Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
“af79b192077a6485fc90be4112f80acce8fb748b”上不存在“paddle/phi/kernels/cross_kernel.h”
未验证 提交 bcf03273 编写于 作者: L littletomatodonkey 提交者： GitHub
浏览文件
操作

add pad func (#26106) 

* add pad func

* add pad

* test=develop, add pad op and apis

* restore pad2d

* test=develop, fix paddl declare

* fix pad interface

* test=develop, fix pad

* test=develop, add all pad api and cos_sim

* test=develop, remove padding default value

* test=develop, rename var to tensor

* test=develop, add more tests

* test=develop, rename tovar to totensor

* test=develop, fix init

* test=develop, add more test

* test=develop, add more tests
上级 a39bd3ac
隐藏空白更改
内联 并排
Showing with 3390 addition and 15 deletion
paddle/fluid/operators/pad3d_op.cc 0 → 100644
浏览文件 @ bcf03273
/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
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 <algorithm>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
using framework::Tensor;
template <typename T>
void ConstPad3DFuncNCDHW(const T* in_data, T* out_data, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left, const int out_d,
const int out_h, const int out_w, const T value) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
out_data[out_d * out_height * out_width + out_h * out_width + out_w] =
(in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width)
? value
: in_data[in_d * in_height * in_width + in_h * in_width + in_w];
}
template <typename T>
void ConstPad3DFuncNDHWC(const T* in_data, T* out_data, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d, const int out_h,
const int out_w, const T value) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
if (in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width) {
for (int c = 0; c < channels; ++c) {
out_data[out_index + c] = value;
}
} else {
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
out_data[out_index + c] = in_data[in_index + c];
}
}
}
template <typename T>
void ReflectPad3DFuncNCDHW(const T* in_data, T* out_data, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h, const int out_w,
const T value) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = std::max(in_d, -in_d); // reflect by 0
in_d = std::min(in_d, 2 * in_depth - in_d - 2); // reflect by in_depth
in_h = std::max(in_h, -in_h); // reflect by 0
in_h = std::min(in_h, 2 * in_height - in_h - 2); // reflect by in_height
in_w = std::max(in_w, -in_w); // reflect by 0
in_w = std::min(in_w, 2 * in_width - in_w - 2); // reflect by in_width
out_data[out_d * out_height * out_width + out_h * out_width + out_w] =
in_data[in_d * in_height * in_width + in_h * in_width + in_w];
}
template <typename T>
void ReflectPad3DFuncNDHWC(const T* in_data, T* out_data, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d, const int out_h,
const int out_w, const T value) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = std::max(in_d, -in_d);
in_d = std::min(in_d, 2 * in_depth - in_d - 2);
in_h = std::max(in_h, -in_h);
in_h = std::min(in_h, 2 * in_height - in_h - 2);
in_w = std::max(in_w, -in_w);
in_w = std::min(in_w, 2 * in_width - in_w - 2);
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
out_data[out_index + c] = in_data[in_index + c];
}
}
template <typename T>
void ReplicatePad3DFuncNCDHW(const T* in_data, T* out_data, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h, const int out_w,
const T value) {
int in_d = std::min(in_depth - 1, std::max(out_d - pad_front, 0));
int in_h = std::min(in_height - 1, std::max(out_h - pad_top, 0));
int in_w = std::min(in_width - 1, std::max(out_w - pad_left, 0));
out_data[out_d * out_height * out_width + out_h * out_width + out_w] =
in_data[in_d * in_height * in_width + in_h * in_width + in_w];
}
template <typename T>
void ReplicatePad3DFuncNDHWC(const T* in_data, T* out_data, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d,
const int out_h, const int out_w, const T value) {
int in_d = std::min(in_depth - 1, std::max(out_d - pad_front, 0));
int in_h = std::min(in_height - 1, std::max(out_h - pad_top, 0));
int in_w = std::min(in_width - 1, std::max(out_w - pad_left, 0));
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
out_data[out_index + c] = in_data[in_index + c];
}
}
template <typename T>
void CircularPad3DFuncNCDHW(const T* in_data, T* out_data, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h, const int out_w,
const T value) {
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
out_data[out_d * out_height * out_width + out_h * out_width + out_w] =
in_data[in_d * in_height * in_width + in_h * in_width + in_w];
}
template <typename T>
void CircularPad3DFuncNDHWC(const T* in_data, T* out_data, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d,
const int out_h, const int out_w, const T value) {
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
out_data[out_index + c] = in_data[in_index + c];
}
}
template <typename T>
void Pad3DNCDHW(const T* in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height, const int out_width,
const int pad_front, const int pad_top, const int pad_left,
T value, T* out_data,
void (*pad_func)(const T*, T*, const int, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const T)) {
for (int n = 0; n < num; ++n) {
for (int c = 0; c < channels; ++c) {
for (int out_d = 0; out_d < out_depth; ++out_d) {
for (int out_h = 0; out_h < out_height; ++out_h) {
for (int out_w = 0; out_w < out_width; ++out_w) {
pad_func(in_data, out_data, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top,
pad_left, out_d, out_h, out_w, value);
}
}
}
in_data += in_depth * in_height * in_width;
out_data += out_depth * out_height * out_width;
}
}
}
template <typename T>
void Pad3DNDHWC(const T* in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height, const int out_width,
const int pad_front, const int pad_top, const int pad_left,
T value, T* out_data,
void (*pad_func)(const T*, T*, const int, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const int, const T)) {
for (int n = 0; n < num; ++n) {
for (int out_d = 0; out_d < out_depth; ++out_d) {
for (int out_h = 0; out_h < out_height; ++out_h) {
for (int out_w = 0; out_w < out_width; ++out_w) {
pad_func(in_data, out_data, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top,
pad_left, out_d, out_h, out_w, value);
}
}
}
in_data += in_depth * in_height * in_width * channels;
out_data += out_depth * out_height * out_width * channels;
}
}
template <typename T>
void ConstPad3DGradNCDHW(T* d_in_data, const T* d_out_data, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left, const int out_d,
const int out_h, const int out_w) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
if (!(in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width)) {
d_in_data[in_d * in_height * in_width + in_h * in_width + in_w] =
d_out_data[out_d * out_height * out_width + out_h * out_width + out_w];
}
}
template <typename T>
void ConstPad3DGradNDHWC(T* d_in_data, const T* d_out_data, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d, const int out_h,
const int out_w) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
if (!(in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width)) {
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
d_in_data[in_index + c] = d_out_data[out_index + c];
}
}
}
template <typename T>
void ReflectPad3DGradNCDHW(T* d_in_data, const T* d_out_data,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d, const int out_h,
const int out_w) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = std::max(in_d, -in_d); // reflect by 0
in_d = std::min(in_d, 2 * in_depth - in_d - 2); // reflect by in_depth
in_h = std::max(in_h, -in_h); // reflect by 0
in_h = std::min(in_h, 2 * in_height - in_h - 2); // reflect by in_height
in_w = std::max(in_w, -in_w); // reflect by 0
in_w = std::min(in_w, 2 * in_width - in_w - 2); // reflect by in_width
d_in_data[in_d * in_height * in_width + in_h * in_width + in_w] +=
d_out_data[out_d * out_height * out_width + out_h * out_width + out_w];
}
template <typename T>
void ReflectPad3DGradNDHWC(T* d_in_data, const T* d_out_data,
const int channels, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h, const int out_w) {
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = std::max(in_d, -in_d);
in_d = std::min(in_d, 2 * in_depth - in_d - 2);
in_h = std::max(in_h, -in_h);
in_h = std::min(in_h, 2 * in_height - in_h - 2);
in_w = std::max(in_w, -in_w);
in_w = std::min(in_w, 2 * in_width - in_w - 2);
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
d_in_data[in_index + c] += d_out_data[out_index + c];
}
}
template <typename T>
void ReplicatePad3DGradNCDHW(T* d_in_data, const T* d_out_data,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d,
const int out_h, const int out_w) {
int in_d = std::min(in_depth - 1, std::max(out_d - pad_front, 0));
int in_h = std::min(in_height - 1, std::max(out_h - pad_top, 0));
int in_w = std::min(in_width - 1, std::max(out_w - pad_left, 0));
d_in_data[in_d * in_height * in_width + in_h * in_width + in_w] +=
d_out_data[out_d * out_height * out_width + out_h * out_width + out_w];
}
template <typename T>
void ReplicatePad3DGradNDHWC(T* d_in_data, const T* d_out_data,
const int channels, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h,
const int out_w) {
int in_d = std::min(in_depth - 1, std::max(out_d - pad_front, 0));
int in_h = std::min(in_height - 1, std::max(out_h - pad_top, 0));
int in_w = std::min(in_width - 1, std::max(out_w - pad_left, 0));
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
d_in_data[in_index + c] += d_out_data[out_index + c];
}
}
template <typename T>
void CircularPad3DGradNCDHW(T* d_in_data, const T* d_out_data,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const int out_d,
const int out_h, const int out_w) {
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
d_in_data[in_d * in_height * in_width + in_h * in_width + in_w] +=
d_out_data[out_d * out_height * out_width + out_h * out_width + out_w];
}
template <typename T>
void CircularPad3DGradNDHWC(T* d_in_data, const T* d_out_data,
const int channels, const int in_depth,
const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const int out_d, const int out_h, const int out_w) {
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
const int out_index =
(out_d * out_height * out_width + out_h * out_width + out_w) * channels;
const int in_index =
(in_d * in_height * in_width + in_h * in_width + in_w) * channels;
for (int c = 0; c < channels; ++c) {
d_in_data[in_index + c] += d_out_data[out_index + c];
}
}
template <typename T>
void Pad3DGradNCDHW(T* d_in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data,
void (*pad_func)(T*, const T*, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const int)) {
for (int n = 0; n < num; ++n) {
for (int c = 0; c < channels; ++c) {
for (int out_d = 0; out_d < out_depth; ++out_d) {
for (int out_h = 0; out_h < out_height; ++out_h) {
