提交 5e52dafd 编写于 作者: J jerrywgz

add roi align

上级 5fc30522
/* Copyright (c) 2018 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 "paddle/fluid/operators/roi_align_op.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
class ROIAlignOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of ROIAlignOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("ROIs"),
"Input(ROIs) of ROIAlignOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of ROIAlignOp should not be null.");
auto input_dims = ctx->GetInputDim("X");
auto rois_dims = ctx->GetInputDim("ROIs");
PADDLE_ENFORCE(input_dims.size() == 4,
"The format of input tensor is NCHW.");
PADDLE_ENFORCE(rois_dims.size() == 2,
"ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
"given as [[x1, y1, x2, y2], …].");
PADDLE_ENFORCE(rois_dims[1] == 4,
"ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
"given as [[x1, y1, x2, y2], …].");
int pooled_height = ctx->Attrs().Get<int>("pooled_height");
int pooled_width = ctx->Attrs().Get<int>("pooled_width");
float spatial_scale = ctx->Attrs().Get<float>("spatial_scale");
PADDLE_ENFORCE_GT(pooled_height, 0,
"The pooled output height must greater than 0");
PADDLE_ENFORCE_GT(pooled_width, 0,
"The pooled output width must greater than 0");
PADDLE_ENFORCE_GT(spatial_scale, 0.0f,
"The spatial scale must greater than 0");
auto out_dims = input_dims;
out_dims[0] = rois_dims[0];
out_dims[1] = input_dims[1];
out_dims[2] = pooled_height;
out_dims[3] = pooled_width;
ctx->SetOutputDim("Out", out_dims);
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
ctx.device_context());
}
};
class ROIAlignGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
"The GRAD@Out of ROIAlignGradOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")),
"The GRAD@X of ROIAlignGradOp should not be null.");
ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
ctx.device_context());
}
};
class ROIAlignOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X",
"(Tensor), "
"the input of ROIAlignOp. "
"The format of input tensor is NCHW. Where N is batch size, "
"C is the number of input channels, "
"H is the height of the feature, and "
"W is the width of the feature.");
AddInput("ROIs",
"(LoDTensor), "
"ROIs (Regions of Interest) to pool over. "
"should be a 2-D LoDTensor of shape (num_rois, 4)"
"given as [[x1, y1, x2, y2], …]. "
"Where batch_id is the id of the data, "
"(x1, y1) is the top left coordinates, and "
"(x2, y2) is the bottom right coordinates.");
AddOutput("Out",
"(Tensor), "
"The output of ROIAlignOp is a 4-D tensor with shape "
"(num_rois, channels, pooled_h, pooled_w).");
AddAttr<float>("spatial_scale",
"(float, default 1.0), "
"Multiplicative spatial scale factor "
"to translate ROI coords from their input scale "
"to the scale used when pooling.")
.SetDefault(1.0);
AddAttr<int>("pooled_height",
"(int, default 1), "
"The pooled output height.")
.SetDefault(1);
AddAttr<int>("pooled_width",
"(int, default 1), "
"The pooled output width.")
