SunGaofeng
7 years ago
committed by
qingqing01
parent
30aad88449
commit
e3c4b0dace
@ -0,0 +1,173 @@
|
|||||||
|
/* Copyright (c) 2016 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/psroi_pool_op.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
using LoDTensor = framework::LoDTensor;
|
||||||
|
|
||||||
|
class PSROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||||
|
public:
|
||||||
|
void Make() override {
|
||||||
|
AddInput("X",
|
||||||
|
"(Tensor), "
|
||||||
|
"the input of PSROIPoolOp. "
|
||||||
|
"The format of input tensor is NCHW. Where N is the batch size, "
|
||||||
|
"C is the number of input channels, "
|
||||||
|
"H is the height of the input feature map, and "
|
||||||
|
"W is the width.");
|
||||||
|
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 (x1, y1) is the top left coordinates, and "
|
||||||
|
"(x2, y2) is the bottom right coordinates. "
|
||||||
|
"The roi batch index can be calculated from LoD.");
|
||||||
|
AddOutput("Out",
|
||||||
|
"(Tensor), "
|
||||||
|
"the output of PSROIPoolOp is a 4-D Tensor with shape "
|
||||||
|
"(num_rois, output_channels, pooled_h, pooled_w).");
|
||||||
|
AddAttr<int>(
|
||||||
|
"output_channels",
|
||||||
|
"(int), "
|
||||||
|
"the number of channels of the output feature map. "
|
||||||
|
"For a task of C classes of objects, output_channels should be "
|
||||||
|
"(C + 1) for classification only.");
|
||||||
|
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);
|
||||||
|
AddComment(R"Doc(
|
||||||
|
**PSROIPool Operator**
|
||||||
|
|
||||||
|
Position sensitive region of interest pooling (also known as PSROIPooling) is to perform
|
||||||
|
position-sensitive average pooling on regions of interest specified by input, takes as
|
||||||
|
input N position-sensitive score maps and a list of num_rois regions of interest.
|
||||||
|
|
||||||
|
PSROIPooling for R-FCN. Please refer to https://arxiv.org/abs/1605.06409 for more details.
|
||||||
|
)Doc");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class PSROIPoolOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("X"),
|
||||||
|
"Input(X) of PSROIPoolOp should not be null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput("ROIs"),
|
||||||
|
"Input(ROIs) of PSROIPoolOp should not be null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasOutput("Out"),
|
||||||
|
"Output(Out) of PSROIPoolOp 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");
|
||||||
|
int output_channels = ctx->Attrs().Get<int>("output_channels");
|
||||||
|
float spatial_scale = ctx->Attrs().Get<float>("spatial_scale");
|
||||||
|
|
||||||
|
PADDLE_ENFORCE(
|
||||||
|
input_dims[1] == output_channels * pooled_height * pooled_width,
|
||||||
|
"the channel of X(%d) should be equal to the product of "
|
||||||
|
"output_channels(%d), pooled_height(%d) and pooled_width(%d)",
|
||||||
|
input_dims[1], output_channels, pooled_height, pooled_width);
|
||||||
|
|
||||||
|
PADDLE_ENFORCE_GT(pooled_height, 0,
|
||||||
|
"The pooled output height must be greater than 0");
|
||||||
|
PADDLE_ENFORCE_GT(pooled_width, 0,
|
||||||
|
"The pooled output width must be greater than 0");
|
||||||
|
PADDLE_ENFORCE_GT(output_channels, 1,
|
||||||
|
"The pooled output channels must greater than 1");
|
||||||
|
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] =
|
||||||
|
output_channels; // input_dims[1] / (pooled_height * pooled_width);
|
||||||
|
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 PSROIPoolGradOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||||
|
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
|
||||||
|
"The gradient of Out should not be null.");
|
||||||
|
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
|
||||||
|
"The gradient of X should not be null.");
|
||||||
|
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("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());
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OPERATOR(psroi_pool, ops::PSROIPoolOp, ops::PSROIPoolOpMaker,
|
||||||
|
paddle::framework::DefaultGradOpDescMaker<true>);
|
||||||
|
REGISTER_OPERATOR(psroi_pool_grad, ops::PSROIPoolGradOp);
|
||||||
|
REGISTER_OP_CPU_KERNEL(
|
||||||
|
psroi_pool,
|
||||||
|
