENFORCE_EQ(
size, data.size(),
paddle::platform::errors::InvalidArgument(
"The size of argument data should"
" be equal to the argument size. Expected %d, but received %d.",
size, data.size()));
for (size_t i = 0; i < data.size(); ++i) { for (size_t i = 0; i < data.size(); ++i) {
in_ptr[i] = paddle::platform::float16(data[i]); in_ptr[i] = paddle::platform::float16(data[i]);
} }
......
paddle/fluid/operators/math/padding.h
浏览文件 @ 9437ce36
...@@ -85,8 +85,9 @@ void PaddingFunctor(int rank, const framework::ExecutionContext& context, ...@@ -85,8 +85,9 @@ void PaddingFunctor(int rank, const framework::ExecutionContext& context,
PadFunction<DeviceContext, T, 6>(context, pads, src, pad_value, out); PadFunction<DeviceContext, T, 6>(context, pads, src, pad_value, out);
break; break;
default: default:
PADDLE_THROW( PADDLE_THROW(platform::errors::Unimplemented(
"PadOp only support tensors with no more than 6 dimensions."); "PadOp only support tensors with no more"
" than 6 dimensions currently."));
} }
} }
...@@ -114,8 +115,9 @@ void PaddingGradFunctor(int rank, const framework::ExecutionContext& context, ...@@ -114,8 +115,9 @@ void PaddingGradFunctor(int rank, const framework::ExecutionContext& context,
PadGradFunction<DeviceContext, T, 6>(context, pads, src, out); PadGradFunction<DeviceContext, T, 6>(context, pads, src, out);
break; break;
default: default:
PADDLE_THROW( PADDLE_THROW(platform::errors::Unimplemented(
"PadOp only support tensors with no more than 6 dimensions."); "PadOp only support tensors with no more"
" than 6 dimensions currently."));
} }
} }
......
paddle/fluid/operators/math/sampler.h
浏览文件 @ 9437ce36
...@@ -19,6 +19,8 @@ limitations under the License. */ ...@@ -19,6 +19,8 @@ limitations under the License. */
#include <random> #include <random>
#include <vector> #include <vector>
#include "paddle/fluid/platform/enforce.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
namespace math { namespace math {
...@@ -31,7 +33,10 @@ namespace math { ...@@ -31,7 +33,10 @@ namespace math {
class Sampler { class Sampler {
public: public:
explicit Sampler(int64_t range, unsigned int seed = 0UL) : range_(range) { explicit Sampler(int64_t range, unsigned int seed = 0UL) : range_(range) {
// PADDLE_ENFORCE_GT(range, 0, "Range should be greater than 0."); PADDLE_ENFORCE_GT(range, 0, platform::errors::InvalidArgument(
"Range should be"
" greater than 0, but recevied %d.",
range));
if (seed == 0) { if (seed == 0) {
std::random_device r; std::random_device r;
seed_ = r(); seed_ = r();
......
paddle/fluid/operators/math/vol2col.cc
浏览文件 @ 9437ce36
...@@ -34,10 +34,16 @@ class Vol2ColFunctor<platform::CPUDeviceContext, T> { ...@@ -34,10 +34,16 @@ class Vol2ColFunctor<platform::CPUDeviceContext, T> {
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* col, const std::vector<int>& paddings, framework::Tensor* col,
const DataLayout data_layout) const { const DataLayout data_layout) const {
PADDLE_ENFORCE_EQ(vol.dims().size(), 4, PADDLE_ENFORCE_EQ(
"The dimension of vol should be 4."); vol.dims().size(), 4,
PADDLE_ENFORCE_EQ(col->dims().size(), 7, platform::errors::InvalidArgument("The dimension of"
"The dimension of col should be 7."); " vol should be 4, but received %d.",
vol.dims().size()));
PADDLE_ENFORCE_EQ(
col->dims().size(), 7,
platform::errors::InvalidArgument("The dimension of"
"col should be 7, but received %d.",
col->dims().size()));
int input_channels = int input_channels =
(data_layout != DataLayout::kNHWC ? vol.dims()[0] : vol.dims()[3]); (data_layout != DataLayout::kNHWC ? vol.dims()[0] : vol.dims()[3]);
...@@ -65,27 +71,33 @@ class Vol2ColFunctor<platform::CPUDeviceContext, T> { ...@@ -65,27 +71,33 @@ class Vol2ColFunctor<platform::CPUDeviceContext, T> {
int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2]; int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2];
int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2]; int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2];
PADDLE_ENFORCE_EQ((input_depth + pad_d_forth + pad_d_back - auto input_depth_tmp = (input_depth + pad_d_forth + pad_d_back -
((dilations[0] * (filter_depth - 1) + 1))) / ((dilations[0] * (filter_depth - 1) + 1))) /
strides[0] + strides[0] +
1, 1;
output_depth, PADDLE_ENFORCE_EQ(
"input_depth and output_depth are " input_depth_tmp, output_depth,
