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

fix some inappropriate expressions in api doc for grid_sampler. test=develop

Browse Source
fix_recordio_link
dengkaipeng 7 years ago
parent 593e1b18d7
commit ff6329bd5f
8 changed files with 436 additions and 409 deletions
Show Stats Download Patch File Download Diff File

172
paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc
Unescape View File

@ -22,107 +22,111 @@ using framework::Tensor;
using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
using DataLayout = platform::DataLayout;
using ScopedSpatialTransformerDescriptor =
platform::ScopedSpatialTransformerDescriptor;
platform::ScopedSpatialTransformerDescriptor;
template <typename T>
using CudnnDataType = platform::CudnnDataType<T>;
template <typename T>
class CUDNNGridSampleOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace");
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto handle = dev_ctx.cudnn_handle();
auto* input = ctx.Input<Tensor>("X");
auto* grid = ctx.Input<Tensor>("Grid");
auto* output = ctx.Output<Tensor>("Output");
int n = input->dims()[0];
int c = input->dims()[1];
int h = input->dims()[2];
int w = input->dims()[3];
const int size[4] = {n, c, h, w};
const T* input_data = input->data<T>();
const T* grid_data = grid->data<T>();
T* output_data = output->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
ScopedSpatialTransformerDescriptor st_desc;
cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace");
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto handle = dev_ctx.cudnn_handle();
auto* input = ctx.Input<Tensor>("X");
auto* grid = ctx.Input<Tensor>("Grid");
auto* output = ctx.Output<Tensor>("Output");
int n = input->dims()[0];
int c = input->dims()[1];
int h = input->dims()[2];
int w = input->dims()[3];
const int size[4] = {n, c, h, w};
const T* input_data = input->data<T>();
const T* grid_data = grid->data<T>();
T* output_data = output->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
ScopedSpatialTransformerDescriptor st_desc;
cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
st_desc.descriptor<T>(4, size);
ScopedTensorDescriptor input_desc;
ScopedTensorDescriptor output_desc;
cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input->dims()));
cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(output->dims()));
CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerForward(
handle, cudnn_st_desc, CudnnDataType<T>::kOne(), cudnn_input_desc, input_data,
grid_data, CudnnDataType<T>::kZero(), cudnn_output_desc, output_data));
}
ScopedTensorDescriptor input_desc;
ScopedTensorDescriptor output_desc;
cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input->dims()));
cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(output->dims()));
CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerForward(
handle, cudnn_st_desc, CudnnDataType<T>::kOne(), cudnn_input_desc,
input_data, grid_data, CudnnDataType<T>::kZero(), cudnn_output_desc,
output_data));
}
};
template <typename T>
class CUDNNGridSampleGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace");
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto handle = dev_ctx.cudnn_handle();
auto* input = ctx.Input<Tensor>("X");
auto* grid = ctx.Input<Tensor>("Grid");
auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* grid_grad = ctx.Output<Tensor>(framework::GradVarName("Grid"));
auto output_grad_dims = output_grad->dims();
const int n = output_grad_dims[0];
const int c = output_grad_dims[1];
const int h = output_grad_dims[2];
const int w = output_grad_dims[3];
const int size[4] = {n, c, h, w};
ScopedSpatialTransformerDescriptor st_dest;
cudnnSpatialTransformerDescriptor_t cudnn_st_dest =
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace");
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto handle = dev_ctx.cudnn_handle();
auto* input = ctx.Input<Tensor>("X");
auto* grid = ctx.Input<Tensor>("Grid");
auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* grid_grad = ctx.Output<Tensor>(framework::GradVarName("Grid"));
auto output_grad_dims = output_grad->dims();
const int n = output_grad_dims[0];
const int c = output_grad_dims[1];
const int h = output_grad_dims[2];
const int w = output_grad_dims[3];
const int size[4] = {n, c, h, w};
ScopedSpatialTransformerDescriptor st_dest;
cudnnSpatialTransformerDescriptor_t cudnn_st_dest =
st_dest.descriptor<T>(4, size);
const T* input_data = input->data<T>();
const T* grid_data = grid->data<T>();
const T* output_grad_data = output_grad->data<T>();
T* input_grad_data = input_grad->mutable_data<T>(output_grad_dims, ctx.GetPlace());
T* grid_grad_data = grid_grad->mutable_data<T>({n, h, w, 2}, ctx.GetPlace());
ScopedTensorDescriptor input_desc;
ScopedTensorDescriptor input_grad_desc;
ScopedTensorDescriptor output_grad_desc;
cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input->dims()));
cudnnTensorDescriptor_t cudnn_input_grad_desc = input_grad_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input_grad->dims()));
cudnnTensorDescriptor_t cudnn_output_grad_desc = output_grad_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(output_grad->dims()));
CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerBackward(
handle, cudnn_st_dest, CudnnDataType<T>::kOne(),
cudnn_input_desc, input_data, CudnnDataType<T>::kZero(),
cudnn_input_grad_desc, input_grad_data, CudnnDataType<T>::kOne(),
cudnn_output_grad_desc, output_grad_data, grid_data,
CudnnDataType<T>::kZero(), grid_grad_data));
}
const T* input_data = input->data<T>();
const T* grid_data = grid->data<T>();
const T* output_grad_data = output_grad->data<T>();
T* input_grad_data =
input_grad->mutable_data<T>(output_grad_dims, ctx.GetPlace());
T* grid_grad_data =
grid_grad->mutable_data<T>({n, h, w, 2}, ctx.GetPlace());
ScopedTensorDescriptor input_desc;
ScopedTensorDescriptor input_grad_desc;
ScopedTensorDescriptor output_grad_desc;
cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input->dims()));
cudnnTensorDescriptor_t cudnn_input_grad_desc =
input_grad_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(input_grad->dims()));
cudnnTensorDescriptor_t cudnn_output_grad_desc =
output_grad_desc.descriptor<T>(
DataLayout::kNCHW, framework::vectorize2int(output_grad->dims()));
CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerBackward(
handle, cudnn_st_dest, CudnnDataType<T>::kOne(), cudnn_input_desc,
input_data, CudnnDataType<T>::kZero(), cudnn_input_grad_desc,
input_grad_data, CudnnDataType<T>::kOne(), cudnn_output_grad_desc,
output_grad_data, grid_data, CudnnDataType<T>::kZero(),
grid_grad_data));
}
};
} // namespace operators
} // namespace paddle
namespace plat = paddle::platform;
REGISTER_OP_KERNEL(grid_sampler, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNGridSampleOpKernel<float>,
paddle::operators::CUDNNGridSampleOpKernel<double>);
REGISTER_OP_KERNEL(grid_sampler, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNGridSampleOpKernel<float>,
paddle::operators::CUDNNGridSampleOpKernel<double>);
REGISTER_OP_KERNEL(grid_sampler_grad, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNGridSampleGradOpKernel<float>,
paddle::operators::CUDNNGridSampleGradOpKernel<double>);
paddle::operators::CUDNNGridSampleGradOpKernel<float>,
paddle::operators::CUDNNGridSampleGradOpKernel<double>);