for (int out_w = 0; out_w < out_width; ++out_w) {
pad_func(d_in_data, d_out_data, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top,
pad_left, out_d, out_h, out_w);
}
}
}
d_in_data += in_depth * in_height * in_width;
d_out_data += out_depth * out_height * out_width;
}
}
}
template <typename T>
void Pad3DGradNDHWC(T* d_in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height,
const int out_width, const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data,
void (*pad_func)(T*, const T*, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const int, const int)) {
for (int n = 0; n < num; ++n) {
for (int out_d = 0; out_d < out_depth; ++out_d) {
for (int out_h = 0; out_h < out_height; ++out_h) {
for (int out_w = 0; out_w < out_width; ++out_w) {
pad_func(d_in_data, d_out_data, channels, in_depth, in_height,
in_width, out_depth, out_height, out_width, pad_front,
pad_top, pad_left, out_d, out_h, out_w);
}
}
}
d_in_data += in_depth * in_height * in_width * channels;
d_out_data += out_depth * out_height * out_width * channels;
}
}
static inline std::vector<int> GetPaddings(
const framework::ExecutionContext& context) {
std::vector<int> paddings(6);
auto* paddings_t = context.Input<Tensor>("Paddings");
if (paddings_t) {
auto paddings_data = paddings_t->data<int>();
std::memcpy(paddings.data(), paddings_data, paddings.size() * sizeof(int));
} else {
auto pads = context.Attr<std::vector<int>>("paddings");
std::copy(pads.begin(), pads.end(), paddings.data());
}
return paddings;
}
template <typename T>
class Pad3dCPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
std::vector<int> pads = GetPaddings(context);
auto mode = context.Attr<std::string>("mode");
auto data_format = context.Attr<std::string>("data_format");
T value = static_cast<T>(context.Attr<float>("value"));
auto* x = context.Input<Tensor>("X");
auto in_dims = x->dims();
const T* in_data = x->data<T>();
auto* out = context.Output<Tensor>("Out");
if (data_format == "NCDHW") {
out->Resize({in_dims[0], in_dims[1], in_dims[2] + pads[4] + pads[5],
in_dims[3] + pads[2] + pads[3],
in_dims[4] + pads[0] + pads[1]});
} else {
out->Resize({in_dims[0], in_dims[1] + pads[4] + pads[5],
in_dims[2] + pads[2] + pads[3],
in_dims[3] + pads[0] + pads[1], in_dims[4]});
}
auto out_dims = out->dims();
T* out_data = out->mutable_data<T>(context.GetPlace());
int channels = in_dims[1];
int in_depth = in_dims[2];
int in_height = in_dims[3];
int in_width = in_dims[4];
int out_depth = out_dims[2];
int out_height = out_dims[3];
int out_width = out_dims[4];
if (data_format == "NDHWC") {
channels = in_dims[4];
in_depth = in_dims[1];
in_height = in_dims[2];
in_width = in_dims[3];
out_depth = out_dims[1];
out_height = out_dims[2];
out_width = out_dims[3];
}
if (mode == "reflect") {
PADDLE_ENFORCE_GT(in_depth, pads[4],
platform::errors::InvalidArgument(
"The depth of Input(X)'s dimension should be "
"greater than pad_front"
" in reflect mode"
", but received depth(%d) and pad_front(%d).",
in_depth, pads[4]));
PADDLE_ENFORCE_GT(in_depth, pads[5],
platform::errors::InvalidArgument(
"The depth of Input(X)'s dimension should be "
"greater than pad_back"
" in reflect mode"
", but received depth(%d) and pad_back(%d).",
in_depth, pads[5]));
PADDLE_ENFORCE_GT(in_height, pads[2],
platform::errors::InvalidArgument(
"The height of Input(X)'s dimension should be "
"greater than pad_top"
" in reflect mode"
", but received depth(%d) and pad_top(%d).",
in_height, pads[2]));
PADDLE_ENFORCE_GT(in_height, pads[3],
platform::errors::InvalidArgument(
"The height of Input(X)'s dimension should be "
"greater than pad_bottom"
" in reflect mode"
", but received depth(%d) and pad_bottom(%d).",
in_height, pads[3]));
PADDLE_ENFORCE_GT(in_width, pads[0],
platform::errors::InvalidArgument(
"The width of Input(X)'s dimension should be "
"greater than pad_left"
" in reflect mode"
", but received depth(%d) and pad_left(%d).",
in_width, pads[0]));
PADDLE_ENFORCE_GT(in_width, pads[1],
platform::errors::InvalidArgument(
"The width of Input(X)'s dimension should be "
"greater than pad_right"
" in reflect mode"
", but received depth(%d) and pad_right(%d).",
in_width, pads[1]));
}
const int pad_left = pads[0];
const int pad_top = pads[2];
const int pad_front = pads[4];
const int num = in_dims[0];
if (data_format == "NCDHW") {
std::map<std::string,
void (*)(const T*, T*, const int, const int, const int,
const int, const int, const int, const int, const int,
const int, const int, const int, const int, const T)>
func_map;
func_map["reflect"] = ReflectPad3DFuncNCDHW;
func_map["replicate"] = ReplicatePad3DFuncNCDHW;
func_map["circular"] = CircularPad3DFuncNCDHW;
func_map["constant"] = ConstPad3DFuncNCDHW;
Pad3DNCDHW(in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
value, out_data, func_map[mode]);
} else {
std::map<std::string, void (*)(const T*, T*, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const int, const int, const T)>
func_map;
func_map["reflect"] = ReflectPad3DFuncNDHWC;
func_map["replicate"] = ReplicatePad3DFuncNDHWC;
func_map["circular"] = CircularPad3DFuncNDHWC;
func_map["constant"] = ConstPad3DFuncNDHWC;
Pad3DNDHWC(in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
value, out_data, func_map[mode]);
}
}
};
template <typename T>
class Pad3dGradCPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
std::vector<int> pads = GetPaddings(context);
auto mode = context.Attr<std::string>("mode");
auto data_format = context.Attr<std::string>("data_format");
auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
auto* d_in = context.Output<Tensor>(framework::GradVarName("X"));
auto d_in_dims = d_in->dims();
auto d_out_dims = d_out->dims();
const T* d_out_data = d_out->data<T>();
T* d_in_data = d_in->mutable_data<T>(context.GetPlace());
math::SetConstant<platform::CPUDeviceContext, T> set_zero;
set_zero(context.template device_context<platform::CPUDeviceContext>(),
d_in, static_cast<T>(0));
const int pad_left = pads[0];
const int pad_top = pads[2];
const int pad_front = pads[4];
const int num = d_in_dims[0];
if (data_format == "NCDHW") {
const int channels = d_in_dims[1];
const int in_depth = d_in_dims[2];
const int in_height = d_in_dims[3];
const int in_width = d_in_dims[4];
const int out_depth = d_out_dims[2];
const int out_height = d_out_dims[3];
const int out_width = d_out_dims[4];
std::map<std::string,
void (*)(T*, const T*, const int, const int, const int,
const int, const int, const int, const int, const int,
const int, const int, const int, const int)>
func_map;
func_map["reflect"] = ReflectPad3DGradNCDHW;
func_map["replicate"] = ReplicatePad3DGradNCDHW;
func_map["circular"] = CircularPad3DGradNCDHW;
func_map["constant"] = ConstPad3DGradNCDHW;
Pad3DGradNCDHW(d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top,
pad_left, d_out_data, func_map[mode]);
} else {
const int channels = d_in_dims[4];
const int in_depth = d_in_dims[1];
const int in_height = d_in_dims[2];
const int in_width = d_in_dims[3];
const int out_depth = d_out_dims[1];
const int out_height = d_out_dims[2];
const int out_width = d_out_dims[3];
std::map<std::string,
void (*)(T*, const T*, const int, const int, const int,
const int, const int, const int, const int, const int,
const int, const int, const int, const int, const int)>
func_map;
func_map["reflect"] = ReflectPad3DGradNDHWC;
func_map["replicate"] = ReplicatePad3DGradNDHWC;
func_map["circular"] = CircularPad3DGradNDHWC;
func_map["constant"] = ConstPad3DGradNDHWC;
Pad3DGradNDHWC(d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top,
pad_left, d_out_data, func_map[mode]);
}
}
};
class Pad3dOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Pad3d");
OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "Pad3d");
auto x_dim = ctx->GetInputDim("X");
PADDLE_ENFORCE_EQ(x_dim.size(), 5,
platform::errors::InvalidArgument(
"The size of Input(X)'s dimension should be equal to "
"5, but received %d. ",
x_dim.size()));
std::vector<int64_t> out_dims(x_dim.size());
auto data_format = ctx->Attrs().Get<std::string>("data_format");
out_dims[0] = x_dim[0];
if (ctx->HasInput("Paddings")) {
auto paddings_dim = ctx->GetInputDim("Paddings");
PADDLE_ENFORCE_EQ(paddings_dim.size(), 1,
platform::errors::InvalidArgument(
"Size of Input(Paddings)'s dimension should be "
"equal to 1, but received %d.",
paddings_dim.size()));
if (ctx->IsRuntime()) {
PADDLE_ENFORCE_EQ(paddings_dim[0], 6,
platform::errors::InvalidArgument(
"Shape of Input(Paddings) should be equal to "
"[6], but received [%d].",
paddings_dim[0]));
}
out_dims[1] = x_dim[1];
out_dims[2] = x_dim[2];
out_dims[3] = x_dim[3];
} else {
auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
PADDLE_ENFORCE_EQ(
paddings.size(), 6,
platform::errors::InvalidArgument(
"Size of paddings should be equal to 4, but received %d.",
static_cast<int>(paddings.size())));
if (data_format == "NCDHW") {
out_dims[1] = x_dim[1]; // channel
out_dims[2] = ((!ctx->IsRuntime()) && (x_dim[2] < 0))
? x_dim[2]
: (x_dim[2] + paddings[4] + paddings[5]); // depth
out_dims[3] = ((!ctx->IsRuntime()) && (x_dim[3] < 0))
? x_dim[3]
: (x_dim[3] + paddings[2] + paddings[3]); // height
out_dims[4] = ((!ctx->IsRuntime()) && (x_dim[4] < 0))
? x_dim[4]
: (x_dim[4] + paddings[0] + paddings[1]); // width
} else { // NDHWC
out_dims[4] = x_dim[4]; // channel
out_dims[1] = ((!ctx->IsRuntime()) && (x_dim[1] < 0))
? x_dim[1]
: (x_dim[1] + paddings[4] + paddings[5]); // depth
out_dims[2] = ((!ctx->IsRuntime()) && (x_dim[2] < 0))
? x_dim[2]
: (x_dim[2] + paddings[2] + paddings[3]); // height
out_dims[3] = ((!ctx->IsRuntime()) && (x_dim[3] < 0))
? x_dim[3]
: (x_dim[3] + paddings[0] + paddings[1]); // width
}
}
ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
ctx->ShareLoD("X", /*->*/ "Out");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
}
};
class Pad3dOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X",
"The input of pad3d op. "
"The input should be a 5-D tensor with formate NCDHW or NDHWC.");
AddOutput("Out",
"The output of pad3d op. "
"A tensor with the same shape as X.");
AddInput("Paddings",
"A 1-D tensor to describe the padding rules."
"paddings=[0, 1, 2, 3, 4, 5] means "
"padding 0 column to left, 1 column to right, "
"2 row to top, 3 row to bottom, 4 depth to front "
"and 5 depth to back. Size of paddings must be 6.")
.AsDispensable();
AddAttr<std::vector<int>>(
"paddings",
"(vector<int>) "
"A list<int> to describe the padding rules."
"paddings=[0, 1, 2, 3, 4, 5] means "
"padding 0 column to left, 1 column to right, "
"2 row to top, 3 row to bottom, 4 depth to front "
"and 5 depth to back. Size of paddings must be 6.");
AddAttr<float>("value",
"(float, default 0.0) "
"The value to fill the padded areas in constant mode.")
.SetDefault(0.0f);
AddAttr<std::string>(
"mode",
"(string, default constant) "
"Four modes: constant(default), reflect, replicate, circular.")
.SetDefault("constant");
AddAttr<std::string>(