"mismatching."); platform::errors::InvalidArgument(
PADDLE_ENFORCE_EQ((input_height + pad_h_up + pad_h_down - "input_depth(%d) and output_depth(%d) are mismatching.",
((dilations[1] * (filter_height - 1) + 1))) / input_depth_tmp, output_depth));
strides[1] + auto input_height_tmp = (input_height + pad_h_up + pad_h_down -
1, ((dilations[1] * (filter_height - 1) + 1))) /
output_height, strides[1] +
"input_height and output_height are " 1;
"mismatching."); PADDLE_ENFORCE_EQ(
PADDLE_ENFORCE_EQ((input_width + pad_w_left + pad_w_right - input_height_tmp, output_height,
((dilations[2] * (filter_width - 1) + 1))) / platform::errors::InvalidArgument(
strides[2] + "input_height(%d) and output_height(%d) are mismatching.",
1, input_height_tmp, output_height));
output_width, auto input_width_tmp = (input_width + pad_w_left + pad_w_right -
"input_width and output_width are " ((dilations[2] * (filter_width - 1) + 1))) /
"mismatching."); strides[2] +
1;
PADDLE_ENFORCE_EQ(
input_width_tmp, output_width,
platform::errors::InvalidArgument(
"input_width(%d) and output_width(%d) are mismatching.",
input_width_tmp, output_width));
const T* vol_data = vol.data<T>(); const T* vol_data = vol.data<T>();
T* col_data = col->data<T>(); T* col_data = col->data<T>();
...@@ -140,10 +152,16 @@ class Col2VolFunctor<platform::CPUDeviceContext, T> { ...@@ -140,10 +152,16 @@ class Col2VolFunctor<platform::CPUDeviceContext, T> {
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* vol, const std::vector<int>& paddings, framework::Tensor* vol,
const DataLayout data_layout) const { const DataLayout data_layout) const {
PADDLE_ENFORCE_EQ(vol->dims().size(), 4, PADDLE_ENFORCE_EQ(
"The dimension of vol should be 4."); vol->dims().size(), 4,
PADDLE_ENFORCE_EQ(col.dims().size(), 7, platform::errors::InvalidArgument("The dimension of vol"
"The dimension of col should be 7."); " should be 4, but received %d.",
vol->dims().size()));
PADDLE_ENFORCE_EQ(
col.dims().size(), 7,
platform::errors::InvalidArgument("The dimension of col"
" should be 7, but received %d.",
col.dims().size()));
int input_channels = int input_channels =
(data_layout != DataLayout::kNHWC ? vol->dims()[0] : vol->dims()[3]); (data_layout != DataLayout::kNHWC ? vol->dims()[0] : vol->dims()[3]);
...@@ -170,27 +188,33 @@ class Col2VolFunctor<platform::CPUDeviceContext, T> { ...@@ -170,27 +188,33 @@ class Col2VolFunctor<platform::CPUDeviceContext, T> {
int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2]; int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2];
int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2]; int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2];
PADDLE_ENFORCE_EQ((input_depth + pad_d_forth + pad_d_back - auto input_depth_tmp = (input_depth + pad_d_forth + pad_d_back -
((dilations[0] * (filter_depth - 1) + 1))) / ((dilations[0] * (filter_depth - 1) + 1))) /
strides[0] + strides[0] +
1, 1;
output_depth, PADDLE_ENFORCE_EQ(input_depth_tmp, output_depth,
"input_depth and output_depth are " platform::errors::InvalidArgument(
"mismatching."); "input_depth(%d)"
PADDLE_ENFORCE_EQ((input_height + pad_h_up + pad_h_down - " and output_depth(%d) are mismatching.",
((dilations[1] * (filter_height - 1) + 1))) / input_depth_tmp, output_depth));
strides[1] + auto input_height_tmp = (input_height + pad_h_up + pad_h_down -
1, ((dilations[1] * (filter_height - 1) + 1))) /
output_height, strides[1] +
"input_height and output_height are " 1;
"mismatching."); PADDLE_ENFORCE_EQ(input_height_tmp, output_height,
PADDLE_ENFORCE_EQ((input_width + pad_w_left + pad_w_right - platform::errors::InvalidArgument(
((dilations[2] * (filter_width - 1) + 1))) / "input_height(%d)"
strides[2] + " and output_height(%d) are mismatching.",
1, input_height_tmp, output_height));
output_width, auto input_width_tmp = (input_width + pad_w_left + pad_w_right -
"input_width and output_width are " ((dilations[2] * (filter_width - 1) + 1))) /
"mismatching."); strides[2] +
1;
PADDLE_ENFORCE_EQ(input_width_tmp, output_width,
platform::errors::InvalidArgument(
"input_width(%d)"
" and output_width(%d) are mismatching.",
input_width_tmp, output_width));
T* vol_data = vol->data<T>(); T* vol_data = vol->data<T>();
const T* col_data = col.data<T>(); const T* col_data = col.data<T>();
......
