MNN/test/op/NormalizeTest.cpp

58 lines
2.2 KiB
C++

//
// NormalizeTest.cpp
// MNNTests
//
// Created by MNN on 2021/10/22.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include <cmath>
#include <MNN/expr/Expr.hpp>
#include <MNN/expr/ExprCreator.hpp>
#include "MNNTestSuite.h"
#include "TestUtils.h"
using namespace MNN::Express;
class NormalizeTest : public MNNTestCase {
public:
static void _refNormalize(float* dst, const float* src, int batch, int channel, int area, float* scale, float eps) {
// Normalize
for (int b=0; b<batch; ++b) {
for (int x=0; x<area; ++x) {
auto dstX = dst + b * area * channel + x;
auto srcX = src + b * area * channel + x;
float sumSquare = 0.0f;
for (int c=0; c<channel; ++c) {
sumSquare += (srcX[area * c] * srcX[area * c]);
}
float normalValue = 1.0f / sqrtf(sumSquare + eps);
for (int c=0; c<channel; ++c) {
dstX[area*c] = srcX[area * c] * normalValue * scale[c];
}
}
}
}
virtual ~NormalizeTest() = default;
virtual bool run(int precision) {
auto input = _Input({1, 2, 2, 1}, NCHW);
// set input data
const float inpudata[] = {-1.0, -2.0, 3.0, 4.0};
auto inputPtr = input->writeMap<float>();
memcpy(inputPtr, inpudata, 4 * sizeof(float));
input = _Convert(input, NC4HW4);
std::vector<float> scaleData = {0.5f, 0.5f};
float eps = 0.00f;
auto output = _Normalize(input, 0, 0, eps, scaleData);
output = _Convert(output, NCHW);
std::vector<float> expectedOutput(4);
_refNormalize(expectedOutput.data(), inpudata, 1, 2, 2, scaleData.data(), eps);
auto gotOutput = output->readMap<float>();
float errorScale = precision <= MNN::BackendConfig::Precision_High ? 1 : 1000;
if (!checkVectorByRelativeError<float>(gotOutput, expectedOutput.data(), 4, 1e-5 * errorScale)) {
MNN_ERROR("NormalizeTest test failed!\n");
return false;
}
return true;
}
};
MNNTestSuiteRegister(NormalizeTest, "op/normalize");