.SetDefault(1);
AddAttr<int>("sampling_ratio",
"(int,default -1),"
"number of sampling points in the interpolation grid"
"If <=0, then grid points are adaptive to roi_width "
"and pooled_w, likewise for height")
.SetDefault(-1);
AddComment(R"DOC(
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(roi_align, ops::ROIAlignOp, ops::ROIAlignOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(roi_align_grad, ops::ROIAlignGradOp);
REGISTER_OP_CPU_KERNEL(
roi_align,
ops::CPUROIAlignOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::CPUROIAlignOpKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
roi_align_grad,
ops::CPUROIAlignGradOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::CPUROIAlignGradOpKernel<paddle::platform::CPUDeviceContext, double>);
/* Copyright (c) 2018 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. */
#pragma once
#include <algorithm>
#include <limits>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
template <class T>
void pre_calc_for_bilinear_interpolate(
const platform::DeviceContext& ctx, const int height, const int width,
const int pooled_height, const int pooled_width, const int iy_upper,
const int ix_upper, T roi_ymin, T roi_xmin, T bin_size_h, T bin_size_w,
int roi_bin_grid_h, int roi_bin_grid_w, Tensor* pre_pos, Tensor* pre_w) {
int pre_calc_index = 0;
int* pre_pos_data = pre_pos->mutable_data<int>(ctx.GetPlace());
T* pre_w_data = pre_w->mutable_data<T>(ctx.GetPlace());
for (int ph = 0; ph < pooled_height; ph++) {
for (int pw = 0; pw < pooled_width; pw++) {
for (int iy = 0; iy < iy_upper; iy++) {
// calculate y of sample points
T y = roi_ymin + ph * bin_size_h +
static_cast<T>(iy + .5f) * bin_size_h /
static_cast<T>(roi_bin_grid_h);
// calculate x of samle points
for (int ix = 0; ix < ix_upper; ix++) {
T x = roi_xmin + pw * bin_size_w +
static_cast<T>(ix + .5f) * bin_size_w /
static_cast<T>(roi_bin_grid_w);
// deal with elements out of map
if (y < -1.0 || y > height || x < -1.0 || x > width) {
for (int i = 0; i < 4; ++i) {
pre_pos_data[i + pre_calc_index * 4] = 0;
pre_w_data[i + pre_calc_index * 4] = 0;
}
pre_calc_index += 1;
continue;
}
if (y <= 0) {
y = 0;
}
if (x <= 0) {
x = 0;
}
int y_low = static_cast<int>(y);
int x_low = static_cast<int>(x);
int y_high;
int x_high;
if (y_low >= height - 1) {
y_high = y_low = height - 1;
y = static_cast<T>(y_low);
} else {
y_high = y_low + 1;
}
if (x_low >= width - 1) {
x_high = x_low = width - 1;
x = static_cast<T>(x_low);
} else {
x_high = x_low + 1;
}
T ly = y - y_low, lx = x - x_low;
T hy = 1. - ly, hx = 1. - lx;
pre_pos_data[pre_calc_index * 4] = y_low * width + x_low;
pre_pos_data[pre_calc_index * 4 + 1] = y_low * width + x_high;
pre_pos_data[pre_calc_index * 4 + 2] = y_high * width + x_low;
pre_pos_data[pre_calc_index * 4 + 3] = y_high * width + x_high;
pre_w_data[pre_calc_index * 4] = hy * hx;
pre_w_data[pre_calc_index * 4 + 1] = hy * lx;
pre_w_data[pre_calc_index * 4 + 2] = ly * hx;
pre_w_data[pre_calc_index * 4 + 3] = ly * lx;
pre_calc_index += 1;
}
}
}
}
}
template <class T>
void bilinear_interpolate_gradient(const int height, const int width, T y, T x,
const T out_grad_this_bin, const T count,
T* batch_grad_data) {
int x_low, y_low, x_high, y_high;
T w1, w2, w3, w4;
if (y < -1.0 || y > height || x < -1.0 || x > width) {
w1 = w2 = w3 = w4 = 0;
x_low = x_high = y_low = y_high = -1;
return;
}
if (y <= 0) {
y = 0;
}
if (x <= 0) {
x = 0;
}
y_low = static_cast<int>(y);
x_low = static_cast<int>(x);
if (y_low >= height - 1) {
y_high = y_low = height - 1;