"data_format",
"(string, default NCDHW) Only used in "
"An optional string from: \"NDHWC\", \"NCDHW\". "
"Defaults to \"NDHWC\". Specify the data format of the input data.")
.SetDefault("NCDHW");
AddComment(R"DOC(
Pad3d Operator.
Pad 3-d images according to 'paddings' and 'mode'.
If mode is 'reflect', paddings[0] and paddings[1] must be no greater
than width-1. The height and depth dimension have the same condition.
Given that X is a channel of image from input:
X = [[[[[1, 2, 3],
[4, 5, 6]]]]]
Case 0:
paddings = [2, 2, 1, 1, 0, 0],
mode = 'constant'
pad_value = 0
Out = [[[[[0. 0. 0. 0. 0. 0. 0.]
[0. 0. 1. 2. 3. 0. 0.]
[0. 0. 4. 5. 6. 0. 0.]
[0. 0. 0. 0. 0. 0. 0.]]]]]
Case 1:
paddings = [2, 2, 1, 1, 0, 0],
mode = 'reflect'
Out = [[[[[6. 5. 4. 5. 6. 5. 4.]
[3. 2. 1. 2. 3. 2. 1.]
[6. 5. 4. 5. 6. 5. 4.]
[3. 2. 1. 2. 3. 2. 1.]]]]]
Case 2:
paddings = [2, 2, 1, 1, 0, 0],
mode = 'replicate'
Out = [[[[[1. 1. 1. 2. 3. 3. 3.]
[1. 1. 1. 2. 3. 3. 3.]
[4. 4. 4. 5. 6. 6. 6.]
[4. 4. 4. 5. 6. 6. 6.]]]]]
Case 3:
paddings = [2, 2, 1, 1, 0, 0],
mode = 'circular'
Out = [[[[[5. 6. 4. 5. 6. 4. 5.]
[2. 3. 1. 2. 3. 1. 2.]
[5. 6. 4. 5. 6. 4. 5.]
[2. 3. 1. 2. 3. 1. 2.]]]]]
)DOC");
}
};
class Pad3dOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Pad3d@Grad");
OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
framework::GradVarName("Out"), "Pad3d@Grad");
auto x_dims = ctx->GetInputDim("X");
auto x_grad_name = framework::GradVarName("X");
if (ctx->HasOutput(x_grad_name)) {
ctx->SetOutputDim(x_grad_name, x_dims);
}
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
ctx, framework::GradVarName("Out")),
ctx.GetPlace());
}
};
template <typename T>
class Pad3dOpGradMaker : public framework::SingleGradOpMaker<T> {
public:
using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
protected:
void Apply(GradOpPtr<T> bind) const override {
bind->SetInput("X", this->Input("X"));
if (this->HasInput("Paddings")) {
bind->SetInput("Paddings", this->Input("Paddings"));
}
bind->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
bind->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
bind->SetAttrMap(this->Attrs());
bind->SetType("pad3d_grad");
}
};
DECLARE_NO_NEED_BUFFER_VARS_INFERER(Pad3dOpGradNoNeedBufferVarsInferer, "X");
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(pad3d, ops::Pad3dOp, ops::Pad3dOpMaker,
ops::Pad3dOpGradMaker<paddle::framework::OpDesc>,
ops::Pad3dOpGradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(pad3d_grad, ops::Pad3dOpGrad,
ops::Pad3dOpGradNoNeedBufferVarsInferer);
REGISTER_OP_CPU_KERNEL(pad3d, ops::Pad3dCPUKernel<float>,
ops::Pad3dCPUKernel<double>, ops::Pad3dCPUKernel<int>,
ops::Pad3dCPUKernel<int64_t>);
REGISTER_OP_CPU_KERNEL(pad3d_grad, ops::Pad3dGradCPUKernel<float>,
ops::Pad3dGradCPUKernel<double>);
paddle/fluid/operators/pad3d_op.cu 0 → 100644
浏览文件 @ bcf03273
/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
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 <algorithm>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/gpu_info.h"
namespace paddle {
namespace operators {
using platform::PADDLE_CUDA_NUM_THREADS;
using framework::Tensor;
template <typename T>
__global__ void Pad3DConstNCDHW(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T value, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int nc = index / out_width;
const int out_w = index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
out_data[index] =
(in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width)
? value
: in_data[nc * in_depth * in_height * in_width +
in_d * in_height * in_width + in_h * in_width + in_w];
}
}
template <typename T>
__global__ void Pad3DConstNDHWC(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T value, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int n = index / channels;
const int c = index % channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
const int in_d = out_d - pad_front;
const int in_h = out_h - pad_top;
const int in_w = out_w - pad_left;
out_data[index] =
(in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
in_h >= in_height || in_w >= in_width)
? value
: in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c];
}
}
template <typename T>
__global__ void Pad3DReflectNCDHW(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int nc = index / out_width;
const int out_w = index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = max(in_d, -in_d); // reflect by 0
in_d = min(in_d, 2 * in_depth - in_d - 2); // reflect by in_depth
in_h = max(in_h, -in_h); // reflect by 0
in_h = min(in_h, 2 * in_height - in_h - 2); // reflect by in_height
in_w = max(in_w, -in_w); // reflect by 0
in_w = min(in_w, 2 * in_width - in_w - 2); // reflect by in_width
out_data[index] =
in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
in_width +
in_w];
}
}
template <typename T>
__global__ void Pad3DReflectNDHWC(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int n = index / channels;
const int c = index % channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = max(in_d, -in_d);
in_d = min(in_d, 2 * in_depth - in_d - 2);
in_h = max(in_h, -in_h);
in_h = min(in_h, 2 * in_height - in_h - 2);
in_w = max(in_w, -in_w);
in_w = min(in_w, 2 * in_width - in_w - 2);
out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c];
}
}
template <typename T>
__global__ void Pad3DReplicateNCDHW(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int nc = index / out_width;
const int out_w = index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
int in_h = min(in_height - 1, max(out_h - pad_top, 0));
int in_w = min(in_width - 1, max(out_w - pad_left, 0));
out_data[index] =
in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
in_width +
in_w];
}
}
template <typename T>
__global__ void Pad3DReplicateNDHWC(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int n = index / channels;
const int c = index % channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
int in_h = min(in_height - 1, max(out_h - pad_top, 0));
int in_w = min(in_width - 1, max(out_w - pad_left, 0));
out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c];
}
}
template <typename T>
__global__ void Pad3DCircularNCDHW(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int nc = index / out_width;
const int out_w = index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
out_data[index] =
in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
in_width +
in_w];
}
}
template <typename T>
__global__ void Pad3DCircularNDHWC(const int nthreads, const T* in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, T* out_data) {
CUDA_KERNEL_LOOP(index, nthreads) {
int n = index / channels;
const int c = index % channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c];
}
}
template <typename T>
__global__ void Pad3DGradConstNCDHW(const int in_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data) {
CUDA_KERNEL_LOOP(in_index, in_size) {
const int in_w = in_index % in_width;
int nc = in_index / in_width;
const int in_h = nc % in_height;
nc /= in_height;
const int in_d = nc % in_depth;
nc /= in_depth;
const int out_d = in_d + pad_front;
const int out_h = in_h + pad_top;
const int out_w = in_w + pad_left;
d_in_data[in_index] =
d_out_data[nc * out_depth * out_height * out_width +
out_d * out_height * out_width + out_h * out_width + out_w];
}
}
template <typename T>
__global__ void Pad3DGradConstNDHWC(const int in_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data) {
CUDA_KERNEL_LOOP(in_index, in_size) {
const int c = in_index % channels;
int n = in_index / channels;
const int in_w = n % in_width;
n /= in_width;
const int in_h = n % in_height;
n /= in_height;
const int in_d = n % in_depth;
n /= in_depth;
const int out_d = in_d + pad_front;
const int out_h = in_h + pad_top;
const int out_w = in_w + pad_left;
d_in_data[in_index] =
d_out_data[n * out_depth * out_height * out_width * channels +
out_d * out_height * out_width * channels +
out_h * out_width * channels + out_w * channels + c];
}
}
template <typename T>
__global__ void Pad3DGradReflectNCDHW(const int out_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
int nc = out_index / out_width;
const int out_w = out_index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = max(in_d, -in_d);
in_h = max(in_h, -in_h);
in_w = max(in_w, -in_w);
in_d = min(in_d, 2 * in_depth - in_d - 2);
in_h = min(in_h, 2 * in_height - in_h - 2);
in_w = min(in_w, 2 * in_width - in_w - 2);
platform::CudaAtomicAdd(
&d_in_data[nc * in_depth * in_height * in_width +
in_d * in_height * in_width + in_h * in_width + in_w],
d_out_data[out_index]);
}
}
template <typename T>
__global__ void Pad3DGradReflectNDHWC(const int out_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height, const int out_width,
const int pad_front, const int pad_top,
const int pad_left, const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
const int c = out_index % channels;
int n = out_index / channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
int in_d = out_d - pad_front;
int in_h = out_h - pad_top;
int in_w = out_w - pad_left;
in_d = max(in_d, -in_d);
in_h = max(in_h, -in_h);
in_w = max(in_w, -in_w);
in_d = min(in_d, in_depth * 2 - in_d - 2);
in_h = min(in_h, in_height * 2 - in_h - 2);
in_w = min(in_w, in_width * 2 - in_w - 2);
platform::CudaAtomicAdd(
&d_in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c],
d_out_data[out_index]);
}
}
template <typename T>
__global__ void Pad3DGradReplicateNCDHW(
const int out_size, T* d_in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height, const int out_width,
const int pad_front, const int pad_top, const int pad_left,
const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
int nc = out_index / out_width;
const int out_w = out_index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
const int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
const int in_h = min(in_height - 1, max(out_h - pad_top, 0));
const int in_w = min(in_width - 1, max(out_w - pad_left, 0));
platform::CudaAtomicAdd(
&d_in_data[nc * in_depth * in_height * in_width +
in_d * in_height * in_width + in_h * in_width + in_w],
d_out_data[out_index]);
}
}
template <typename T>
__global__ void Pad3DGradReplicateNDHWC(
const int out_size, T* d_in_data, const int num, const int channels,
const int in_depth, const int in_height, const int in_width,
const int out_depth, const int out_height, const int out_width,
const int pad_front, const int pad_top, const int pad_left,
const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
const int c = out_index % channels;
int n = out_index / channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
const int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
const int in_h = min(in_height - 1, max(out_h - pad_top, 0));
const int in_w = min(in_width - 1, max(out_w - pad_left, 0));
platform::CudaAtomicAdd(
&d_in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c],
d_out_data[out_index]);
