I have a onnx case RoiAlign UT that can passed on onnxruntime. (The file is too big size, I put it on your FTP server path:to_verisilicon/shuliu/roialign.zip)
I add a UT for RoiAlign based on the onnx case : (need put the roialign folder under host_build/src/tim)
roi_align_test.cc
/****************************************************************************
*
* Copyright (c) 2022 Vivante Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
* DEALINGS IN THE SOFTWARE.
*
*****************************************************************************/
#include "tim/vx/ops/roi_align.h"
#include "gtest/gtest.h"
#include "test_utils.h"
#include "tim/vx/context.h"
#include "tim/vx/graph.h"
#include "tim/vx/types.h"
#include <iostream>
#include <fstream>
#include <numeric>
TEST(ROI_Align, shape_25_25_256_1_float32) {
auto ctx = tim::vx::Context::Create();
auto graph = ctx->CreateGraph();
uint32_t height = 25;
uint32_t width = 25;
uint32_t channels = 256;
uint32_t batch = 1;
uint32_t num_rois = 1000;
uint32_t depth = channels;
int32_t out_height = 7;
int32_t out_width = 7;
float height_ratio = 0.3125f;
float width_ratio = 0.3125f;
int32_t height_sample_num = 2;
int32_t width_sample_num = 2;
tim::vx::ShapeType input_shape({width, height, channels, batch}); //whcn
tim::vx::ShapeType regions_shape({4, num_rois});
tim::vx::ShapeType batch_index_shape({num_rois});
tim::vx::ShapeType output_shape(
{(uint32_t)out_width, (uint32_t)out_height, depth, num_rois});
tim::vx::TensorSpec input_spec(tim::vx::DataType::FLOAT32, input_shape,
tim::vx::TensorAttribute::INPUT);
tim::vx::TensorSpec regions_spec(tim::vx::DataType::FLOAT32, regions_shape,
tim::vx::TensorAttribute::INPUT);
tim::vx::TensorSpec batch_index_spec(tim::vx::DataType::INT32,
batch_index_shape,
tim::vx::TensorAttribute::INPUT);
tim::vx::TensorSpec output_spec(tim::vx::DataType::FLOAT32, output_shape,
tim::vx::TensorAttribute::OUTPUT);
auto input_count = std::accumulate(input_shape.begin(),
input_shape.end(), 1, std::multiplies<int64_t>());
float *input_data = new float[input_count];
std::ifstream in("roialign/test_data_set_0/input_0.bin", std::ios::in | std::ios::binary);
if(in.is_open())
{
in.read((char *)input_data, input_count*sizeof(float));
std::cout<<"input data:"<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<input_data[i]<<" ";
}
std::cout<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<input_data[input_count-10+i]<<" ";
}
std::cout<<std::endl;
in.close();
}
else
{
std::cout<<"open input_0.bin fail!"<<std::endl;
}
auto regions_count = std::accumulate(regions_shape.begin(),
regions_shape.end(), 1, std::multiplies<int64_t>());
float *regions_data = new float[regions_count];
std::ifstream in1("roialign/test_data_set_0/input_1.bin", std::ios::in | std::ios::binary);
if(in1.is_open())
{
in1.read((char *)regions_data, regions_count*sizeof(float));
std::cout<<"regions data:"<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<regions_data[i]<<" ";
}
std::cout<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<regions_data[regions_count-10+i]<<" ";
}
std::cout<<std::endl;
in1.close();
}
else
{
std::cout<<"open input_1.bin fail!"<<std::endl;
}
auto batch_index_count = std::accumulate(batch_index_shape.begin(),
batch_index_shape.end(), 1, std::multiplies<int64_t>());
int64_t *batch_index_data = new int64_t[batch_index_count];
std::ifstream in2("roialign/test_data_set_0/input_2.bin", std::ios::in | std::ios::binary);
if(in2.is_open())
{
in2.read((char *)batch_index_data, batch_index_count*sizeof(int64_t));
std::cout<<"batch_index data:"<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<batch_index_data[i]<<" ";
}
std::cout<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<batch_index_data[batch_index_count-10+i]<<" ";
}
std::cout<<std::endl;
in2.close();
}
else
{
std::cout<<"open input_2.bin fail!"<<std::endl;
}
auto out_count = std::accumulate(output_shape.begin(),
output_shape.end(), 1, std::multiplies<int64_t>());
float *golden = new float[out_count];
std::vector<float> golden_float(out_count);
std::ifstream out("roialign/test_data_set_0/output_0.bin", std::ios::in | std::ios::binary);
if(out.is_open())
{
out.read((char *)golden, out_count*sizeof(float));
for(auto i=0; i<out_count; i++)
{
golden_float[i] = golden[i];
}
std::cout<<"output data:"<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<golden[i]<<" ";
}
std::cout<<std::endl;
for(int i=0; i<10; i++)
{
std::cout<<golden[out_count-10+i]<<" ";
}
std::cout<<std::endl;
out.close();
}
else
{
std::cout<<"open output_0.bin fail!"<<std::endl;
}
auto input_tensor = graph->CreateTensor(input_spec);
auto regions_tensor = graph->CreateTensor(regions_spec);
auto batch_index_tensor =
graph->CreateTensor(batch_index_spec);
auto output_tensor = graph->CreateTensor(output_spec);
auto roi_align = graph->CreateOperation<tim::vx::ops::ROI_Align>(
out_height, out_width, height_ratio, width_ratio, height_sample_num,
width_sample_num);
(*roi_align)
.BindInput(input_tensor)
.BindInput(regions_tensor)
.BindInput(batch_index_tensor)
.BindOutput(output_tensor);
EXPECT_TRUE(graph->Compile());
input_tensor->CopyDataToTensor(input_data, sizeof(input_data)*sizeof(float));
regions_tensor->CopyDataToTensor(regions_data, sizeof(regions_data)*sizeof(float));
batch_index_tensor->CopyDataToTensor(batch_index_data, sizeof(batch_index_data)*sizeof(float));
EXPECT_TRUE(graph->Run());
std::vector<float> output(num_rois * out_height * out_width * depth);
EXPECT_TRUE(output_tensor->CopyDataFromTensor(output.data()));
EXPECT_EQ(golden_float, output);
}
After running, it is inconsistent with the expected result.
![[QST] custom op : Some question about kernel resource input data type](https://avatars.githubusercontent.com/u/14870906?v=4)
8 Feb 14, 2022
4 Apr 5, 2021
17 Jun 19, 2022
145 Jul 31, 2022
390 Aug 3, 2022
3 Dec 27, 2021
1.1k Aug 2, 2022
6 Oct 27, 2020
18 Apr 20, 2021
4.4k Jul 30, 2022
1 Dec 3, 2021
32 Jul 31, 2022
8 Feb 12, 2022
10 Apr 21, 2022
2 May 28, 2022
72 Jul 4, 2022
20 Jun 16, 2022
1 Feb 26, 2022
25 Mar 8, 2022
5.5k Aug 4, 2022
366 Aug 7, 2022
25 Apr 14, 2022
200 Aug 6, 2022
108 Aug 5, 2022
194 Jul 19, 2022
509 Aug 5, 2022
34 May 5, 2022
16 Jun 3, 2021
47 Jul 20, 2022