188
paddle/fluid/operators/grid_sampler_op.cc
Unescape View File

@ -24,70 +24,76 @@ namespace operators {
using Tensor = framework::Tensor;
class GridSampleOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of GridSampleOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grid"),
"Input(Grid) of GridSampleOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Output"),
"Output(Output) of GridSampleOp should not be null.");
auto x_dims = ctx->GetInputDim("X");
auto grid_dims = ctx->GetInputDim("Grid");
PADDLE_ENFORCE(x_dims.size() == 4, "Input(X) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims.size() == 4, "Input(Grid) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0], "Input(X) and Input(Grid) dims[0] should be equal.");
PADDLE_ENFORCE_EQ(grid_dims[1], x_dims[2], "Input(X) dims[2] and Input(Grid) dims[1] should be equal.");
PADDLE_ENFORCE_EQ(grid_dims[2], x_dims[3], "Input(X) dims[3] and Input(Grid) dims[2] should be equal.");
ctx->SetOutputDim("Output", x_dims);
ctx->ShareLoD("X", "Output");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of GridSampleOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grid"),
"Input(Grid) of GridSampleOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Output"),
"Output(Output) of GridSampleOp should not be null.");
auto x_dims = ctx->GetInputDim("X");
auto grid_dims = ctx->GetInputDim("Grid");
PADDLE_ENFORCE(x_dims.size() == 4,
"Input(X) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims.size() == 4,
"Input(Grid) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0],
"Input(X) and Input(Grid) dims[0] should be equal.");
PADDLE_ENFORCE_EQ(
grid_dims[1], x_dims[2],
"Input(X) dims[2] and Input(Grid) dims[1] should be equal.");
PADDLE_ENFORCE_EQ(
grid_dims[2], x_dims[3],
"Input(X) dims[3] and Input(Grid) dims[2] should be equal.");
ctx->SetOutputDim("Output", x_dims);
ctx->ShareLoD("X", "Output");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_);
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
framework::DataLayout::kAnyLayout, library_);
}
};
class GridSampleOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput(
"X",
"(Tensor) The input data of GridSampleOp, "
"This is a 4-D tensor with shape of [N, C, H, W]");
AddInput(
"Grid",
"(Tensor) The input grid of GridSampleOp generated by AffineGridOp, "
"This is a 4-D tensor with shape of [N, H, W, 2] is the concatenation "
"of x and y coordinates with shape [N, H, W] in last dimention");
AddOutput(
"Output",
"(Tensor) Output tensor with shape [N, C, H, W]");
AddAttr<bool>(
"use_cudnn",
"(bool, default true) Only used in cudnn kernel, need install cudnn")
.SetDefault(true);
AddComment(R"DOC(
It sample input X by grid gennerate by AffineGridOp. The grid of shape
[N, H, W, 2] is the concatenation of (x, y) coordinates with shape
[N, H, W] each, with x indexing the 4th-D(W) of input feature map and y to
indexng the 3rd-D(H), finally results is the bilinear interpolation value
of 4 nearest corner points.
public:
void Make() override {
AddInput("X",
"(Tensor) The input data of GridSampleOp, "
"This is a 4-D tensor with shape of [N, C, H, W]");
AddInput(
"Grid",
"(Tensor) The input grid of GridSampleOp generated by AffineGridOp, "
"This is a 4-D tensor with shape of [N, H, W, 2] is the concatenation "
"of x and y coordinates with shape [N, H, W] in last dimention");
AddOutput("Output", "(Tensor) Output tensor with shape [N, C, H, W]");
AddAttr<bool>(
"use_cudnn",
"(bool, default true) Only used in cudnn kernel, need install cudnn")
.SetDefault(true);
AddComment(R"DOC(
This operation samples input X by using bilinear interpolation based on
flow field grid, which is usually gennerated by affine_grid. The grid of
shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates
with shape [N, H, W] each, where grid_x is indexing the 4th dimension
(in width dimension) of input data x and grid_y is indexng the 3rd
dimention (in height dimension), finally results is the bilinear
interpolation value of 4 nearest corner points.
Step 1:
Get (x, y) grid coordinates and scale to [0, H-1/W-1].
@ -127,11 +133,11 @@ class GridSampleOpMaker : public framework::OpProtoAndCheckerMaker {
output = wn * d_e * d_s + en * d_w * d_s
+ ws * d_e * d_n + es * d_w * d_n
)DOC");
}
}
};
class GridSampleOpGrad : public framework::OperatorWithKernel {
public:
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
auto input_dims = ctx->GetInputDim("X");
@ -144,43 +150,43 @@ class GridSampleOpGrad : public framework::OperatorWithKernel {
}
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()),
ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_);
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
framework::DataLayout::kAnyLayout, library_);
}
};
class GridSampleGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* op = new framework::OpDesc();
op->SetType("grid_sampler_grad");
op->SetInput("X", Input("X"));
op->SetInput("Grid", Input("Grid"));
op->SetInput(framework::GradVarName("Output"), OutputGrad("Output"));
op->SetAttrMap(Attrs());
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
op->SetOutput(framework::GradVarName("Grid"), InputGrad("Grid"));
return std::unique_ptr<framework::OpDesc>(op);
}
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* op = new framework::OpDesc();
op->SetType("grid_sampler_grad");
op->SetInput("X", Input("X"));
op->SetInput("Grid", Input("Grid"));
op->SetInput(framework::GradVarName("Output"), OutputGrad("Output"));
op->SetAttrMap(Attrs());
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
op->SetOutput(framework::GradVarName("Grid"), InputGrad("Grid"));
return std::unique_ptr<framework::OpDesc>(op);
}
};
} // namespace operators
} // namespace paddle
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(grid_sampler, ops::GridSampleOp, ops::GridSampleOpMaker,