}
}
template <typename T>
__global__ void Pad3DGradCircularNCDHW(const int out_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
int nc = out_index / out_width;
const int out_w = out_index % out_width;
const int out_h = nc % out_height;
nc /= out_height;
const int out_d = nc % out_depth;
nc /= out_depth;
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
platform::CudaAtomicAdd(
&d_in_data[nc * in_depth * in_height * in_width +
in_d * in_height * in_width + in_h * in_width + in_w],
d_out_data[out_index]);
}
}
template <typename T>
__global__ void Pad3DGradCircularNDHWC(const int out_size, T* d_in_data,
const int num, const int channels,
const int in_depth, const int in_height,
const int in_width, const int out_depth,
const int out_height,
const int out_width, const int pad_front,
const int pad_top, const int pad_left,
const T* d_out_data) {
CUDA_KERNEL_LOOP(out_index, out_size) {
const int c = out_index % channels;
int n = out_index / channels;
const int out_w = n % out_width;
n /= out_width;
const int out_h = n % out_height;
n /= out_height;
const int out_d = n % out_depth;
n /= out_depth;
int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;
platform::CudaAtomicAdd(
&d_in_data[n * in_depth * in_height * in_width * channels +
in_d * in_height * in_width * channels +
in_h * in_width * channels + in_w * channels + c],
d_out_data[out_index]);
}
}
static inline std::vector<int> GetPaddings(
const framework::ExecutionContext& context) {
std::vector<int> paddings(6);
auto* paddings_data = context.Input<Tensor>("Paddings");
if (paddings_data) {
Tensor pads;
framework::TensorCopySync(*paddings_data, platform::CPUPlace(), &pads);
auto pads_data = pads.data<int>();
std::memcpy(paddings.data(), pads_data, paddings.size() * sizeof(int));
} else {
auto pads = context.Attr<std::vector<int>>("paddings");
std::copy(pads.begin(), pads.end(), paddings.data());
}
return paddings;
}
template <typename T>
class Pad3dCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
std::vector<int> pads = GetPaddings(context);
auto mode = context.Attr<std::string>("mode");
auto data_format = context.Attr<std::string>("data_format");
T value = static_cast<T>(context.Attr<float>("value"));
auto* x = context.Input<Tensor>("X");
auto in_dims = x->dims();
const T* in_data = x->data<T>();
auto* out = context.Output<Tensor>("Out");
auto out_dims = out->dims();
if (data_format == "NCDHW") {
out_dims[0] = in_dims[0];
out_dims[1] = in_dims[1];
out_dims[2] = in_dims[2] + pads[4] + pads[5];
out_dims[3] = in_dims[3] + pads[2] + pads[3];
out_dims[4] = in_dims[4] + pads[0] + pads[1];
} else {
out_dims[0] = in_dims[0];
out_dims[1] = in_dims[1] + pads[4] + pads[5];
out_dims[2] = in_dims[2] + pads[2] + pads[3];
out_dims[3] = in_dims[3] + pads[0] + pads[1];
out_dims[4] = in_dims[4];
}
T* out_data = out->mutable_data<T>(out_dims, context.GetPlace());
int channels = in_dims[1];
int in_depth = in_dims[2];
int in_height = in_dims[3];
int in_width = in_dims[4];
int out_depth = out_dims[2];
int out_height = out_dims[3];
int out_width = out_dims[4];
if (data_format == "NDHWC") {
channels = in_dims[4];
in_depth = in_dims[1];
in_height = in_dims[2];
in_width = in_dims[3];
out_depth = out_dims[1];
out_height = out_dims[2];
out_width = out_dims[3];
}
if (mode == "reflect") {
PADDLE_ENFORCE_GT(in_depth, pads[4],
platform::errors::InvalidArgument(
"The depth of Input(X)'s dimension should be "
"greater than pad_front"
" in reflect mode"
", but received depth(%d) and pad_front(%d).",
in_depth, pads[4]));
PADDLE_ENFORCE_GT(in_depth, pads[5],
platform::errors::InvalidArgument(
"The depth of Input(X)'s dimension should be "
"greater than pad_back"
" in reflect mode"
", but received depth(%d) and pad_back(%d).",
in_depth, pads[5]));
PADDLE_ENFORCE_GT(in_height, pads[2],
platform::errors::InvalidArgument(
"The height of Input(X)'s dimension should be "
"greater than pad_top"
" in reflect mode"
", but received depth(%d) and pad_top(%d).",
in_height, pads[2]));
PADDLE_ENFORCE_GT(in_height, pads[3],
platform::errors::InvalidArgument(
"The height of Input(X)'s dimension should be "
"greater than pad_bottom"
" in reflect mode"
", but received depth(%d) and pad_bottom(%d).",
in_height, pads[3]));
PADDLE_ENFORCE_GT(in_width, pads[0],
platform::errors::InvalidArgument(
"The width of Input(X)'s dimension should be "
"greater than pad_left"
" in reflect mode"
", but received depth(%d) and pad_left(%d).",
in_width, pads[0]));
PADDLE_ENFORCE_GT(in_width, pads[1],
platform::errors::InvalidArgument(
"The width of Input(X)'s dimension should be "
"greater than pad_right"
" in reflect mode"
", but received depth(%d) and pad_right(%d).",
in_width, pads[1]));
}
const int pad_left = pads[0];
const int pad_top = pads[2];
const int pad_front = pads[4];
const int num = in_dims[0];
auto stream = context.cuda_device_context().stream();
int block = PADDLE_CUDA_NUM_THREADS;
const int out_size = out->numel();
int grid = (out_size + block - 1) / block;
if (data_format == "NCDHW") {
if (mode == "reflect") {
Pad3DReflectNCDHW<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else if (mode == "replicate") {
Pad3DReplicateNCDHW<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else if (mode == "circular") {
Pad3DCircularNCDHW<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else {
Pad3DConstNCDHW<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
value, out_data);
}
} else {
if (mode == "reflect") {
Pad3DReflectNDHWC<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else if (mode == "replicate") {
Pad3DReplicateNDHWC<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else if (mode == "circular") {
Pad3DCircularNDHWC<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
out_data);
} else {
Pad3DConstNDHWC<T><<<grid, block, 0, stream>>>(
out_size, in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
value, out_data);
}
}
}
};
template <typename T>
class Pad3dGradCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
std::vector<int> pads = GetPaddings(context);
auto mode = context.Attr<std::string>("mode");
auto data_format = context.Attr<std::string>("data_format");
auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
auto* d_in = context.Output<Tensor>(framework::GradVarName("X"));
auto d_in_dims = d_in->dims();
auto d_out_dims = d_out->dims();
const T* d_out_data = d_out->data<T>();
T* d_in_data = d_in->mutable_data<T>(context.GetPlace());
math::SetConstant<platform::CUDADeviceContext, T> set_zero;
set_zero(context.template device_context<platform::CUDADeviceContext>(),
d_in, static_cast<T>(0));
const int pad_left = pads[0];
const int pad_top = pads[2];
const int pad_front = pads[4];
const int num = d_in_dims[0];
auto stream = context.cuda_device_context().stream();
int block = PADDLE_CUDA_NUM_THREADS;
const int out_size = d_out->numel();
const int in_size = d_in->numel();
int grid = (out_size + block - 1) / block;
if (data_format == "NCDHW") {
const int channels = d_in_dims[1];
const int in_depth = d_in_dims[2];
const int in_height = d_in_dims[3];
const int in_width = d_in_dims[4];
const int out_depth = d_out_dims[2];
const int out_height = d_out_dims[3];
const int out_width = d_out_dims[4];
if (mode == "reflect") {
Pad3DGradReflectNCDHW<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else if (mode == "replicate") {
Pad3DGradReplicateNCDHW<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else if (mode == "circular") {
Pad3DGradCircularNCDHW<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else {
grid = (in_size + block - 1) / block;
Pad3DGradConstNCDHW<T><<<grid, block, 0, stream>>>(
in_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
}
} else {
const int channels = d_in_dims[4];
const int in_depth = d_in_dims[1];
const int in_height = d_in_dims[2];
const int in_width = d_in_dims[3];
const int out_depth = d_out_dims[1];
const int out_height = d_out_dims[2];
const int out_width = d_out_dims[3];
if (mode == "reflect") {
Pad3DGradReflectNDHWC<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else if (mode == "replicate") {
Pad3DGradReplicateNDHWC<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else if (mode == "circular") {
Pad3DGradCircularNDHWC<T><<<grid, block, 0, stream>>>(
out_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
} else {
grid = (in_size + block - 1) / block;
Pad3DGradConstNDHWC<T><<<grid, block, 0, stream>>>(
in_size, d_in_data, num, channels, in_depth, in_height, in_width,
out_depth, out_height, out_width, pad_front, pad_top, pad_left,
d_out_data);
}
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(pad3d, ops::Pad3dCUDAKernel<plat::float16>,
ops::Pad3dCUDAKernel<float>,
ops::Pad3dCUDAKernel<double>, ops::Pad3dCUDAKernel<int>,
ops::Pad3dCUDAKernel<int64_t>);
REGISTER_OP_CUDA_KERNEL(pad3d_grad, ops::Pad3dGradCUDAKernel<plat::float16>,
ops::Pad3dGradCUDAKernel<float>,
ops::Pad3dGradCUDAKernel<double>);
python/paddle/fluid/tests/unittests/test_cosine_similarity_api.py 0 → 100644
浏览文件 @ bcf03273
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
import unittest
import numpy as np
from op_test import OpTest
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
import paddle.fluid.core as core
from paddle.fluid import Program, program_guard, Executor, default_main_program
class TestCosineSimilarityAPI(unittest.TestCase):
def setUp(self):
self.places = [paddle.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(paddle.CUDAPlace(0))
def _get_numpy_out(self, x1, x2, dim=1, eps=1e-8):
w12 = np.sum(x1 * x2, axis=dim)
w1 = np.sum(x1 * x1, axis=dim)
w2 = np.sum(x2 * x2, axis=dim)
n12 = np.sqrt(np.clip(w1 * w2, eps * eps, None))
cos_sim = w12 / n12
return cos_sim
def check_static_result(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
shape = [10, 15]
dim = 1
eps = 1e-8
np.random.seed(0)
np_x1 = np.random.rand(*shape).astype(np.float32)
np_x2 = np.random.rand(*shape).astype(np.float32)
x1 = paddle.data(name="x1", shape=shape)
x2 = paddle.data(name="x2", shape=shape)
result = F.cosine_similarity(x1, x2, dim=dim, eps=eps)
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x1": np_x1,
"x2": np_x2},
fetch_list=[result])
np_out = self._get_numpy_out(np_x1, np_x2, dim=dim, eps=eps)
self.assertTrue(np.allclose(fetches[0], np_out))
def test_static(self):
for place in self.places:
self.check_static_result(place=place)
def test_dygraph_1(self):
paddle.disable_static()
shape = [10, 15]
dim = 1
eps = 1e-8
np.random.seed(1)
np_x1 = np.random.rand(*shape).astype(np.float32)
np_x2 = np.random.rand(*shape).astype(np.float32)
np_out = self._get_numpy_out(np_x1, np_x2, dim=dim, eps=eps)
tesnor_x1 = paddle.to_variable(np_x1)
tesnor_x2 = paddle.to_variable(np_x2)
y = F.cosine_similarity(tesnor_x1, tesnor_x2, dim=dim, eps=eps)
self.assertTrue(np.allclose(y.numpy(), np_out))
def test_dygraph_2(self):
paddle.disable_static()
shape = [12, 13]
dim = 0
eps = 1e-6
np.random.seed(1)
np_x1 = np.random.rand(*shape).astype(np.float32)