y = static_cast<T>(y_low);
} else {
y_high = y_low + 1;
}
if (x_low >= width - 1) {
x_high = x_low = width - 1;
x = static_cast<T>(x_low);
} else {
x_high = x_low + 1;
}
T ly = y - y_low, lx = x - x_low;
T hy = 1. - ly, hx = 1. - lx;
w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
T diff1 = out_grad_this_bin * w1 / count;
T diff2 = out_grad_this_bin * w2 / count;
T diff3 = out_grad_this_bin * w3 / count;
T diff4 = out_grad_this_bin * w4 / count;
if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
*(batch_grad_data + y_low * width + x_low) += diff1;
*(batch_grad_data + y_low * width + x_high) += diff2;
*(batch_grad_data + y_high * width + x_low) += diff3;
*(batch_grad_data + y_high * width + x_high) += diff4;
}
return;
}
template <typename DeviceContext, typename T>
class CPUROIAlignOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* in = ctx.Input<framework::Tensor>("X");
auto* rois = ctx.Input<framework::LoDTensor>("ROIs");
auto* out = ctx.Output<framework::Tensor>("Out");
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
auto spatial_scale = ctx.Attr<float>("spatial_scale");
auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
auto& dev_ctx = ctx.template device_context<DeviceContext>();
auto in_dims = in->dims();
int64_t batch_size = in_dims[0];
int64_t channels = in_dims[1];
int64_t height = in_dims[2];
int64_t width = in_dims[3];
int64_t rois_num = rois->dims()[0];
auto in_stride = framework::stride(in_dims);
auto roi_stride = framework::stride(rois->dims());
auto out_stride = framework::stride(out->dims());
const T* input_data = in->data<T>();
framework::Tensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
int* roi_batch_id_data =
roi_batch_id_list.mutable_data<int>(ctx.GetPlace());
auto rois_lod = rois->lod().back();
int rois_batch_size = rois_lod.size() - 1;
PADDLE_ENFORCE_EQ(
rois_batch_size, batch_size,
"The rois_batch_size and imgs batch_size must be the same.");
int rois_num_with_lod = rois_lod[rois_batch_size];
PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod,
"The rois_num from input and lod must be the same.");
for (int n = 0; n < rois_batch_size; ++n) {
for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
roi_batch_id_data[i] = n;
}
}
T* output_data = out->mutable_data<T>(ctx.GetPlace());
const T* rois_data = rois->data<T>();
for (int n = 0; n < rois_num; ++n) {
int roi_batch_id = roi_batch_id_data[n];
T roi_xmin = rois_data[0] * spatial_scale;
T roi_ymin = rois_data[1] * spatial_scale;
T roi_xmax = rois_data[2] * spatial_scale;
T roi_ymax = rois_data[3] * spatial_scale;
T roi_width = std::max(roi_xmax - roi_xmin, static_cast<T>(1.));
T roi_height = std::max(roi_ymax - roi_ymin, static_cast<T>(1.));
T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
const T* batch_data = input_data + roi_batch_id * in_stride[0];
int roi_bin_grid_h = (sampling_ratio > 0)
? sampling_ratio
: ceil(roi_height / pooled_height);
int roi_bin_grid_w = (sampling_ratio > 0)
? sampling_ratio
: ceil(roi_width / pooled_width);
const T count = roi_bin_grid_h * roi_bin_grid_w;
Tensor pre_pos;
Tensor pre_w;
int pre_size = count * out_stride[1];
pre_pos.Resize({pre_size, 4});
pre_w.Resize({pre_size, 4});
pre_calc_for_bilinear_interpolate(
dev_ctx, height, width, pooled_height, pooled_width, roi_bin_grid_h,
roi_bin_grid_w, roi_ymin, roi_xmin, bin_size_h, bin_size_w,
roi_bin_grid_h, roi_bin_grid_w, &pre_pos, &pre_w);
const int* pre_pos_data = pre_pos.data<int>();
const T* pre_w_data = pre_w.data<T>();
for (int c = 0; c < channels; c++) {