335
paddle/fluid/operators/grid_sampler_op.h
View File

File diff suppressed because it is too large Load Diff

10
paddle/fluid/platform/cudnn_helper.h
Unescape View File

@ -342,7 +342,7 @@ class ScopedPoolingDescriptor {
};
class ScopedSpatialTransformerDescriptor {
public:
public:
ScopedSpatialTransformerDescriptor() {
PADDLE_ENFORCE(dynload::cudnnCreateSpatialTransformerDescriptor(&desc_));
}
@ -354,13 +354,13 @@ class ScopedSpatialTransformerDescriptor {
inline cudnnSpatialTransformerDescriptor_t descriptor(const int nbDims,
const int dimA[]) {
PADDLE_ENFORCE(dynload::cudnnSetSpatialTransformerNdDescriptor(
desc_, CUDNN_SAMPLER_BILINEAR, CudnnDataType<T>::type, nbDims, dimA));
desc_, CUDNN_SAMPLER_BILINEAR, CudnnDataType<T>::type, nbDims, dimA));
return desc_;
}
private:
cudnnSpatialTransformerDescriptor_t desc_;
DISABLE_COPY_AND_ASSIGN(ScopedSpatialTransformerDescriptor);
private:
cudnnSpatialTransformerDescriptor_t desc_;
DISABLE_COPY_AND_ASSIGN(ScopedSpatialTransformerDescriptor);
};
inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) {