np_x2 = np.random.rand(*shape).astype(np.float32)
np_out = self._get_numpy_out(np_x1, np_x2, dim=dim, eps=eps)
tesnor_x1 = paddle.to_variable(np_x1)
tesnor_x2 = paddle.to_variable(np_x2)
y = F.cosine_similarity(tesnor_x1, tesnor_x2, dim=dim, eps=eps)
self.assertTrue(np.allclose(y.numpy(), np_out))
def test_dygraph_3(self):
paddle.disable_static()
shape1 = [10, 12, 10]
shape2 = [10, 1, 10]
dim = 2
eps = 1e-6
np.random.seed(1)
np_x1 = np.random.rand(*shape1).astype(np.float32)
np_x2 = np.random.rand(*shape2).astype(np.float32)
np_out = self._get_numpy_out(np_x1, np_x2, dim=dim, eps=eps)
tesnor_x1 = paddle.to_variable(np_x1)
tesnor_x2 = paddle.to_variable(np_x2)
y = F.cosine_similarity(tesnor_x1, tesnor_x2, dim=dim, eps=eps)
self.assertTrue(np.allclose(y.numpy(), np_out))
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_pad3d_op.py 0 → 100644
浏览文件 @ bcf03273
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
import unittest
import numpy as np
from op_test import OpTest
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
import paddle.fluid.core as core
from paddle.fluid import Program, program_guard, Executor, default_main_program
class TestPad3dOp(OpTest):
def setUp(self):
paddle.enable_static()
self.value = 0.0
self.variable_paddings = False
self.initTestCase()
self.op_type = "pad3d"
self.inputs = {'X': np.random.random(self.shape).astype("float64")}
self.attrs = {}
if self.variable_paddings:
self.attrs['paddings'] = []
self.inputs['Paddings'] = np.array(self.paddings).flatten().astype(
"int32")
else:
self.attrs['paddings'] = np.array(self.paddings).flatten().astype(
"int32")
self.attrs['value'] = self.value
self.attrs['mode'] = self.mode
self.attrs['data_format'] = self.data_format
if self.data_format == "NCDHW":
paddings = [
(0, 0),
(0, 0),
(self.paddings[4], self.paddings[5]),
(self.paddings[2], self.paddings[3]),
(self.paddings[0], self.paddings[1]),
]
else:
paddings = [
(0, 0),
(self.paddings[4], self.paddings[5]),
(self.paddings[2], self.paddings[3]),
(self.paddings[0], self.paddings[1]),
(0, 0),
]
if self.mode == "constant":
out = np.pad(self.inputs['X'],
paddings,
mode=self.mode,
constant_values=self.value)
elif self.mode == "reflect":
out = np.pad(self.inputs['X'], paddings, mode=self.mode)
elif self.mode == "replicate":
out = np.pad(self.inputs['X'], paddings, mode="edge")
elif self.mode == "circular":
out = np.pad(self.inputs['X'], paddings, mode="wrap")
self.outputs = {'Out': out}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X'], 'Out')
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [0, 0, 0, 0, 0, 0]
self.mode = "constant"
self.data_format = "NCDHW"
self.pad_value = 0.0
class TestCase1(TestPad3dOp):
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [0, 1, 2, 3, 4, 5]
self.mode = "constant"
self.data_format = "NCDHW"
self.value = 1.0
class TestCase2(TestPad3dOp):
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [1, 1, 1, 1, 1, 1]
self.mode = "constant"
self.data_format = "NDHWC"
self.value = 1.0
class TestCase3(TestPad3dOp):
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [0, 1, 1, 0, 2, 3]
self.mode = "reflect"
self.data_format = "NCDHW"
class TestCase4(TestPad3dOp):
def initTestCase(self):
self.shape = (4, 4, 4, 4, 4)
self.paddings = [0, 1, 2, 1, 2, 3]
self.mode = "reflect"
self.data_format = "NDHWC"
class TestCase5(TestPad3dOp):
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [0, 1, 2, 3, 2, 1]
self.mode = "replicate"
self.data_format = "NCDHW"
class TestCase6(TestPad3dOp):
def initTestCase(self):
self.shape = (4, 4, 4, 4, 4)
self.paddings = [5, 4, 2, 1, 2, 3]
self.mode = "replicate"
self.data_format = "NDHWC"
class TestCase7(TestPad3dOp):
def initTestCase(self):
self.shape = (2, 3, 4, 5, 6)
self.paddings = [0, 1, 2, 3, 2, 1]
self.mode = "circular"
self.data_format = "NCDHW"
class TestCase8(TestPad3dOp):
def initTestCase(self):
self.shape = (4, 4, 4, 4, 4)
self.paddings = [0, 1, 2, 1, 2, 3]
self.mode = "circular"
self.data_format = "NDHWC"
class TestPadAPI(unittest.TestCase):
def setUp(self):
self.places = [paddle.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(paddle.CUDAPlace(0))
def check_static_result_1(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
input_shape = (1, 2, 3, 4, 5)
pad = [1, 2, 1, 1, 3, 4]
mode = "constant"
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
result = F.pad(x=x, pad=pad, value=value, mode=mode)
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x": input_data},
fetch_list=[result])
np_out = self._get_numpy_out(input_data, pad, mode, value)
self.assertTrue(np.allclose(fetches[0], np_out))
def check_static_result_2(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
input_shape = (2, 3, 4, 5, 6)
pad = [1, 2, 1, 1, 1, 2]
mode = "reflect"
input_data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
result1 = F.pad(x=x, pad=pad, mode=mode, data_format="NCDHW")
result2 = F.pad(x=x, pad=pad, mode=mode, data_format="NDHWC")
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x": input_data},
fetch_list=[result1, result2])
np_out1 = self._get_numpy_out(
input_data, pad, mode, data_format="NCDHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, data_format="NDHWC")
self.assertTrue(np.allclose(fetches[0], np_out1))
self.assertTrue(np.allclose(fetches[1], np_out2))
def check_static_result_3(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
input_shape = (2, 3, 4, 5, 6)
pad = [1, 2, 1, 1, 3, 4]
mode = "replicate"
input_data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
result1 = F.pad(x=x, pad=pad, mode=mode, data_format="NCDHW")
result2 = F.pad(x=x, pad=pad, mode=mode, data_format="NDHWC")
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x": input_data},
fetch_list=[result1, result2])
np_out1 = self._get_numpy_out(
input_data, pad, mode, data_format="NCDHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, data_format="NDHWC")
self.assertTrue(np.allclose(fetches[0], np_out1))
self.assertTrue(np.allclose(fetches[1], np_out2))
def check_static_result_4(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
input_shape = (2, 3, 4, 5, 6)
pad = [1, 2, 1, 1, 3, 4]
mode = "circular"
input_data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
result1 = F.pad(x=x, pad=pad, mode=mode, data_format="NCDHW")
result2 = F.pad(x=x, pad=pad, mode=mode, data_format="NDHWC")
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x": input_data},
fetch_list=[result1, result2])
np_out1 = self._get_numpy_out(
input_data, pad, mode, data_format="NCDHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, data_format="NDHWC")
self.assertTrue(np.allclose(fetches[0], np_out1))
self.assertTrue(np.allclose(fetches[1], np_out2))
def _get_numpy_out(self,
input_data,
pad,
mode,
value=0,
data_format="NCDHW"):
if data_format == "NCDHW":
pad = [
(0, 0),
(0, 0),
(pad[4], pad[5]),
(pad[2], pad[3]),
(pad[0], pad[1]),
]
elif data_format == "NDHWC":
pad = [
(0, 0),
(pad[4], pad[5]),
(pad[2], pad[3]),
(pad[0], pad[1]),
(0, 0),
]
elif data_format == "NCHW":
pad = [
(0, 0),
(0, 0),
(pad[2], pad[3]),
(pad[0], pad[1]),
]
elif data_format == "NHWC":
pad = [
(0, 0),
(pad[2], pad[3]),
(pad[0], pad[1]),
(0, 0),
]
elif data_format == "NCL":
pad = [
(0, 0),
(0, 0),
(pad[0], pad[1]),
]
elif data_format == "NLC":
pad = [
(0, 0),
(pad[0], pad[1]),
(0, 0),
]
if mode == "constant":
out = np.pad(input_data, pad, mode=mode, constant_values=value)
elif mode == "reflect":
out = np.pad(input_data, pad, mode=mode)
elif mode == "replicate":
out = np.pad(input_data, pad, mode="edge")
elif mode == "circular":
out = np.pad(input_data, pad, mode="wrap")
return out
def test_static(self):
for place in self.places:
self.check_static_result_1(place=place)
self.check_static_result_2(place=place)
self.check_static_result_3(place=place)
self.check_static_result_4(place=place)
def test_dygraph_1(self):
paddle.disable_static()
input_shape = (1, 2, 3, 4, 5)
pad = [1, 2, 1, 1, 3, 4]
mode = "constant"
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
np_out1 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NCDHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NDHWC")
tensor_data = paddle.to_tensor(input_data)
y1 = F.pad(tensor_data,
pad=pad,
mode=mode,
value=value,
data_format="NCDHW")
y2 = F.pad(tensor_data,
pad=pad,
mode=mode,
value=value,
data_format="NDHWC")
self.assertTrue(np.allclose(y1.numpy(), np_out1))
self.assertTrue(np.allclose(y2.numpy(), np_out2))
def test_dygraph_2(self):
paddle.disable_static()
input_shape = (2, 3, 4, 5)
pad = [1, 1, 3, 4]
mode = "constant"
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
np_out1 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NCHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NHWC")
tensor_data = paddle.to_tensor(input_data)
tensor_pad = paddle.to_tensor(pad, dtype="int32")
y1 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NCHW")
y2 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NHWC")
self.assertTrue(np.allclose(y1.numpy(), np_out1))
self.assertTrue(np.allclose(y2.numpy(), np_out2))
def test_dygraph_2(self):
paddle.disable_static()
input_shape = (2, 3, 4, 5)
pad = [1, 1, 3, 4]
mode = "constant"
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
np_out1 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NCHW")
np_out2 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NHWC")
tensor_data = paddle.to_tensor(input_data)
tensor_pad = paddle.to_tensor(pad, dtype="int32")
y1 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NCHW")
y2 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NHWC")
self.assertTrue(np.allclose(y1.numpy(), np_out1))
self.assertTrue(np.allclose(y2.numpy(), np_out2))
def test_dygraph_3(self):
paddle.disable_static()
input_shape = (3, 4, 5)
pad = [3, 4]
mode = "constant"
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
np_out1 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NCL")
np_out2 = self._get_numpy_out(
input_data, pad, mode, value, data_format="NLC")
tensor_data = paddle.to_tensor(input_data)
tensor_pad = paddle.to_tensor(pad, dtype="int32")
y1 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NCL")
y2 = F.pad(tensor_data,
pad=tensor_pad,
mode=mode,
value=value,
data_format="NLC")
self.assertTrue(np.allclose(y1.numpy(), np_out1))
self.assertTrue(np.allclose(y2.numpy(), np_out2))
class TestPad1dAPI(unittest.TestCase):
def _get_numpy_out(self,
input_data,
pad,
mode,
value=0.0,
data_format="NCL"):
if data_format == "NCL":
pad = [
(0, 0),
(0, 0),
(pad[0], pad[1]),
]
else:
pad = [
(0, 0),
(pad[0], pad[1]),
(0, 0),
]
if mode == "constant":
out = np.pad(input_data, pad, mode=mode, constant_values=value)
elif mode == "reflect":
out = np.pad(input_data, pad, mode=mode)
elif mode == "replicate":
out = np.pad(input_data, pad, mode="edge")
return out
def setUp(self):
self.places = [paddle.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(paddle.CUDAPlace(0))
def test_class(self):
paddle.disable_static()