ops::CPUPSROIPoolOpKernel<paddle::platform::CPUDeviceContext, float>,
|
||||||
|
ops::CPUPSROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
|
||||||
|
REGISTER_OP_CPU_KERNEL(
|
||||||
|
psroi_pool_grad,
|
||||||
|
ops::CPUPSROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, float>,
|
||||||
|
ops::CPUPSROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, double>);
|
||||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,134 @@
|
|||||||
|
# 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 math
|
||||||
|
import numpy as np
|
||||||
|
import unittest
|
||||||
|
from op_test import OpTest
|
||||||
|
|
||||||
|
|
||||||
|
class TestPSROIPoolOp(OpTest):
|
||||||
|
def set_data(self):
|
||||||
|
self.init_test_case()
|
||||||
|
self.make_rois()
|
||||||
|
self.calc_psroi_pool()
|
||||||
|
self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)}
|
||||||
|
self.attrs = {
|
||||||
|
'output_channels': self.output_channels,
|
||||||
|
'spatial_scale': self.spatial_scale,
|
||||||
|
'pooled_height': self.pooled_height,
|
||||||
|
'pooled_width': self.pooled_width
|
||||||
|
}
|
||||||
|
self.outputs = {'Out': self.outs}
|
||||||
|
|
||||||
|
def init_test_case(self):
|
||||||
|
self.batch_size = 3
|
||||||
|
self.channels = 3 * 2 * 2
|
||||||
|
self.height = 6
|
||||||
|
self.width = 4
|
||||||
|
|
||||||
|
self.x_dim = [self.batch_size, self.channels, self.height, self.width]
|
||||||
|
|
||||||
|
self.spatial_scale = 1.0 / 4.0
|
||||||
|
self.output_channels = 3
|
||||||
|
self.pooled_height = 2
|
||||||
|
self.pooled_width = 2
|
||||||
|
|
||||||
|
self.x = np.random.random(self.x_dim).astype('float32')
|
||||||
|
|
||||||
|
def make_rois(self):
|
||||||
|
rois = []
|
||||||
|
self.rois_lod = [[]]
|
||||||
|
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 calc_psroi_pool(self):
|
||||||
|
output_shape = (self.rois_num, self.output_channels, self.pooled_height,
|
||||||
|
self.pooled_width)
|
||||||
|
out_data = np.zeros(output_shape)
|
||||||
|
for i in range(self.rois_num):
|
||||||
|
roi = self.rois[i]
|
||||||
|
roi_batch_id = int(roi[0])
|
||||||
|
roi_start_w = round(roi[1]) * self.spatial_scale
|
||||||
|
roi_start_h = round(roi[2]) * self.spatial_scale
|
||||||
|
roi_end_w = (round(roi[3]) + 1.) * self.spatial_scale
|
||||||
|
roi_end_h = (round(roi[4]) + 1.) * self.spatial_scale
|
||||||
|
|
||||||
|
roi_height = max(roi_end_h - roi_start_h, 0.1)
|
||||||
|
roi_width = max(roi_end_w - roi_start_w, 0.1)
|
||||||
|
|
||||||
|
bin_size_h = roi_height / float(self.pooled_height)
|
||||||
|
bin_size_w = roi_width / float(self.pooled_width)
|
||||||
|
|
||||||
|
x_i = self.x[roi_batch_id]
|
||||||
|
|
||||||
|
for c in range(self.output_channels):
|
||||||
|
for ph in range(self.pooled_height):
|
||||||
|
for pw in range(self.pooled_width):
|
||||||
|
hstart = int(
|
||||||
|
math.floor(float(ph) * bin_size_h + roi_start_h))
|
||||||
|
wstart = int(
|
||||||
|
math.floor(float(pw) * bin_size_w + roi_start_w))
|
||||||
|
hend = int(
|
||||||
|
math.ceil(
|
||||||
|
float(ph + 1) * bin_size_h + roi_start_h))
|
||||||
|
wend = int(
|
||||||
|
math.ceil(
|
||||||
|
float(pw + 1) * bin_size_w + roi_start_w))
|
||||||
|
hstart = min(max(hstart, 0), self.height)
|
||||||
|
hend = min(max(hend, 0), self.height)
|
||||||
|
wstart = min(max(wstart, 0), self.width)
|
||||||
|
wend = min(max(wend, 0), self.width)
|
||||||
|
|
||||||
|
c_in = (c * self.pooled_height + ph
|
||||||
|
) * self.pooled_width + pw
|
||||||
|
is_empty = (hend <= hstart) or (wend <= wstart)
|
||||||
|
out_sum = 0.
|
||||||
|
for ih in range(hstart, hend):
|
||||||
|
for iw in range(wstart, wend):
|
||||||
|
out_sum += x_i[c_in, ih, iw]
|
||||||
|
bin_area = (hend - hstart) * (wend - wstart)
|
||||||
|
out_data[i, c, ph, pw] = 0. if is_empty else (
|
||||||
|
out_sum / float(bin_area))
|
||||||
|
self.outs = out_data.astype('float32')
|
||||||
|
|
||||||
|
def setUp(self):
|
||||||
|
self.op_type = 'psroi_pool'
|
||||||
|
self.set_data()
|
||||||
|
|
||||||
|
def test_check_output(self):
|
||||||
|
self.check_output()
|
||||||
|
|
||||||
|
def test_check_grad(self):
|
||||||
|
self.check_grad(['X'], 'Out')
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
unittest.main()
|
||||||
Loading…
Reference in new issue