paddle/fluid/operators/math/vol2col.cu
浏览文件 @ 9437ce36
...@@ -90,10 +90,16 @@ class Vol2ColFunctor<platform::CUDADeviceContext, T> { ...@@ -90,10 +90,16 @@ class Vol2ColFunctor<platform::CUDADeviceContext, T> {
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* col, const std::vector<int>& paddings, framework::Tensor* col,
const DataLayout data_layout) const { const DataLayout data_layout) const {
PADDLE_ENFORCE_EQ(vol.dims().size(), 4, PADDLE_ENFORCE_EQ(
"The dimension of vol should be 4."); vol.dims().size(), 4,
PADDLE_ENFORCE_EQ(col->dims().size(), 7, platform::errors::InvalidArgument("The dimension of"
"The dimension of col should be 7."); " vol should be 4, but received %d.",
vol.dims().size()));
PADDLE_ENFORCE_EQ(
col->dims().size(), 7,
platform::errors::InvalidArgument("The dimension of"
"col should be 7, but received %d.",
col->dims().size()));
int input_channels = int input_channels =
(data_layout != DataLayout::kNHWC ? vol.dims()[0] : vol.dims()[3]); (data_layout != DataLayout::kNHWC ? vol.dims()[0] : vol.dims()[3]);
...@@ -117,27 +123,33 @@ class Vol2ColFunctor<platform::CUDADeviceContext, T> { ...@@ -117,27 +123,33 @@ class Vol2ColFunctor<platform::CUDADeviceContext, T> {
int pad_h_down = paddings_size_is_6 ? paddings[3] : paddings[1]; int pad_h_down = paddings_size_is_6 ? paddings[3] : paddings[1];
int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2]; int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2];
int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2]; int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2];
PADDLE_ENFORCE_EQ((input_depth + pad_d_forth + pad_d_back - auto input_depth_tmp = (input_depth + pad_d_forth + pad_d_back -
((dilations[0] * (filter_depth - 1) + 1))) / ((dilations[0] * (filter_depth - 1) + 1))) /
strides[0] + strides[0] +
1, 1;
output_depth, PADDLE_ENFORCE_EQ(
"input_depth and output_depth are " input_depth_tmp, output_depth,
"mismatching."); platform::errors::InvalidArgument(
PADDLE_ENFORCE_EQ((input_height + pad_h_up + pad_h_down - "input_depth(%d) and output_depth(%d) are mismatching.",
((dilations[1] * (filter_height - 1) + 1))) / input_depth_tmp, output_depth));
strides[1] + auto input_height_tmp = (input_height + pad_h_up + pad_h_down -
1, ((dilations[1] * (filter_height - 1) + 1))) /
output_height, strides[1] +
"input_height and output_height are " 1;
"mismatching."); PADDLE_ENFORCE_EQ(
PADDLE_ENFORCE_EQ((input_width + pad_w_left + pad_w_right - input_height_tmp, output_height,
((dilations[2] * (filter_width - 1) + 1))) / platform::errors::InvalidArgument(
strides[2] + "input_height(%d) and output_height(%d) are mismatching.",
1, input_height_tmp, output_height));
output_width, auto input_width_tmp = (input_width + pad_w_left + pad_w_right -
"input_width and output_width are " ((dilations[2] * (filter_width - 1) + 1))) /
"mismatching."); strides[2] +
1;
PADDLE_ENFORCE_EQ(
input_width_tmp, output_width,
platform::errors::InvalidArgument(
"input_width(%d) and output_width(%d) are mismatching.",
input_width_tmp, output_width));
int num_outputs = int num_outputs =
input_channels * output_depth * output_height * output_width; input_channels * output_depth * output_height * output_width;