int pre_calc_index = 0;
for (int ph = 0; ph < pooled_height; ph++) {
for (int pw = 0; pw < pooled_width; pw++) {
const int pool_index = ph * pooled_width + pw;
T output_val = 0;
for (int iy = 0; iy < roi_bin_grid_h; iy++) {
for (int ix = 0; ix < roi_bin_grid_w; ix++) {
for (int i = 0; i < 4; i++) {
int pos = pre_pos_data[pre_calc_index * 4 + i];
T w = pre_w_data[pre_calc_index * 4 + i];
output_val += w * batch_data[pos];
}
pre_calc_index += 1;
}
}
output_val /= count;
output_data[pool_index] = output_val;
}
}
batch_data += in_stride[1];
output_data += out_stride[1];
}
rois_data += roi_stride[0];
}
return;
}
};
template <typename DeviceContext, typename T>
class CPUROIAlignGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* in = ctx.Input<framework::Tensor>("X");
auto* rois = ctx.Input<framework::LoDTensor>("ROIs");
auto* out_grad =
ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* in_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
auto spatial_scale = ctx.Attr<float>("spatial_scale");
auto sampling_ratio = ctx.Attr<int>("sampling_ratio");
auto in_dims = in->dims();
if (in_grad) {
int64_t channels = in_dims[1];
int64_t height = in_dims[2];
int64_t width = in_dims[3];
int rois_num = rois->dims()[0];
framework::Tensor roi_batch_id_list;
roi_batch_id_list.Resize({rois_num});
int* roi_batch_id_data =
roi_batch_id_list.mutable_data<int>(ctx.GetPlace());
auto rois_lod = rois->lod().back();
int rois_batch_size = rois_lod.size() - 1;
for (int n = 0; n < rois_batch_size; ++n) {
for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
roi_batch_id_data[i] = n;
}
}
const T* rois_data = rois->data<T>();
const T* out_grad_data = out_grad->data<T>();
T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
auto in_stride = framework::stride(in->dims());
auto roi_stride = framework::stride(rois->dims());
auto out_stride = framework::stride(out_grad->dims());
for (int n = 0; n < rois_num; ++n) {
int roi_batch_idx = roi_batch_id_data[n];
T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0];
const T* batch_out_grad_data =
out_grad_data + roi_batch_idx * out_stride[0];
T roi_xmin = rois_data[0] * spatial_scale;
T roi_ymin = rois_data[1] * spatial_scale;
T roi_xmax = rois_data[2] * spatial_scale;
T roi_ymax = rois_data[3] * spatial_scale;
T roi_width = std::max(roi_xmax - roi_xmin, static_cast<T>(1.));
T roi_height = std::max(roi_ymax - roi_ymin, static_cast<T>(1.));
T bin_size_h =
static_cast<T>(roi_height) / static_cast<T>(pooled_height);
T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
for (int c = 0; c < channels; ++c) {
for (int ph = 0; ph < pooled_height; ++ph) {
for (int pw = 0; pw < pooled_width; ++pw) {
int pool_index = ph * pooled_width + pw;
T out_grad_this_bin = batch_out_grad_data[pool_index];
int roi_bin_grid_h = (sampling_ratio > 0)
? sampling_ratio
: ceil(roi_height / pooled_height);
int roi_bin_grid_w = (sampling_ratio > 0)
? sampling_ratio
: ceil(roi_width / pooled_width);
T count = roi_bin_grid_h * roi_bin_grid_w;
for (int iy = 0; iy < roi_bin_grid_h; iy++) {
const T y = roi_ymin + ph * bin_size_h +
static_cast<T>(iy + .5f) * bin_size_h /
static_cast<T>(roi_bin_grid_h);
for (int ix = 0; ix < roi_bin_grid_w; ix++) {
const T x = roi_xmin + pw * bin_size_w +
static_cast<T>(ix + .5f) * bin_size_w /
static_cast<T>(roi_bin_grid_w);
bilinear_interpolate_gradient(height, width, y, x,
out_grad_this_bin, count,
batch_grad_data);
}
}
}
}
batch_grad_data += in_stride[1];
batch_out_grad_data += out_stride[1];
}
rois_data += roi_stride[0];
}
}
return;
}
};