90
paddle/fluid/platform/dynload/cudnn.h
Unescape View File

@ -65,51 +65,51 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
* include all needed cudnn functions in HPPL
* different cudnn version has different interfaces
**/
#define CUDNN_DNN_ROUTINE_EACH(__macro) \
__macro(cudnnSetTensor4dDescriptor); \
__macro(cudnnSetTensor4dDescriptorEx); \
__macro(cudnnSetTensorNdDescriptor); \
__macro(cudnnGetTensorNdDescriptor); \
__macro(cudnnGetConvolutionNdForwardOutputDim); \
__macro(cudnnGetConvolutionForwardAlgorithm); \
__macro(cudnnCreateTensorDescriptor); \
__macro(cudnnDestroyTensorDescriptor); \
__macro(cudnnCreateFilterDescriptor); \
__macro(cudnnSetFilter4dDescriptor); \
__macro(cudnnSetFilterNdDescriptor); \
__macro(cudnnGetFilterNdDescriptor); \
__macro(cudnnSetPooling2dDescriptor); \
__macro(cudnnSetPoolingNdDescriptor); \
__macro(cudnnGetPoolingNdDescriptor); \
__macro(cudnnDestroyFilterDescriptor); \
__macro(cudnnCreateConvolutionDescriptor); \
__macro(cudnnCreatePoolingDescriptor); \
__macro(cudnnDestroyPoolingDescriptor); \
__macro(cudnnSetConvolution2dDescriptor); \
__macro(cudnnDestroyConvolutionDescriptor); \
__macro(cudnnSetConvolutionNdDescriptor); \
__macro(cudnnGetConvolutionNdDescriptor); \
__macro(cudnnDeriveBNTensorDescriptor); \
__macro(cudnnCreateSpatialTransformerDescriptor); \
__macro(cudnnSetSpatialTransformerNdDescriptor); \
__macro(cudnnDestroySpatialTransformerDescriptor);\
__macro(cudnnSpatialTfGridGeneratorForward); \
__macro(cudnnSpatialTfGridGeneratorBackward); \
__macro(cudnnSpatialTfSamplerForward); \
__macro(cudnnSpatialTfSamplerBackward); \
__macro(cudnnCreate); \
__macro(cudnnDestroy); \
__macro(cudnnSetStream); \
__macro(cudnnActivationForward); \
__macro(cudnnConvolutionForward); \
__macro(cudnnConvolutionBackwardBias); \
__macro(cudnnGetConvolutionForwardWorkspaceSize); \
__macro(cudnnTransformTensor); \
__macro(cudnnPoolingForward); \
__macro(cudnnPoolingBackward); \
__macro(cudnnSoftmaxBackward); \
__macro(cudnnSoftmaxForward); \
__macro(cudnnGetVersion); \
#define CUDNN_DNN_ROUTINE_EACH(__macro) \
__macro(cudnnSetTensor4dDescriptor); \
__macro(cudnnSetTensor4dDescriptorEx); \
__macro(cudnnSetTensorNdDescriptor); \
__macro(cudnnGetTensorNdDescriptor); \
__macro(cudnnGetConvolutionNdForwardOutputDim); \
__macro(cudnnGetConvolutionForwardAlgorithm); \
__macro(cudnnCreateTensorDescriptor); \
__macro(cudnnDestroyTensorDescriptor); \
__macro(cudnnCreateFilterDescriptor); \
__macro(cudnnSetFilter4dDescriptor); \
__macro(cudnnSetFilterNdDescriptor); \
__macro(cudnnGetFilterNdDescriptor); \
__macro(cudnnSetPooling2dDescriptor); \
__macro(cudnnSetPoolingNdDescriptor); \
__macro(cudnnGetPoolingNdDescriptor); \
__macro(cudnnDestroyFilterDescriptor); \
__macro(cudnnCreateConvolutionDescriptor); \
__macro(cudnnCreatePoolingDescriptor); \
__macro(cudnnDestroyPoolingDescriptor); \
__macro(cudnnSetConvolution2dDescriptor); \
__macro(cudnnDestroyConvolutionDescriptor); \