...@@ -241,10 +253,16 @@ class Col2VolFunctor<platform::CUDADeviceContext, T> { ...@@ -241,10 +253,16 @@ class Col2VolFunctor<platform::CUDADeviceContext, T> {
const std::vector<int>& strides, const std::vector<int>& strides,
const std::vector<int>& paddings, framework::Tensor* vol, const std::vector<int>& paddings, framework::Tensor* vol,
const DataLayout data_layout) const { const DataLayout data_layout) const {
PADDLE_ENFORCE_EQ(vol->dims().size(), 4, PADDLE_ENFORCE_EQ(
"The dimension of vol should be 4."); vol->dims().size(), 4,
PADDLE_ENFORCE_EQ(col.dims().size(), 7, platform::errors::InvalidArgument("The dimension of vol"
"The dimension of col should be 7."); " should be 4, but received %d.",
vol->dims().size()));
PADDLE_ENFORCE_EQ(
col.dims().size(), 7,
platform::errors::InvalidArgument("The dimension of col"
" should be 7, but received %d.",
col.dims().size()));
int input_channels = int input_channels =
(data_layout != DataLayout::kNHWC ? vol->dims()[0] : vol->dims()[3]); (data_layout != DataLayout::kNHWC ? vol->dims()[0] : vol->dims()[3]);
...@@ -269,27 +287,33 @@ class Col2VolFunctor<platform::CUDADeviceContext, T> { ...@@ -269,27 +287,33 @@ class Col2VolFunctor<platform::CUDADeviceContext, T> {
int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2]; int pad_w_left = paddings_size_is_6 ? paddings[4] : paddings[2];
int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2]; int pad_w_right = paddings_size_is_6 ? paddings[5] : paddings[2];
PADDLE_ENFORCE_EQ((input_depth + pad_d_forth + pad_d_back - auto input_depth_tmp = (input_depth + pad_d_forth + pad_d_back -
((dilations[0] * (filter_depth - 1) + 1))) / ((dilations[0] * (filter_depth - 1) + 1))) /
strides[0] + strides[0] +
1, 1;
output_depth, PADDLE_ENFORCE_EQ(input_depth_tmp, output_depth,
"input_depth and output_depth are " platform::errors::InvalidArgument(
"mismatching."); "input_depth(%d)"
PADDLE_ENFORCE_EQ((input_height + pad_h_up + pad_h_down - " and output_depth(%d) are mismatching.",
((dilations[1] * (filter_height - 1) + 1))) / input_depth_tmp, output_depth));
strides[1] + auto input_height_tmp = (input_height + pad_h_up + pad_h_down -
1, ((dilations[1] * (filter_height - 1) + 1))) /
output_height, strides[1] +
"input_height and output_height are " 1;
"mismatching."); PADDLE_ENFORCE_EQ(input_height_tmp, output_height,
PADDLE_ENFORCE_EQ((input_width + pad_w_left + pad_w_right - platform::errors::InvalidArgument(
((dilations[2] * (filter_width - 1) + 1))) / "input_height(%d)"
strides[2] + " and output_height(%d) are mismatching.",
1, input_height_tmp, output_height));
output_width, auto input_width_tmp = (input_width + pad_w_left + pad_w_right -
"input_width and output_width are " ((dilations[2] * (filter_width - 1) + 1))) /
"mismatching."); strides[2] +
1;
PADDLE_ENFORCE_EQ(input_width_tmp, output_width,
platform::errors::InvalidArgument(
"input_width(%d)"
" and output_width(%d) are mismatching.",
input_width_tmp, output_width));
int num_kernels = input_channels * input_depth * input_height * input_width; int num_kernels = input_channels * input_depth * input_height * input_width;
......
Markdown is supported
0% 或 .
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 或