} // namespace operators
} // namespace paddle
# Copyright (c) 2018 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.
from __future__ import print_function
import unittest
import numpy as np
import math
import sys
from op_test import OpTest
class TestROIAlignOp(OpTest):
def set_data(self):
self.init_test_case()
self.make_rois()
self.calc_roi_align()
self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)}
self.attrs = {
'spatial_scale': self.spatial_scale,
'pooled_height': self.pooled_height,
'pooled_width': self.pooled_width,
'sampling_ratio': self.sampling_ratio
}
self.outputs = {'Out': self.out_data}
def init_test_case(self):
self.batch_size = 1
self.channels = 3
self.height = 8
self.width = 6
# n, c, h, w
self.x_dim = (self.batch_size, self.channels, self.height, self.width)
self.spatial_scale = 1.0 / 1.0
self.pooled_height = 2
self.pooled_width = 2
self.sampling_ratio = 2
self.x = np.random.random(self.x_dim).astype('float32')
def pre_calc(self, x_i, roi_xmin, roi_ymin, roi_bin_grid_h, roi_bin_grid_w,
bin_size_h, bin_size_w):
count = roi_bin_grid_h * roi_bin_grid_w
bilinear_pos = np.zeros(
[self.channels, self.pooled_height, self.pooled_width, count, 4],
np.int32)
bilinear_w = np.zeros(
[self.pooled_height, self.pooled_width, count, 4], np.float32)
for ph in range(self.pooled_width):
for pw in range(self.pooled_height):
c = 0
for iy in range(roi_bin_grid_h):
y = roi_ymin + ph * bin_size_h + (iy + 0.5) * \
bin_size_h / roi_bin_grid_h
for ix in range(roi_bin_grid_w):
x = roi_xmin + pw * bin_size_w + (ix + 0.5) * \
bin_size_w / roi_bin_grid_w
if y < -1.0 or y > self.height or \
x < -1.0 or x > self.width:
continue
if y <= 0:
y = 0
if x <= 0:
x = 0
y_low = int(y)
x_low = int(x)
if y_low >= self.height - 1:
y = y_high = y_low = self.height - 1
else:
y_high = y_low + 1
if x_low >= self.width - 1:
x = x_high = x_low = self.width - 1
else:
x_high = x_low = self.width - 1
ly = y - y_low
lx = x - x_low
hy = 1 - ly
hx = 1 - lx
for ch in range(self.channels):
bilinear_pos[ch, ph, pw, c, 0] = x_i[ch, y_low,
x_low]
bilinear_pos[ch, ph, pw, c, 1] = x_i[ch, y_low,
x_high]
bilinear_pos[ch, ph, pw, c, 2] = x_i[ch, y_high,
x_low]
bilinear_pos[ch, ph, pw, c, 3] = x_i[ch, y_high,
x_high]
bilinear_w[ph, pw, c, 0] = hy * hx
bilinear_w[ph, pw, c, 1] = hy * lx
bilinear_w[ph, pw, c, 2] = ly * hx
bilinear_w[ph, pw, c, 3] = ly * lx
c = c + 1
return bilinear_pos, bilinear_w
def calc_roi_align(self):
self.out_data = np.zeros((self.rois_num, self.channels,
self.pooled_height, self.pooled_width))
for i in range(self.rois_num):
roi = self.rois[i]
roi_batch_id = int(roi[0])
x_i = self.x[roi_batch_id]
roi_xmin = roi[1] * self.spatial_scale
roi_ymin = roi[2] * self.spatial_scale
roi_xmax = roi[3] * self.spatial_scale
roi_ymax = roi[4] * self.spatial_scale
roi_width = int(max(roi_xmax - roi_xmin, 1))
roi_height = int(max(roi_ymax - roi_ymin, 1))
bin_size_h = float(roi_height) / float(self.pooled_height)
bin_size_w = float(roi_width) / float(self.pooled_width)
roi_bin_grid_h = self.sampling_ratio if self.sampling_ratio > 0 else \
math.ceil(roi_height / pooled_height)
roi_bin_grid_w = self.sampling_ratio if self.sampling_ratio > 0 else \
math.ceil(roi_width / pooled_width)
count = int(roi_bin_grid_h * roi_bin_grid_w)
pre_size = count * self.pooled_width * self.pooled_height
bilinear_pos, bilinear_w = self.pre_calc(x_i, roi_xmin, roi_ymin,
int(roi_bin_grid_h),
int(roi_bin_grid_w),
bin_size_h, bin_size_w)
for ch in range(self.channels):
align_per_bin = (bilinear_pos[ch] * bilinear_w).sum(axis=-1)
output_val = align_per_bin.mean(axis=-1)
self.out_data[i, ch, :, :] = output_val
def make_rois(self):
rois = []
self.rois_lod = [[0]]
for bno in range(self.batch_size):
self.rois_lod[0].append(bno + 1)
for i in range(bno + 1):
x1 = np.random.random_integers(
0, self.width // self.spatial_scale - self.pooled_width)
y1 = np.random.random_integers(
0, self.height // self.spatial_scale - self.pooled_height)
x2 = np.random.random_integers(x1 + self.pooled_width,
self.width // self.spatial_scale)
y2 = np.random.random_integers(
y1 + self.pooled_height, self.height // self.spatial_scale)
roi = [bno, x1, y1, x2, y2]
rois.append(roi)
self.rois_num = len(rois)
self.rois = np.array(rois).astype("float32")
def setUp(self):
self.op_type = "roi_align"
self.set_data()
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['X'], 'Out')
Markdown is supported
0% .
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