for place in self.places:
input_shape = (3, 4, 5)
pad = [1, 2]
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
pad_reflection = nn.ReflectionPad1d(padding=pad)
pad_replication = nn.ReplicationPad1d(padding=pad)
pad_constant = nn.ConstantPad1d(padding=pad, value=value)
data = paddle.to_tensor(input_data)
output = pad_reflection(data)
np_out = self._get_numpy_out(
input_data, pad, "reflect", data_format="NCL")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_replication(data)
np_out = self._get_numpy_out(
input_data, pad, "replicate", data_format="NCL")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_constant(data)
np_out = self._get_numpy_out(
input_data, pad, "constant", value=value, data_format="NCL")
self.assertTrue(np.allclose(output.numpy(), np_out))
class TestPad2dAPI(unittest.TestCase):
def _get_numpy_out(self,
input_data,
pad,
mode,
value=0.0,
data_format="NCHW"):
if data_format == "NCHW":
pad = [
(0, 0),
(0, 0),
(pad[2], pad[3]),
(pad[0], pad[1]),
]
else:
pad = [
(0, 0),
(pad[2], pad[3]),
(pad[0], pad[1]),
(0, 0),
]
if mode == "constant":
out = np.pad(input_data, pad, mode=mode, constant_values=value)
elif mode == "reflect":
out = np.pad(input_data, pad, mode=mode)
elif mode == "replicate":
out = np.pad(input_data, pad, mode="edge")
return out
def setUp(self):
self.places = [paddle.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(paddle.CUDAPlace(0))
def test_class(self):
paddle.disable_static()
for place in self.places:
input_shape = (3, 4, 5, 6)
pad = [1, 2, 2, 1]
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
pad_reflection = nn.ReflectionPad2d(padding=pad)
pad_replication = nn.ReplicationPad2d(padding=pad)
pad_constant = nn.ConstantPad2d(padding=pad, value=value)
pad_zero = nn.ZeroPad2d(padding=pad)
data = paddle.to_tensor(input_data)
output = pad_reflection(data)
np_out = self._get_numpy_out(
input_data, pad, "reflect", data_format="NCHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_replication(data)
np_out = self._get_numpy_out(
input_data, pad, "replicate", data_format="NCHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_constant(data)
np_out = self._get_numpy_out(
input_data, pad, "constant", value=value, data_format="NCHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_zero(data)
np_out = self._get_numpy_out(
input_data, pad, "constant", value=0, data_format="NCHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
class TestPad3dAPI(unittest.TestCase):
def _get_numpy_out(self,
input_data,
pad,
mode,
value=0.0,
data_format="NCDHW"):
if data_format == "NCDHW":
pad = [
(0, 0),
(0, 0),
(pad[4], pad[5]),
(pad[2], pad[3]),
(pad[0], pad[1]),
]
else:
pad = [
(0, 0),
(pad[4], pad[5]),
(pad[2], pad[3]),
(pad[0], pad[1]),
(0, 0),
]
if mode == "constant":
out = np.pad(input_data, pad, mode=mode, constant_values=value)
elif mode == "reflect":
out = np.pad(input_data, pad, mode=mode)
elif mode == "replicate":
out = np.pad(input_data, pad, mode="edge")
return out
def setUp(self):
self.places = [paddle.CPUPlace()]
if core.is_compiled_with_cuda():
self.places.append(paddle.CUDAPlace(0))
def test_class(self):
paddle.disable_static()
for place in self.places:
input_shape = (3, 4, 5, 6, 7)
pad = [1, 2, 2, 1, 1, 0]
value = 100
input_data = np.random.rand(*input_shape).astype(np.float32)
pad_replication = nn.ReplicationPad3d(padding=pad)
pad_constant = nn.ConstantPad3d(padding=pad, value=value)
data = paddle.to_tensor(input_data)
output = pad_replication(data)
np_out = self._get_numpy_out(
input_data, pad, "replicate", data_format="NCDHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
output = pad_constant(data)
np_out = self._get_numpy_out(
input_data, pad, "constant", value=value, data_format="NCDHW")
self.assertTrue(np.allclose(output.numpy(), np_out))
class TestPad3dOpError(unittest.TestCase):
def test_errors(self):
def test_variable():
input_shape = (1, 2, 3, 4, 5)
data = np.random.rand(*input_shape).astype(np.float32)
F.pad(x=data, paddings=[1, 1, 1, 1, 1, 1])
def test_reflect_1():
input_shape = (1, 2, 3, 4, 5)
data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
y = F.pad(x, pad=[5, 6, 1, 1, 1, 1], value=1, mode='reflect')
place = paddle.CPUPlace()
exe = Executor(place)
outputs = exe.run(feed={'x': data}, fetch_list=[y.name])
def test_reflect_2():
input_shape = (1, 2, 3, 4, 5)
data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
y = F.pad(x, pad=[1, 1, 4, 3, 1, 1], value=1, mode='reflect')
place = paddle.CPUPlace()
exe = Executor(place)
outputs = exe.run(feed={'x': data}, fetch_list=[y.name])
def test_reflect_3():
input_shape = (1, 2, 3, 4, 5)
data = np.random.rand(*input_shape).astype(np.float32)
x = paddle.data(name="x", shape=input_shape)
y = F.pad(x, pad=[1, 1, 1, 1, 2, 3], value=1, mode='reflect')
place = paddle.CPUPlace()
exe = Executor(place)
outputs = exe.run(feed={'x': data}, fetch_list=[y.name])
self.assertRaises(TypeError, test_variable)
self.assertRaises(Exception, test_reflect_1)
self.assertRaises(Exception, test_reflect_2)
self.assertRaises(Exception, test_reflect_3)
if __name__ == '__main__':
unittest.main()
python/paddle/nn/__init__.py
浏览文件 @ bcf03273
...@@ -65,6 +65,16 @@ from .layer.activation import HSigmoid #DEFINE_ALIAS ...@@ -65,6 +65,16 @@ from .layer.activation import HSigmoid #DEFINE_ALIAS
from .layer.common import BilinearTensorProduct #DEFINE_ALIAS from .layer.common import BilinearTensorProduct #DEFINE_ALIAS
from .layer.common import Pool2D #DEFINE_ALIAS from .layer.common import Pool2D #DEFINE_ALIAS
from .layer.common import Pad2D #DEFINE_ALIAS from .layer.common import Pad2D #DEFINE_ALIAS
from .layer.common import ReflectionPad1d #DEFINE_ALIAS
from .layer.common import ReplicationPad1d #DEFINE_ALIAS
from .layer.common import ConstantPad1d #DEFINE_ALIAS
from .layer.common import ReflectionPad2d #DEFINE_ALIAS
from .layer.common import ReplicationPad2d #DEFINE_ALIAS
from .layer.common import ConstantPad2d #DEFINE_ALIAS
from .layer.common import ZeroPad2d #DEFINE_ALIAS
from .layer.common import ReplicationPad3d #DEFINE_ALIAS
from .layer.common import ConstantPad3d #DEFINE_ALIAS
from .layer.common import CosineSimilarity #DEFINE_ALIAS
from .layer.common import Embedding #DEFINE_ALIAS from .layer.common import Embedding #DEFINE_ALIAS
from .layer.common import Linear #DEFINE_ALIAS from .layer.common import Linear #DEFINE_ALIAS
from .layer.common import Flatten #DEFINE_ALIAS from .layer.common import Flatten #DEFINE_ALIAS
......
python/paddle/nn/functional/__init__.py
浏览文件 @ bcf03273
...@@ -58,6 +58,7 @@ from .common import one_hot #DEFINE_ALIAS ...@@ -58,6 +58,7 @@ from .common import one_hot #DEFINE_ALIAS
from .common import pad #DEFINE_ALIAS from .common import pad #DEFINE_ALIAS
from .common import pad_constant_like #DEFINE_ALIAS from .common import pad_constant_like #DEFINE_ALIAS
from .common import pad2d #DEFINE_ALIAS from .common import pad2d #DEFINE_ALIAS
from .common import cosine_similarity #DEFINE_ALIAS
from .common import unfold #DEFINE_ALIAS from .common import unfold #DEFINE_ALIAS
# from .common import bilinear_tensor_product #DEFINE_ALIAS # from .common import bilinear_tensor_product #DEFINE_ALIAS
from .common import assign #DEFINE_ALIAS from .common import assign #DEFINE_ALIAS
......
python/paddle/nn/functional/common.py
浏览文件 @ bcf03273
...@@ -13,17 +13,24 @@ ...@@ -13,17 +13,24 @@
# limitations under the License. # limitations under the License.
import warnings import warnings
import paddle.fluid.core as core
from ...fluid.framework import in_dygraph_mode, core
from paddle.fluid.layer_helper import LayerHelper from paddle.fluid.layer_helper import LayerHelper
from paddle.fluid.layers.tensor import Variable, fill_constant from paddle.fluid.layers.tensor import Variable, fill_constant, zeros, concat
# TODO: define the common functions to build a neural network # TODO: define the common functions to build a neural network
from ...fluid.layers import dropout #DEFINE_ALIAS from ...fluid.layers import dropout #DEFINE_ALIAS
from ...fluid.layers import label_smooth #DEFINE_ALIAS from ...fluid.layers import label_smooth #DEFINE_ALIAS
from ...fluid import one_hot #DEFINE_ALIAS from ...fluid import one_hot #DEFINE_ALIAS
from ...fluid.layers import pad #DEFINE_ALIAS
from ...fluid.layers import pad2d #DEFINE_ALIAS from ...fluid.layers import pad2d #DEFINE_ALIAS
from ...fluid.layers import unfold #DEFINE_ALIAS from ...fluid.layers import unfold #DEFINE_ALIAS
from ...fluid.layers import assign #DEFINE_ALIAS from ...fluid.layers import assign #DEFINE_ALIAS
from ...fluid.layers import squeeze #DEFINE_ALIAS
from ...fluid.layers import unsqueeze #DEFINE_ALIAS
from ...fluid.layers import elementwise_mul #DEFINE_ALIAS
from ...tensor import clamp #DEFINE_ALIAS
from ...tensor import sum #DEFINE_ALIAS
from ...tensor import sqrt #DEFINE_ALIAS
#from ...fluid.layers import fc #DEFINE_ALIAS #from ...fluid.layers import fc #DEFINE_ALIAS
from ...fluid.layers import pad_constant_like #DEFINE_ALIAS from ...fluid.layers import pad_constant_like #DEFINE_ALIAS
...@@ -40,7 +47,8 @@ __all__ = [ ...@@ -40,7 +47,8 @@ __all__ = [
'unfold', 'unfold',
# 'bilinear_tensor_product', # 'bilinear_tensor_product',
'assign', 'assign',
'interpolate' 'interpolate',
'cosine_similarity',
] ]
...@@ -446,3 +454,235 @@ def interpolate(input, ...@@ -446,3 +454,235 @@ def interpolate(input,
outputs={"Out": out}, outputs={"Out": out},
attrs=attrs) attrs=attrs)
return out return out
def pad(x, pad, mode='constant', value=0, data_format="NCHW", name=None):
"""
Pad tensor according to 'pad' and 'mode'.
If mode is 'reflect', pad[0] and pad[1] must be no greater
than width-1. The height and depth dimension has the same condition.
Parameters:
x (Tensor): The input tensor with data type float32/double/int32/int64_t.
pad (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. 1. If input dimension is 3, then the pad has the form (pad_left,
pad_right). 2. If the input dimension is 4, then the pad has the form (pad_left, pad_right,
pad_top, pad_bottom). 3. If the input dimension is 5, then the pad has the form
(pad_left, pad_right, pad_top, pad_bottom, pad_front, pad_back).
mode (str): Four modes: 'constant' (default), 'reflect', 'replicate', 'circular'.
When in 'constant' mode, this op uses a constant value to pad the input tensor.
When in 'reflect' mode, uses reflection of the input boundaries to pad the input tensor.
When in 'replicate' mode, uses input boundaries to pad the input tensor.
When in 'circular' mode, uses circular input to pad the input tensor.
Default is 'constant'
value (float32): The value to fill the padded areas in 'constant' mode . Default is 0.0
data_format (str): An string from: "NCL", "NLC", NHWC", "NCHW", "NCDHW", "NDHWC". Specify the data format of
the input data.
Default is "NCHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns: a Tensor padded according to pad and mode and data type is same as input.
Return Type: Tensor