__macro(cudnnSetConvolutionNdDescriptor); \
__macro(cudnnGetConvolutionNdDescriptor); \
__macro(cudnnDeriveBNTensorDescriptor); \
__macro(cudnnCreateSpatialTransformerDescriptor); \
__macro(cudnnSetSpatialTransformerNdDescriptor); \
__macro(cudnnDestroySpatialTransformerDescriptor); \
__macro(cudnnSpatialTfGridGeneratorForward); \
__macro(cudnnSpatialTfGridGeneratorBackward); \
__macro(cudnnSpatialTfSamplerForward); \
__macro(cudnnSpatialTfSamplerBackward); \
__macro(cudnnCreate); \
__macro(cudnnDestroy); \
__macro(cudnnSetStream); \
__macro(cudnnActivationForward); \
__macro(cudnnConvolutionForward); \
__macro(cudnnConvolutionBackwardBias); \
__macro(cudnnGetConvolutionForwardWorkspaceSize); \
__macro(cudnnTransformTensor); \
__macro(cudnnPoolingForward); \
__macro(cudnnPoolingBackward); \
__macro(cudnnSoftmaxBackward); \
__macro(cudnnSoftmaxForward); \
__macro(cudnnGetVersion); \
__macro(cudnnGetErrorString);
CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)

29
python/paddle/fluid/layers/nn.py
Unescape View File

@ -7586,11 +7586,13 @@ def hash(input, hash_size, num_hash=1, name=None):
@templatedoc()
def grid_sampler(x, grid, name=None):
"""
It sample input X by grid gennerate by AffineGridOp. The grid of shape
[N, H, W, 2] is the concatenation of (x, y) coordinates with shape
[N, H, W] each, with x indexing the 4th-D(W) of input feature map and y to
indexng the 3rd-D(H), finally results is the bilinear interpolation value
of 4 nearest corner points.
This operation samples input X by using bilinear interpolation based on
flow field grid, which is usually gennerated by affine_grid. The grid of
shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates
with shape [N, H, W] each, where grid_x is indexing the 4th dimension
(in width dimension) of input data x and grid_y is indexng the 3rd
dimention (in height dimension), finally results is the bilinear
interpolation value of 4 nearest corner points.
Step 1:
Get (x, y) grid coordinates and scale to [0, H-1/W-1].
@ -7636,7 +7638,16 @@ def grid_sampler(x, grid, name=None):
name (str, default None): The name of this layer.
Returns:
out(Variable): Output data indices by grid from x of shape [N, C, H, W].
out(Variable): Output of shape [N, C, H, W] data samples input X
using bilnear interpolation based on input grid.
Exmples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[3, 10, 32, 32], dtype='float32')
theta = fluid.layers.data(name='theta', shape=[3, 2, 3], dtype='float32')
grid = fluid.layers.affine_grid(input=theta, size=[3, 10, 32, 32]})
out = fluid.layers.grid_sampler(x=x, grid=grid)
"""
helper = LayerHelper("grid_sampler", **locals())
@ -7649,10 +7660,6 @@ def grid_sampler(x, grid, name=None):
out = helper.create_tmp_variable(x.dtype)
ipts = {'X': x, 'Grid': grid}
helper.apppend_op(
type='grid_sampler',
inputs=ipts,
outputs={'Output', out})
helper.apppend_op(type='grid_sampler', inputs=ipts, outputs={'Output', out})
return out