Examples:
.. code-block:: text
x = [[[[[1., 2., 3.],
[4., 5., 6.]]]]]
Case 0:
pad = [2, 2, 1, 1, 0, 0],
mode = 'constant'
value = 0
Out = [[[[[0. 0. 0. 0. 0. 0. 0.]
[0. 0. 1. 2. 3. 0. 0.]
[0. 0. 4. 5. 6. 0. 0.]
[0. 0. 0. 0. 0. 0. 0.]]]]]
Case 1:
pad = [2, 2, 1, 1, 0, 0],
mode = 'reflect'
Out = [[[[[6. 5. 4. 5. 6. 5. 4.]
[3. 2. 1. 2. 3. 2. 1.]
[6. 5. 4. 5. 6. 5. 4.]
[3. 2. 1. 2. 3. 2. 1.]]]]]
Case 2:
pad = [2, 2, 1, 1, 0, 0],
mode = 'replicate'
Out = [[[[[1. 1. 1. 2. 3. 3. 3.]
[1. 1. 1. 2. 3. 3. 3.]
[4. 4. 4. 5. 6. 6. 6.]
[4. 4. 4. 5. 6. 6. 6.]]]]]
Case 3:
pad = [2, 2, 1, 1, 0, 0],
mode = 'circular'
Out = [[[[[5. 6. 4. 5. 6. 4. 5.]
[2. 3. 1. 2. 3. 1. 2.]
[5. 6. 4. 5. 6. 4. 5.]
[2. 3. 1. 2. 3. 1. 2.]]]]]
Code Examples:
.. code-block:: python
import numpy as np
import paddle
import paddle.nn.functional as F
paddle.disable_static()
# example 1
x_shape = (1, 1, 3)
x = np.arange(np.prod(x_shape), dtype=np.float32).reshape(x_shape) + 1
tensor_x = paddle.to_tensor(x)
y = F.pad(tensor_x, pad=[2, 3], value=1, mode='constant')
print(y.numpy())
# [[[1. 1. 1. 2. 3. 1. 1. 1.]]]
# example 2
x_shape = (1, 1, 2, 3)
x = np.arange(np.prod(x_shape), dtype=np.float32).reshape(x_shape) + 1
tensor_x = paddle.to_tensor(x)
y = F.pad(tensor_x, pad=[1, 2, 1, 1], value=1, mode='circular')
print(y.numpy())
# [[[[6. 4. 5. 6. 4. 5.]
# [3. 1. 2. 3. 1. 2.]
# [6. 4. 5. 6. 4. 5.]
# [3. 1. 2. 3. 1. 2.]]]]
"""
assert mode in ['reflect', 'replicate', 'constant', 'circular'], \
"mode should be one of constant, reflect, replicate, circular, but got {}.".format(mode)
data_format = data_format.upper()
assert data_format in ["NCL", "NCHW", "NCDHW", "NLC", "NHWC", "NDHWC"], \
"data_format should be in one of [NCL, NCHW, NCDHW, NLC, NHWC, NDHWC], " \
"but got {}".format(data_format)
x_dim = len(x.shape)
original_data_format = data_format
unsqueezed_dim = []
if isinstance(pad, Variable):
if data_format in ["NCL", "NCHW", "NCDHW"]:
data_format = "NCDHW"
if x_dim == 3:
pad = concat([zeros((4, ), dtype="int32"), pad], axis=0)
unsqueezed_dim = [3, 4]
x = unsqueeze(x, axes=unsqueezed_dim)
elif x_dim == 4:
pad = concat([pad, zeros((2, ), dtype="int32")], axis=0)
unsqueezed_dim = [2]
x = unsqueeze(x, axes=unsqueezed_dim)
elif data_format in ["NLC", "NHWC", "NDHWC"]:
data_format = "NDHWC"
if x_dim == 3:
pad = concat([zeros((4, ), dtype="int32"), pad], axis=0)
unsqueezed_dim = [2, 3]
x = unsqueeze(x, axes=unsqueezed_dim)
elif x_dim == 4:
pad = concat([pad, zeros((2, ), dtype="int32")], axis=0)
unsqueezed_dim = [1]
x = unsqueeze(x, axes=unsqueezed_dim)
else:
if data_format in ["NCL", "NCHW", "NCDHW"]:
data_format = "NCDHW"
if x_dim == 3:
pad = [0, 0, 0, 0] + pad
unsqueezed_dim = [3, 4]
x = unsqueeze(x, axes=unsqueezed_dim)
elif x_dim == 4:
pad = pad + [0, 0]
unsqueezed_dim = [2]
x = unsqueeze(x, axes=unsqueezed_dim)
elif data_format in ["NLC", "NHWC", "NDHWC"]:
data_format = "NDHWC"
if x_dim == 3:
pad = [0, 0, 0, 0] + pad
unsqueezed_dim = [2, 3]
x = unsqueeze(x, axes=unsqueezed_dim)
elif x_dim == 4:
pad = pad + [0, 0]
unsqueezed_dim = [1]
x = unsqueeze(x, axes=unsqueezed_dim)
if in_dygraph_mode():
if isinstance(pad, Variable):
pad = pad.numpy()
out = core.ops.pad3d(x, "paddings", pad, "mode", mode, "value", value,
"data_format", data_format, "name", name)
else:
attrs = {'mode': mode, 'value': value, 'data_format': data_format}
inputs = {'X': [x]}
if isinstance(pad, Variable):
inputs['Paddings'] = [pad]
attrs['paddings'] = []
else:
attrs['paddings'] = pad
helper = LayerHelper('pad3d', **locals())
dtype = helper.input_dtype(input_param_name='input')
out = helper.create_variable_for_type_inference(dtype)
helper.append_op(
type='pad3d', inputs=inputs, outputs={"Out": out}, attrs=attrs)
if len(unsqueezed_dim) != 0:
out = squeeze(out, axes=unsqueezed_dim)
return out
def cosine_similarity(x1, x2, dim=1, eps=1e-8):
"""
Compute cosine similarity between x1 and x2 along dim.
Parameters:
x1 (Tensor): First input. float32/double.
x2 (Tensor): Second input. float32/double.
dim (int): Dimension of vectors to compute cosine similarity. Default is 1.
eps(float): Small value to avoid division by zero. Default is 1e-8.
Returns: a Tensor representing cosine similarity between x1 and x2 along dim.
Return Type: Tensor
Examples:
.. code-block:: text
Case 0:
x1 = [[0.8024077 0.9927354 0.27238318 0.8344984 ]
[0.48949873 0.5797396 0.65444374 0.66510963]
[0.1031398 0.9614342 0.08365563 0.6796464 ]
[0.10760343 0.7461209 0.7726148 0.5801006 ]]
x2 = [[0.62913156 0.1536727 0.9847992 0.04591406]
[0.9098952 0.15715368 0.8671125 0.3156102 ]
[0.4427798 0.54136837 0.5276275 0.32394758]
[0.3769419 0.8535014 0.48041078 0.9256797 ]]
dim = 1
eps = 1e-8
Out: [0.5275037 0.8368967 0.75037485 0.9245899]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
np.random.seed(0)
x1 = np.random.rand(2,3)
x2 = np.random.rand(2,3)
x1 = paddle.to_tensor(x1)
x2 = paddle.to_tensor(x2)
result = paddle.nn.functional.cosine_similarity(x1, x2, dim=0)
print(result.numpy())
# [0.99806249 0.9817672 0.94987036]
"""
w12 = sum(elementwise_mul(x1, x2), dim=dim)
w1 = sum(elementwise_mul(x1, x1), dim=dim)
w2 = sum(elementwise_mul(x2, x2), dim=dim)
n12 = sqrt(clamp(w1 * w2, min=eps * eps))
cos_sim = w12 / n12
return cos_sim
python/paddle/nn/layer/__init__.py
浏览文件 @ bcf03273
...@@ -38,6 +38,16 @@ from .activation import HSigmoid #DEFINE_ALIAS ...@@ -38,6 +38,16 @@ from .activation import HSigmoid #DEFINE_ALIAS
from .common import BilinearTensorProduct #DEFINE_ALIAS from .common import BilinearTensorProduct #DEFINE_ALIAS
from .common import Pool2D #DEFINE_ALIAS from .common import Pool2D #DEFINE_ALIAS
from .common import Pad2D #DEFINE_ALIAS from .common import Pad2D #DEFINE_ALIAS
from .common import ReflectionPad1d #DEFINE_ALIAS
from .common import ReplicationPad1d #DEFINE_ALIAS
from .common import ConstantPad1d #DEFINE_ALIAS
from .common import ReflectionPad2d #DEFINE_ALIAS
from .common import ReplicationPad2d #DEFINE_ALIAS
from .common import ConstantPad2d #DEFINE_ALIAS
from .common import ZeroPad2d #DEFINE_ALIAS
from .common import ReplicationPad3d #DEFINE_ALIAS
from .common import ConstantPad3d #DEFINE_ALIAS
from .common import CosineSimilarity #DEFINE_ALIAS
from .common import Embedding #DEFINE_ALIAS from .common import Embedding #DEFINE_ALIAS
from .common import Linear #DEFINE_ALIAS from .common import Linear #DEFINE_ALIAS
from .common import Flatten #DEFINE_ALIAS from .common import Flatten #DEFINE_ALIAS
......
python/paddle/nn/layer/common.py
浏览文件 @ bcf03273
...@@ -22,8 +22,22 @@ from ...fluid.dygraph import layers ...@@ -22,8 +22,22 @@ from ...fluid.dygraph import layers
from .. import functional as F from .. import functional as F
__all__ = [ __all__ = [
'BilinearTensorProduct', 'Pool2D', 'Embedding', 'Linear', 'UpSample', 'BilinearTensorProduct',
'Pad2D' 'Pool2D',
'Embedding',
'Linear',
'UpSample',
'Pad2D',
'ReflectionPad1d',
'ReplicationPad1d',
'ConstantPad1d',
'ReflectionPad2d',
'ReplicationPad2d',
'ConstantPad2d',
'ZeroPad2d',
'ConstantPad3d',
'ReplicationPad3d',
'CosineSimilarity',
] ]
...@@ -258,12 +272,10 @@ class Pad2D(layers.Layer): ...@@ -258,12 +272,10 @@ class Pad2D(layers.Layer):
""" """
:alias_main: paddle.nn.Pad2D :alias_main: paddle.nn.Pad2D
:alias: paddle.nn.Pad2D,paddle.nn.layer.Pad2D,paddle.nn.layer.common.Pad2D :alias: paddle.nn.Pad2D,paddle.nn.layer.Pad2D,paddle.nn.layer.common.Pad2D
This interface is used to construct a callable object of the ``Pad2D`` class. This interface is used to construct a callable object of the ``Pad2D`` class.
The Pad2D layer pads the input tensor boundaries according to 'paddings' and 'mode'. The Pad2D layer pads the input tensor boundaries according to 'paddings' and 'mode'.
If mode is 'reflect', paddings[0] and paddings[1] must be no greater If mode is 'reflect', paddings[0] and paddings[1] must be no greater
than height-1. And the width dimension has the same condition. than height-1. And the width dimension has the same condition.
Parameters: Parameters:
paddings (int | List[int32]): The padding size. If padding is a int, uses the same paddings (int | List[int32]): The padding size. If padding is a int, uses the same
padding in all boundaries, if padding is a List, it must contain four integers, padding in all boundaries, if padding is a List, it must contain four integers,
...@@ -278,16 +290,12 @@ class Pad2D(layers.Layer): ...@@ -278,16 +290,12 @@ class Pad2D(layers.Layer):
data_format (str): An string from: "NHWC", "NCHW". Specify the data format of data_format (str): An string from: "NHWC", "NCHW". Specify the data format of
the input data. the input data.
Default is "NCHW" Default is "NCHW"
Returns: Returns:
None None
Examples: Examples:
.. code-block:: text .. code-block:: text
Input = [[[[1., 2., 3.], Input = [[[[1., 2., 3.],
[4., 5., 6.]]]] [4., 5., 6.]]]]
Case 0: Case 0:
paddings = [0, 1, 2, 3], paddings = [0, 1, 2, 3],
mode = 'constant' mode = 'constant'
...@@ -295,24 +303,20 @@ class Pad2D(layers.Layer): ...@@ -295,24 +303,20 @@ class Pad2D(layers.Layer):
Out = [[[[0., 0., 1., 2., 3., 0., 0., 0.], Out = [[[[0., 0., 1., 2., 3., 0., 0., 0.],
[0., 0., 4., 5., 6., 0., 0., 0.], [0., 0., 4., 5., 6., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0.]]]] [0., 0., 0., 0., 0., 0., 0., 0.]]]]
Case 1: Case 1:
paddings = [0, 1, 2, 1], paddings = [0, 1, 2, 1],
mode = 'reflect' mode = 'reflect'
Out = [[[[3., 2., 1., 2., 3., 2.], Out = [[[[3., 2., 1., 2., 3., 2.],
[6., 5., 4., 5., 6., 5.], [6., 5., 4., 5., 6., 5.],
[3., 2., 1., 2., 3., 2.]]]] [3., 2., 1., 2., 3., 2.]]]]
Case 2: Case 2:
paddings = [0, 1, 2, 1], paddings = [0, 1, 2, 1],
mode = 'edge' mode = 'edge'
Out = [[[[1., 1., 1., 2., 3., 3.], Out = [[[[1., 1., 1., 2., 3., 3.],
[4., 4., 4., 5., 6., 6.], [4., 4., 4., 5., 6., 6.],
[4., 4., 4., 5., 6., 6.]]]] [4., 4., 4., 5., 6., 6.]]]]
Code Examples: Code Examples:
.. code-block:: python .. code-block:: python
import paddle.fluid as fluid import paddle.fluid as fluid
import paddle.nn as nn import paddle.nn as nn
import numpy as np import numpy as np
...@@ -342,3 +346,621 @@ class Pad2D(layers.Layer): ...@@ -342,3 +346,621 @@ class Pad2D(layers.Layer):
mode=self._mode, mode=self._mode,
pad_value=self._pad_value, pad_value=self._pad_value,
data_format=self._data_format) data_format=self._data_format)
class ReflectionPad1d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ReflectionPad1d`` class.
Uses reflection of the input boundaries to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right).
data_format (str): An string from: "NCL", "NLC". Specify the data format of the input data.
Default is "NCL"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[1., 2., 3.],
[4., 5., 6.]]]
padding = [1, 2],
Out = [[[2. 1. 2. 3. 2. 1.]
[5. 4. 5. 6. 5. 4.]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 2, 3)
pad = [1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ReflectionPad1d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[2. 1. 2. 3. 2. 1.]