16
python/paddle/fluid/tests/unittests/test_grid_sampler_op.py
Unescape View File

@ -12,7 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
@ -23,11 +22,11 @@ def AffineGrid(theta, size):
h = size[2]
w = size[3]
h_idx = np.repeat(
np.linspace(-1, 1, h)[np.newaxis, :], w, axis=0).T[:, :, np.newaxis]
np.linspace(-1, 1, h)[np.newaxis, :], w, axis=0).T[:, :, np.newaxis]
w_idx = np.repeat(
np.linspace(-1, 1, w)[np.newaxis, :], h, axis=0)[:, :, np.newaxis]
np.linspace(-1, 1, w)[np.newaxis, :], h, axis=0)[:, :, np.newaxis]
grid = np.concatenate(
[w_idx, h_idx, np.ones([h, w, 1])], axis=2) # h * w * 3
[w_idx, h_idx, np.ones([h, w, 1])], axis=2) # h * w * 3
grid = np.repeat(grid[np.newaxis, :], size[0], axis=0) # n * h * w *3
ret = np.zeros([n, h * w, 2])
@ -37,6 +36,7 @@ def AffineGrid(theta, size):
return ret.reshape([n, h, w, 2]).astype("float32")
def getGridPointValue(data, x, y):
data_shape = data.shape
N = data_shape[0]
@ -47,13 +47,15 @@ def getGridPointValue(data, x, y):
for i in range(N):
for j in range(H):
for k in range(W):
if y[i, j, k] < 0 or y[i, j, k] > H - 1 or x[i, j, k] < 0 or x[i, j, k] > W - 1:
if y[i, j, k] < 0 or y[i, j, k] > H - 1 or x[i, j, k] < 0 or x[
i, j, k] > W - 1:
out[i, :, j, k] = 0
else:
out[i, :, j, k] = data[i, :, y[i, j, k], x[i, j, k]]
return out
def GridSampler(data, grid):
dims = data.shape
N = dims[0]
@ -71,7 +73,7 @@ def GridSampler(data, grid):
x0 = np.floor(x).astype('int32')
x1 = x0 + 1
y0 = np.floor(y).astype('int32')
y0 = np.floor(y).astype('int32')
y1 = y0 + 1
wa = np.tile(((x1 - x) * (y1 - y)).reshape((N, 1, H, W)), (1, C, 1, 1))
@ -87,6 +89,7 @@ def GridSampler(data, grid):
out = (wa * va + wb * vb + wc * vc + wd * vd).astype('float32')
return out
class TestGridSamplerOp(OpTest):
def setUp(self):
self.initTestCase()
@ -115,5 +118,6 @@ class TestGridSamplerOp(OpTest):
self.grid_shape = (2, 7, 3, 2)
self.theta_shape = (2, 2, 3)
if __name__ == "__main__":
unittest.main()

5
python/paddle/fluid/tests/unittests/test_layers.py
Unescape View File

@ -868,13 +868,12 @@ class TestBook(unittest.TestCase):
def test_affine_grid_gen(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[2, 5, 7, 3 ], dtype='float32')
grid = layers.data(name='grid', shape=[2, 5, 7, 2], dtype='float32' )
x = layers.data(name='x', shape=[2, 5, 7, 3], dtype='float32')
grid = layers.data(name='grid', shape=[2, 5, 7, 2], dtype='float32')
out = layers.grid_sampler(x, grid)
self.assertIsNotNone(out)
print(str(program))
if __name__ == '__main__':
unittest.main()

Write Preview
Loading…
Cancel
Save