# [5. 4. 5. 6. 5. 4.]]]
"""
def __init__(self, padding, data_format="NCL", name=None):
super(ReflectionPad1d, self).__init__()
self._mode = "reflect"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class ReplicationPad1d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ReplicationPad1d`` class.
Uses input boundaries to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right).
data_format (str): An string from: "NCL", "NLC". Specify the data format of the input data.
Default is "NCL"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[1., 2., 3.],
[4., 5., 6.]]]
padding = [1, 2],
Out = [[[2. 1. 2. 3. 2. 1.]
[5. 4. 5. 6. 5. 4.]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 2, 3)
pad = [1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ReplicationPad1d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[1. 1. 2. 3. 3. 3.]
# [1. 4. 5. 6. 6. 6.]]]
"""
def __init__(self, padding, data_format="NCL", name=None):
super(ReplicationPad1d, self).__init__()
self._mode = "replicate"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class ConstantPad1d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ConstantPad1d`` class.
Uses a constant value to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right).
value (float32): The value to fill the padded areas. Default is 0.0
data_format (str): An string from: "NCL", "NLC". Specify the data format of the input data.
Default is "NCL"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[1., 2., 3.],
[4., 5., 6.]]]
padding = [1, 2],
value = 0.0
Out = [[[0. 1. 2. 3. 0. 0.]
[0. 4. 5. 6. 0. 0.]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 2, 3)
pad = [1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ConstantPad1d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[0. 1. 2. 3. 0. 0.]
# [0. 4. 5. 6. 0. 0.]]]
"""
def __init__(self, padding, value=0.0, data_format="NCL", name=None):
super(ConstantPad1d, self).__init__()
self._mode = "constant"
self._data_format = data_format
self._pad = padding
self._value = value
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
value=self._value,
data_format=self._data_format,
name=self._name)
class ConstantPad2d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ConstantPad2d`` class.
Uses a constant value to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom).
value (float32): The value to fill the padded areas. Default is 0.0
data_format (str): An string from: "NCHW", "NHWC". Specify the data format of the input data.
Default is "NCHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[1., 2., 3.],
[4., 5., 6.]]]]
padding = [1, 1, 0, 0]
value = 0.0
Out = [[[[0. 1. 2. 3. 0.]
[0. 4. 5. 6. 0.]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 2, 3)
pad = [1, 0, 1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ConstantPad2d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[0. 0. 0. 0.]
# [0. 1. 2. 3.]
# [0. 4. 5. 6.]
# [0. 0. 0. 0.]
# [0. 0. 0. 0.]]]]
"""
def __init__(self, padding, value=0.0, data_format="NCHW", name=None):
super(ConstantPad2d, self).__init__()
self._mode = "constant"
self._data_format = data_format
self._pad = padding
self._value = value
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
value=self._value,
data_format=self._data_format,
name=self._name)
class ZeroPad2d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ZeroPad2d`` class.
Uses 0 to pad the input tensor.
Parameters:
padding (Variable | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom).
data_format (str): An string from: "NCHW", "NHWC". Specify the data format of the input data.
Default is "NCHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[1., 2., 3.],
[4., 5., 6.]]]]
padding = [1, 1, 0, 0]
Out = [[[[0. 1. 2. 3. 0.]
[0. 4. 5. 6. 0.]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 2, 3)
pad = [1, 0, 1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ZeroPad2d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[0. 0. 0. 0.]
# [0. 1. 2. 3.]
# [0. 4. 5. 6.]
# [0. 0. 0. 0.]
# [0. 0. 0. 0.]]]]
"""
def __init__(self, padding, data_format="NCHW", name=None):
super(ZeroPad2d, self).__init__()
self._mode = "constant"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class ReplicationPad2d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ReplicationPad2d`` class.
Uses input boundaries to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom).
data_format (str): An string from: "NCHW", "NHWC". Specify the data format of the input data.
Default is "NCHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[1., 2., 3.],
[4., 5., 6.]]]]
padding = [1, 1, 0, 0]
Out = [[[[1. 1. 2. 3. 3.]
[4. 4. 5. 6. 6.]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 2, 3)
pad = [1, 0, 1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ReplicationPad2d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[1. 1. 2. 3.]
# [1. 1. 2. 3.]
# [4. 4. 5. 6.]
# [4. 4. 5. 6.]
# [4. 4. 5. 6.]]]]
"""
def __init__(self, padding, data_format="NCHW", name=None):
super(ReplicationPad2d, self).__init__()
self._mode = "replicate"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class ReflectionPad2d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ReflectionPad2d`` class.
Uses reflection of the input boundaries to pad the input tensor.
Parameters:
padding (Variable | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom).
data_format (str): An string from: "NCHW", "NHWC". Specify the data format of the input data.
Default is "NCHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[1., 2., 3.],
[4., 5., 6.]]]]
padding = [1, 1, 0, 0]
Out = [[[[2. 1. 2. 3. 2.]
[5. 4. 5. 6. 5.]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 4, 3)
pad = [1, 0, 1, 2]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ReflectionPad2d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[ 5. 4. 5. 6.]
# [ 2. 1. 2. 3.]
# [ 5. 4. 5. 6.]
# [ 8. 7. 8. 9.]
# [11. 10. 11. 12.]
# [ 8. 7. 8. 9.]
# [ 5. 4. 5. 6.]]]]
"""
def __init__(self, padding, data_format="NCHW", name=None):
super(ReflectionPad2d, self).__init__()
self._mode = "reflect"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class ConstantPad3d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ConstantPad3d`` class.
Uses a constant value to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom, pad_front, pad_back).
value (float32): The value to fill the padded areas. Default is 0.0
data_format (str): An string from: "NCDHW", "NDHWC". Specify the data format of the input data.
Default is "NCDHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[[1., 2., 3.],
[4., 5., 6.]]]]]
padding = [1, 2, 0, 0, 0, 0]
value = 0.0
Out = [[[[[0. 1. 2. 3. 0. 0.]
[0. 4. 5. 6. 0. 0.]]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 1, 2, 3)
pad = [1, 0, 1, 2, 0, 0]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ConstantPad3d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[[0. 0. 0. 0.]
# [0. 1. 2. 3.]
# [0. 4. 5. 6.]
# [0. 0. 0. 0.]
# [0. 0. 0. 0.]]]]]
"""
def __init__(self, padding, value=0.0, data_format="NCDHW", name=None):
super(ConstantPad3d, self).__init__()
self._mode = "constant"
self._data_format = data_format
self._pad = padding
self._value = value
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
value=self._value,
data_format=self._data_format,
name=self._name)
class ReplicationPad3d(layers.Layer):
"""
This interface is used to construct a callable object of the ``ReplicationPad3d`` class.
Uses input boundaries to pad the input tensor.
Parameters:
padding (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom, pad_front, pad_back).
data_format (str): An string from: "NCDHW", "NDHWC". Specify the data format of the input data.
Default is "NCDHW"
name (str, optional) : The default value is None. Normally there is no need for
user to set this property. For more information, please refer to :ref:`api_guide_Name`.
Returns:
None
Examples:
.. code-block:: text
x = [[[[[1., 2., 3.],
[4., 5., 6.]]]]]
padding = [1, 2, 0, 0, 0, 0]
Out = [[[[[1. 1. 2. 3. 3. 3.]
[4. 4. 5. 6. 6. 6.]]]]]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
input_shape = (1, 1, 1, 2, 3)
pad = [1, 0, 1, 2, 0, 0]
data = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape) + 1
my_pad = nn.ReplicationPad3d(padding=pad)
data = paddle.to_tensor(data)
result = my_pad(data)
print(result.numpy())
# [[[[[1. 1. 2. 3.]
# [1. 1. 2. 3.]
# [4. 4. 5. 6.]
# [4. 4. 5. 6.]
# [4. 4. 5. 6.]]]]]
"""
def __init__(self, padding, data_format="NCDHW", name=None):
super(ReplicationPad3d, self).__init__()
self._mode = "replicate"
self._data_format = data_format
self._pad = padding
self._name = name
def forward(self, x):
return F.pad(x,
pad=self._pad,
mode=self._mode,
data_format=self._data_format,
name=self._name)
class CosineSimilarity(layers.Layer):
"""
This interface is used to compute cosine similarity between x1 and x2 along dim.
Parameters:
dim (int): Dimension of vectors to compute cosine similarity. Default is 1.
eps(float): Small value to avoid division by zero. Default is 1e-8.
Returns:
None
Examples:
.. code-block:: text
Case 0:
x1 = [[0.8024077 0.9927354 0.27238318 0.8344984 ]
[0.48949873 0.5797396 0.65444374 0.66510963]
[0.1031398 0.9614342 0.08365563 0.6796464 ]
[0.10760343 0.7461209 0.7726148 0.5801006 ]]
x2 = [[0.62913156 0.1536727 0.9847992 0.04591406]
[0.9098952 0.15715368 0.8671125 0.3156102 ]
[0.4427798 0.54136837 0.5276275 0.32394758]
[0.3769419 0.8535014 0.48041078 0.9256797 ]]
dim = 1
eps = 1e-8
Out: [0.5275037 0.8368967 0.75037485 0.9245899]
Code Examples:
.. code-block:: python
import paddle
import paddle.nn as nn
import numpy as np
paddle.disable_static()
np.random.seed(0)
x1 = np.random.rand(2,3)
x2 = np.random.rand(2,3)
x1 = paddle.to_tensor(x1)
x2 = paddle.to_tensor(x2)
cos_sim_func = nn.CosineSimilarity(dim=0)
result = cos_sim_func(x1, x2)
print(result.numpy())
# [0.99806249 0.9817672 0.94987036]
"""
def __init__(self, dim=1, eps=1e-8):
super(CosineSimilarity, self).__init__()
self._dim = dim
self._eps = eps
def forward(self, x1, x2):
return F.cosine_similarity(x1, x2, dim=self._dim, eps=self._eps)
python/paddle/tensor/__init__.py
浏览文件 @ bcf03273
...@@ -111,6 +111,7 @@ from .math import cumsum #DEFINE_ALIAS ...@@ -111,6 +111,7 @@ from .math import cumsum #DEFINE_ALIAS
from .math import elementwise_add #DEFINE_ALIAS from .math import elementwise_add #DEFINE_ALIAS
from .math import elementwise_div #DEFINE_ALIAS from .math import elementwise_div #DEFINE_ALIAS
from .math import elementwise_floordiv #DEFINE_ALIAS from .math import elementwise_floordiv #DEFINE_ALIAS
from .math import elementwise_mul #DEFINE_ALIAS
from .math import elementwise_mod #DEFINE_ALIAS from .math import elementwise_mod #DEFINE_ALIAS
from .math import elementwise_pow #DEFINE_ALIAS from .math import elementwise_pow #DEFINE_ALIAS
from .math import elementwise_sub #DEFINE_ALIAS from .math import elementwise_sub #DEFINE_ALIAS
......
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